rippled/docs/telemetry-runbook.md

xrpld Telemetry Operator Runbook

Table of Contents

• Overview
• Quick Start
• Configuration Reference
• Exporting to Grafana Cloud
• Span Reference
  • RPC Spans
  • Transaction Spans
  • Transaction Queue Spans
  • PathFinding Spans
  • Consensus Spans
  • Ledger Spans
  • Peer Spans
• Protocol Span Flow
  • Master overview
  • Client and peer ingress
  • Shared transaction apply pipeline
  • Consensus round
  • Accept, build, and finalize the ledger
  • Side flows: pathfinding and ledger acquire
  • Where telemetry parenting differs from protocol flow
• Insights and Sample Queries
  • Transaction Workflow Analysis
  • DEX (OfferCreate / OfferCancel)
  • Apply Pipeline by Stage
  • Transaction Queue Health
  • RPC Debugging
  • PathFinding Performance
  • Consensus Health
  • Cross-Subsystem Correlation
• Cross-Node Trace Propagation
• Prometheus Metrics (Spanmetrics)
• System Metrics (OTel native -- beast::insight)
• Grafana Dashboards
• Alerting
• Log-Trace Correlation
• Troubleshooting
• Performance Tuning
• Disabling Telemetry
• Validating Telemetry Stack
• Performance Benchmarking

Overview

xrpld supports OpenTelemetry distributed tracing to provide visibility into RPC requests, transaction processing, and consensus rounds.

This runbook covers operating a running node and querying its traces. For building xrpld with telemetry support and the internal architecture, see build/telemetry.md. For plain-language definitions of the XRP Ledger terms used in the dashboards, see the telemetry glossary.

Quick Start

1. Start the observability stack

docker compose -f docker/telemetry/docker-compose.yml up -d

This starts:

• OTel Collector on ports 4317 (gRPC) and 4318 (HTTP), and 13133 (health)
• Tempo on http://localhost:3200 (trace backend)
• Prometheus on http://localhost:9090
• Loki on http://localhost:3100 (log aggregation)
• Grafana on http://localhost:3000 (Tempo pre-configured as datasource)

2. Enable telemetry in xrpld

Add to your xrpld.cfg:

[telemetry]
enabled=1
endpoint=http://localhost:4318/v1/traces

3. Build with telemetry support

Follow BUILD.md, adding -o telemetry=True so Conan pulls opentelemetry-cpp. From a build directory (.build/):

conan install .. --output-folder . --build missing -o telemetry=True --settings build_type=Release
cmake -DCMAKE_TOOLCHAIN_FILE:FILEPATH=build/generators/conan_toolchain.cmake -DCMAKE_BUILD_TYPE=Release -Dxrpld=ON -Dtelemetry=ON ..
cmake --build . --target xrpld

Conan also writes a conan-release CMake preset, so cmake --preset conan-release -Dtelemetry=ON works instead of the explicit toolchain line. There is no preset named default.

4. Run against a live network

Two ready-made configs connect a tracking node (no validator credentials) to a public network with all tracing and native metrics enabled:

Config	Network
docker/telemetry/xrpld-telemetry.cfg	Devnet
docker/telemetry/xrpld-telemetry-mainnet.cfg	Mainnet

.build/xrpld --conf docker/telemetry/xrpld-telemetry-mainnet.cfg

Both set [insight] server=otel (native metrics → collector → Prometheus, which drives the dashboards) and service_instance_id, exposed by Prometheus as the service_instance_id label that the $node dashboard variable filters on. The mainnet config logs to /var/log/xrpld/mainnet/debug.log — the path the collector's filelog receiver tails for log-trace correlation.

Metrics begin flowing as soon as the node connects to peers (server_state ≥ connected); full ledger and consensus panels populate after sync (server_state = full). Check progress with:

curl -s http://localhost:5015 -d '{"method":"server_info"}' |
    jq '.result.info | {server_state, peers, complete_ledgers}'

Mainnet sync is bandwidth- and disk-heavy. For a quick check use the devnet config or the standalone test in docker/telemetry/TESTING.md, which generates spans without waiting for a live sync.

Configuration Reference

Option	Default	Description
enabled	0	Master switch for telemetry
endpoint	http://localhost:4318/v1/traces	OTLP/HTTP endpoint
service_name	xrpld	OpenTelemetry service name resource attribute
service_instance_id	node public key	OpenTelemetry service instance ID resource attribute
trace_rpc	1	Enable RPC request tracing
trace_transactions	1	Enable transaction tracing
trace_consensus	1	Enable consensus tracing
trace_peer	1	Enable peer message tracing (high volume)
trace_ledger	1	Enable ledger tracing
consensus_trace_strategy	deterministic	Consensus trace ID strategy (deterministic or attribute)
batch_size	512	Max spans per batch export
batch_delay_ms	5000	Delay between batch exports
max_queue_size	2048	Max spans queued before dropping
use_tls	0	Use TLS for exporter connection
tls_ca_cert	(empty)	Path to CA certificate bundle
tls_client_cert	(empty)	Client cert (PEM) for mutual TLS; empty = one-way TLS
tls_client_key	(empty)	Private key (PEM) for tls_client_cert

consensus_trace_strategy is not validated. The parser copies the raw string through (TelemetryConfig.cpp:155-156) and the only equality test in the code is strategy == "attribute" (RCLConsensus.cpp:1296). Any other value — including a typo such as determinstic — silently selects the deterministic branch. There is no warning in the log. The two accepted values are documented at include/xrpl/telemetry/Telemetry.h:287-292.

Exporting to Grafana Cloud

The collector can ship traces, metrics, and logs to a hosted Grafana Cloud stack instead of (or alongside) the local Tempo/Prometheus/Loki backends. This is a runtime choice — no xrpld rebuild and no change to the base stack. xrpld still exports to the local collector exactly as before; the collector adds one OTLP/HTTP exporter that forwards all three signals to the Grafana Cloud OTLP gateway, which fans them out to hosted Tempo, Mimir, and Loki.

Credentials

Find these under Grafana Cloud → Connections → OpenTelemetry (OTLP):

Value	Used as	Notes
GRAFANA_CLOUD_OTLP_ENDPOINT	exporter endpoint	Full gateway URL incl. /otlp path
GRAFANA_CLOUD_INSTANCE_ID	Basic-auth user	Numeric stack/instance id
GRAFANA_CLOUD_API_TOKEN	Basic-auth pass	Access-policy token with *:write for all signals

Enable

1. Copy the template and fill in the three values:

   cp docker/telemetry/.env.grafanacloud.example docker/telemetry/.env.grafanacloud
   # edit .env.grafanacloud — this file is gitignored, never commit tokens
   

2. Bring the stack up with the base file and the Grafana Cloud override:

   docker compose -f docker/telemetry/docker-compose.yml \
       -f docker/telemetry/docker-compose.grafanacloud.yaml up -d
   

To return to local-only export, bring the stack up with just the base docker-compose.yml.

Files

File	Role
otel-collector-config.grafanacloud.yaml	Collector config: local backends plus a Grafana Cloud OTLP exporter on all three pipelines
docker-compose.grafanacloud.yaml	Override that mounts that config and injects the credentials
.env.grafanacloud.example	Credential template (copy to .env.grafanacloud)

Local + cloud vs cloud-only

The prepared config dual-exports: data goes to both the local stack and Grafana Cloud, so the on-box backends remain a fallback. For cloud-only, remove the local exporters (debug, otlp/tempo, prometheus, otlphttp/loki) from the respective pipelines in otel-collector-config.grafanacloud.yaml, leaving only otlphttp/grafanacloud.

Note

: shipping logs to Grafana Cloud requires keeping xrpld file logging on (at least warning level) so the collector's filelog receiver has a debug.log to tail. Traces and metrics are unaffected by log level.

Importing dashboards to Grafana Cloud

Shipping data (above) is independent of installing the dashboards. The local stack auto-provisions dashboards from a mounted folder (grafana/provisioning/dashboards/dashboards.yaml, type: file); Grafana Cloud cannot read your filesystem, so its dashboards must be imported over the HTTP API or the UI.

The dashboard JSON in docker/telemetry/grafana/dashboards/ references its backends through datasource template variables (${DS_PROMETHEUS}, ${DS_TEMPO}) rather than fixed UIDs. On import, Grafana binds each variable to a datasource of the matching type — auto-selecting it when only one exists (the usual case: one Mimir, one Tempo). This is what makes the same files work unchanged on both the local stack and Cloud.

If you add a dashboard exported with hardcoded datasource UIDs, replace them with ${DS_PROMETHEUS} / ${DS_TEMPO} before committing.

To import:

1. In Grafana Cloud, go to Dashboards → New → Import.
2. Upload a file from docker/telemetry/grafana/dashboards/ (or paste its JSON), then click Load.
3. At the datasource prompt, confirm the auto-selected Prometheus/Mimir datasource — and Tempo for dashboards that query traces — then Import.
4. Repeat per dashboard. Only consensus-health uses Tempo; the rest need only the Prometheus/Mimir datasource.

Span Reference

All spans instrumented in xrpld, grouped by subsystem:

RPC Spans

Span Name	Source File	Attributes	Description
rpc.http_request	ServerHandler.cpp	request_payload_size	Top-level HTTP RPC request
rpc.ws_upgrade	ServerHandler.cpp	—	WebSocket upgrade handshake
rpc.ws_message	ServerHandler.cpp	command, rpc_status	WebSocket RPC message
rpc.process	ServerHandler.cpp	is_batch, batch_size	RPC processing (child of rpc.http_request/ws_message)
rpc.command.<name>	RPCHandler.cpp	command, version, rpc_role, rpc_status, load_type	Per-command span (e.g., rpc.command.server_info)

On rpc.ws_message, rpc_status is set on four of the five error paths (resource threshold exceeded, bad API version / missing command, caught exception, and an error in the command result — ServerHandler.cpp:489, :522, :571, :608). The exception is the invalid-JSON / oversized-request path, which opens its own rpc.ws_message span and calls only setError(), writing no rpc_status at all (ServerHandler.cpp:392-395) — those rejections are visible solely through status_code="ERROR". The success path calls setOk() and writes no rpc_status either, so there is never an rpc_status="success" series for this span: count successes as total minus error, or filter on status_code. rpc.command.* is unaffected — it sets rpc_status on both outcomes.

Transaction Spans

Span Name	Source File	Attributes	Description
tx.process	NetworkOPs.cpp	tx_hash, local, path, tx_type, fee, sequence, ter_result, applied, current_ledger_seq	Transaction submission and processing
tx.receive	PeerImp.cpp	peer_id, tx_hash, tx_type, peer_version, suppressed, tx_status, current_ledger_seq	Transaction received from peer relay
tx.apply	BuildLedger.cpp	tx_count, tx_failed	Transaction set applied per ledger
tx.preflight	applySteps.cpp	stage, tx_type, ter_result	Stateless checks stage
tx.preclaim	applySteps.cpp	stage, tx_type, ter_result, current_ledger_seq, current_ledger_hash	Ledger-aware checks stage
tx.transactor	Transactor.cpp	stage, tx_type, ter_result, applied, current_ledger_seq, current_ledger_hash	Apply stage (transactor runs)

The three apply-pipeline spans (tx.preflight, tx.preclaim, tx.transactor) share a deterministic trace_id from txID[0:16], so they group under one trace per transaction. The stage attribute (preflight / preclaim / apply) drives the collector spanmetrics stage dimension, giving per-stage RED metrics on the Transaction Overview dashboard.

current_ledger_seq is the current (open/in-flight) ledger index a span acted on — the ledger being worked on, not an established one — so a transaction's txID-keyed spans can be joined to the ledger trace it targeted (span.current_ledger_seq). The view-bearing stages (tx.preclaim, tx.transactor) also carry current_ledger_hash (the current ledger's parent hash); tx.preflight is stateless and omits both.

tx.apply carries no ledger_seq of its own — the sequence is set on its parent ledger.build (BuildLedger.cpp:90), so read it from the parent rather than filtering tx.apply on it.

Transaction Queue Spans

Span Name	Source File	Attributes	Description
txq.enqueue	TxQ.cpp	tx_hash, tx_type, current_ledger_seq, current_ledger_hash	Enqueue decision; parents to tx.process on the submission path (explicit context), a root on the open-ledger rebuild path — current_ledger_seq correlates it to the ledger in both cases
txq.apply_direct	TxQ.cpp	--	Direct apply attempt (bypassing queue)
txq.batch_clear	TxQ.cpp	--	Batch clear of queued transactions for an account
txq.accept	TxQ.cpp	queue_size, ledger_changed	Ledger-close accept loop over queued transactions
txq.accept_tx	TxQ.cpp	tx_hash, retries_remaining, ter_code, txq_status	Per-transaction apply during accept
txq.cleanup	TxQ.cpp	ledger_seq	Post-close cleanup of expired queue entries

PathFinding Spans

Span Name	Source File	Attributes	Description
pathfind.request	PathFind.cpp / RipplePathFind.cpp	pathfind_source_account, pathfind_dest_account	Path-find RPC entry (accounts hashed; set when present)
pathfind.compute	PathRequest.cpp	pathfind_fast, pathfind_dest_currency	Path computation for one request (doUpdate)
pathfind.discover	PathRequest.cpp	pathfind_search_level, pathfind_num_paths, pathfind_num_source_assets	Graph exploration (one per RPC call in findPaths)
pathfind.update_all	PathRequestManager.cpp	pathfind_ledger_index, pathfind_num_requests	Async recomputation of active requests on ledger close

Consensus Spans

Span Name	Source File	Attributes	Description
consensus.round	RCLConsensus.cpp	consensus_ledger_id, ledger_seq, consensus_mode, trace_strategy, consensus_round_id	Root span for a consensus round (deterministic or random trace ID)
consensus.phase.open	Consensus.h	open_duration_ms, peer_positions_at_close (both only if the span is still live at closeLedger())	Open phase duration (child of round)
consensus.proposal.send	RCLConsensus.cpp	consensus_round, is_bow_out	Consensus proposal broadcast
consensus.ledger_close	RCLConsensus.cpp	ledger_seq, consensus_mode	Ledger close event
consensus.establish	Consensus.h	converge_percent, establish_count, proposers	Establish phase duration (child of round)
consensus.update_positions	Consensus.h	converge_percent, proposers, disputes_count, avalanche_threshold (only when peer positions exist), have_close_time_consensus, close_time_threshold	Position update and dispute resolution (see Events below)
consensus.check	Consensus.h	agree_count, disagree_count, converge_percent, have_close_time_consensus, threshold_percent, proposers_finished, consensus_stalled, establish_count, consensus_result	Consensus threshold check
consensus.accept	RCLConsensus.cpp	proposers, round_time_ms, quorum, disputes_count, consensus_state	Ledger accepted by consensus
consensus.accept.apply	RCLConsensus.cpp	ledger_seq, close_time, close_time_correct, close_resolution_ms, consensus_state, proposing, round_time_ms, parent_close_time, close_time_self, close_time_vote_bins, resolution_direction, tx_count, disputes_resolved_count	Ledger application with close time details (see Events below)
consensus.validation.send	RCLConsensus.cpp	ledger_seq, proposing, ledger_hash, full_validation, validation_sign_time	Validation sent after accept (follows-from link)
consensus.mode_change	RCLConsensus.cpp	mode_old, mode_new	Consensus mode transition
consensus.proposal.receive	PeerImp.cpp	proposal_trusted, consensus_round, prev_ledger_prefix, position_hash_prefix	Proposal received from peer (extracts parent context from TraceContext when present; falls back to standalone span for older peers)
consensus.validation.receive	PeerImp.cpp	validation_trusted, ledger_seq (only when the validation carries sfLedgerSequence), full_validation, validation_sign_time	Validation received from peer (extracts parent context from TraceContext when present; falls back to standalone span for older peers)

Consensus Span Events

Parent Span	Event Name	Event Attributes	Description
consensus.update_positions	dispute.resolve	tx_id, dispute_our_vote, dispute_yays, dispute_nays	Emitted per dispute when votes are tallied
consensus.accept.apply	tx.included	tx_id	Emitted per transaction included in the accepted ledger
consensus.round	phase.open	--	Round entered the open phase (also re-fired on recovery)
consensus.round	phase.recovery	--	Round started with StartRoundReason::Recovered
consensus.round	phase.establish	--	Round entered the establish phase on close
consensus.round	phase.accepted	--	Round reached the accepted phase
consensus.round	outcome.yes	--	Round settled with consensus reached
consensus.round	outcome.moved_on	--	Round abandoned; the network moved on without us
consensus.round	outcome.expired	--	Round expired without settling

The nine events above are the complete set. The seven on consensus.round carry no event attributes — they are timestamps marking phase entry and the terminal outcome, so a round's whole life reads off one span's event list. Phase entry additionally rewrites the round's span-level consensus_phase attribute, which is why phase.recovery is the one phase event that leaves consensus_phase unchanged (it fires with an empty label). Evidence: RCLConsensus.cpp:1344, 1386, 1400; outcomes are chosen from result_->state at Consensus.h:1517-1525.

Close Time Queries (Tempo TraceQL)

Span attributes are filtered with span.<attr> inside {}. Combine conditions with &&.

# Find rounds where validators disagreed on close time
{name="consensus.accept.apply" && span.close_time_correct = false}

# Find consensus failures (moved_on)
{name="consensus.accept.apply" && span.consensus_state = "moved_on"}

# Find slow ledger applications (>5s)
{name="consensus.accept.apply" && duration > 5000ms}

# Find specific ledger's consensus details
{name="consensus.accept.apply" && span.ledger_seq = 92345678}

# Find all spans in a consensus round (deterministic trace strategy)
{name="consensus.round" && span.consensus_round_id = "<round_id>"}

# Find dispute resolutions
{name="consensus.update_positions"} >> {event:name="dispute.resolve"}

Ledger Spans

Span Name	Source File	Attributes	Description
ledger.build	BuildLedger.cpp	ledger_seq, close_time, close_time_correct, close_resolution_ms	Ledger build during consensus
ledger.validate	LedgerMaster.cpp	ledger_seq, validations	Ledger promoted to validated
ledger.store	LedgerMaster.cpp	ledger_seq	Ledger stored in history
ledger.acquire	InboundLedger.cpp	ledger_seq, acquire_reason, timeouts, peer_count, outcome	Fetch a missing ledger from peers (parent varies — see known issues)

ledger.acquire sets only ledger_seq and acquire_reason when the span opens in init(). outcome has three values, written on two different paths:

outcome	Written where	Meaning
complete	done()	The ledger was fetched.
failed	done()	The acquisition ended on its own without the ledger. Usually it gave up after timeouts_ > kLedgerTimeoutRetriesMax (= 6), but trigger() also fails immediately on an unusable state or transaction map, so a failed span can carry timeouts=0. Carries span status Error.
aborted	destructor	The acquisition was abandoned before finishing — the sweep evicted it a minute after anything last asked for it, or ledgers_ was cleared wholesale by shutdown or by clearFailures(). Status is left Unset, because the shutdown case is benign.

peer_count is written only on the done() path, so it is absent on aborted spans: reading it would go through Overlay, which a destructor running at teardown cannot depend on still existing. timeouts is written on both paths.

A missing outcome has two causes, and neither is a lost span. The common one is that init() satisfied the ledger straight from the local store, so the acquire never went to the network. The other is a hard failure inside tryDB(): a stored header that cannot be this ledger, or a zero account hash, sets failed_ and init() returns without ever calling done(), so no outcome is written. The destructor does not fill the gap either — its if (!isDone()) guard is already false once failed_ is set, because isDone() is complete_ || failed_. Such a span carries ledger_seq and acquire_reason only. Since aborted exists, a missing outcome is no longer how an abandoned acquisition presents.

When reading acquire duration, exclude or split out outcome="aborted". Those spans stay open from init() until the object is destroyed, so they measure how long the acquisition stayed outstanding rather than fetch latency, and will skew a percentile that mixes them with complete. Only on the sweep path is that duration bounded below by the one-minute threshold. The shutdown and clearFailures() paths abort at whatever age the acquisition happened to have, so an aborted span can also be arbitrarily short.

ledger.build does not carry tx_count / tx_failed. Those two live on its child tx.apply span, which is where the set is actually applied (BuildLedger.cpp:191) — join on the trace, not on one span.

Gap: no ledger.* span carries a ledger hash. The attribute constant ledger_span::attr::ledgerHash is declared (LedgerSpanNames.h:41) but is never set by any call site, so ledger.build / ledger.store / ledger.validate / ledger.acquire are identifiable by ledger_seq only. A query filtering on span.ledger_hash over a ledger.* span returns nothing. ledger_hash is set on consensus.validation.send and on the peer spans, so use those when a hash is required.

Peer Spans

Span Name	Source File	Attributes	Description
peer.proposal.receive	PeerImp.cpp	peer_id, proposal_trusted	Proposal received from peer
peer.validation.receive	PeerImp.cpp	peer_id, validation_trusted	Validation received from peer

Both peer receive spans are kConsumer inbound entry points started as fresh trace roots. They never inherit an ambient span left active on the peer thread, so they do not nest under an unrelated transaction's trace. The distributed child span that links back to the sending node is the separate consensus.*.receive / tx.receive span (see Cross-Node Trace Propagation).

---

Protocol Span Flow

This section maps every span type onto the real xrpld control flow and XRPL protocol order (verified against code and docs/consensus.md) — what the code actually executes next, in what order, with which loops and branches. Spans are drawn as labels on real operations, not as their OpenTelemetry parent links; the SDK's span parenting is listed separately in Where telemetry parenting differs from protocol flow.

These diagrams are the canonical key for linking the span hierarchy — every node and every branch is labelled with the span that represents that state or transition, so a span can be wired to its true protocol parent/child by reading the graph. They are therefore drawn exact, not simplified: every real loop, retry, recovery, and drop branch is shown even when it adds clutter.

Naming and edge conventions:

• Rectangle [ ] — a state/operation that emits a span; the first line is the exact span.name, the parenthetical below is the operation.
• Rounded ( ) with (no span) — a real protocol step that emits no span; shown so the flow stays continuous and is never mistaken for a missing span.
• Solid arrow — the code calls or sequences directly into the next operation. When the transition itself emits a span, the edge is labelled →[span.name]; otherwise it carries the branch condition.
• ↻ — the edge repeats (per tx, per peer, per dispute, per pass, per round).
• Dashed arrow — a conditional branch or an async job hand-off.
• Dotted ⇢ ctx — trace context crosses a node boundary over a protobuf peer message (sender.span ⇢ receiver.span); a different node continues the trace — not an in-process call.
• Red-bordered node — a terminal drop / abandon state.

Master overview

Five ingress origins feed two shared engines — the per-transaction apply pipeline and the consensus round — which converge on ledger build → store → validate. Pathfinding and ledger-acquire are side flows. A single ledger can take many consensus rounds to settle (see Consensus round).

flowchart TB
    classDef ingress fill:#1d4ed8,stroke:#1e3a8a,color:#fff;
    classDef engine  fill:#047857,stroke:#064e3b,color:#fff;
    classDef consensus fill:#b45309,stroke:#7c2d12,color:#fff;
    classDef ledger  fill:#6d28d9,stroke:#4c1d95,color:#fff;
    classDef side    fill:#0e7490,stroke:#155e75,color:#fff;
    classDef plain   fill:#334155,stroke:#0f172a,color:#fff;

    subgraph ING["Ingress (protocol entry points)"]
        direction TB
        RPC["rpc.http_request / rpc.ws_message<br/>rpc.ws_upgrade / grpc.MethodName<br/>(client transport in)"]:::ingress
        SUB(["submit command<br/>(no span)"]):::plain
        PRELAY["tx.receive<br/>(peer relay in)"]:::ingress
        PMSG["peer.proposal.receive<br/>peer.validation.receive<br/>(peer overlay in)"]:::ingress
    end

    TXP["tx.process<br/>(NetworkOPs::processTransaction)"]:::engine
    OPEN["txq.enqueue<br/>(open-ledger apply + TxQ decision)"]:::engine
    PIPE["tx.preflight → tx.preclaim → tx.transactor<br/>(SHARED apply pipeline)"]:::engine

    subgraph CONS["Consensus round"]
        direction TB
        ROUND["consensus.round<br/>(Open → Establish → Accepted)"]:::consensus
        ACC["consensus.accept → consensus.accept.apply"]:::consensus
    end

    subgraph LGR["Ledger finalize"]
        direction TB
        BUILD["ledger.build<br/>(tx.apply over agreed set)"]:::ledger
        STORE["ledger.store<br/>(built, NOT yet final)"]:::ledger
        VAL["ledger.validate<br/>(promoted at quorum)"]:::ledger
    end

    subgraph SIDE["Side flows"]
        direction TB
        PF["pathfind.update_all<br/>pathfind.request/compute/discover"]:::side
        ACQ["ledger.acquire<br/>(fetch missing / correct prior)"]:::side
    end

    RPC -.->|submit / submit_multisigned| SUB --> TXP
    RPC -.->|path_find / ripple_path_find| PF
    PRELAY --> TXP
    TXP --> OPEN
    OPEN -->|↻ up to 3 passes| PIPE
    OPEN -.->|txq.accept re-apply queued tx each close ↻| PIPE
    PMSG -->|peerProposal / recvValidation| ROUND
    ROUND --> ACC --> BUILD
    BUILD -->|↻ each tx × up to 3 passes| PIPE
    BUILD --> STORE
    PMSG -. "trusted validations arrive async → checkAccept quorum" .-> VAL
    ROUND -. "avalanche rounds ↻ (threshold 50→65→70→95%)" .-> ROUND
    ACC -->|endConsensus ↻ next round until a ledger validates| ROUND
    ROUND -.->|wrong-ledger: request correct prior| ACQ
    ACQ -.->|switch-ledger: resume round on correct prior| ROUND
    ACQ --> STORE
    VAL -.->|missing ledger| ACQ
    BUILD -.->|every close re-runs| PF

Client and peer ingress

RPC submit and peer relay converge at tx.process, the single NetworkOPs entry. gRPC serves ledger queries only — it has no submit path and never runs doCommand.

flowchart TB
    classDef span fill:#1d4ed8,stroke:#1e3a8a,color:#fff;
    classDef plain fill:#334155,stroke:#0f172a,color:#fff;
    classDef drop fill:#7f1d1d,stroke:#ef4444,color:#fff;

    HTTP["rpc.http_request<br/>(HTTP entry)"]:::span
    PROC["rpc.process<br/>(parse + batch)"]:::span
    CMD["rpc.command.NAME<br/>(one command)"]:::span
    WSU["rpc.ws_upgrade<br/>(WS handshake)"]:::span
    WSM["rpc.ws_message<br/>(one frame)"]:::span
    GRPC["grpc.MethodName<br/>(ledger query)"]:::span
    GH(["handler_ ctx<br/>(no span)"]):::plain
    SUBMIT(["doSubmit<br/>(no span)"]):::plain
    TXP["tx.process<br/>(NetworkOPs::processTransaction)"]:::span
    RELAYOUT(["Overlay::relay fan-out to N peers<br/>(no span; if applied / terQUEUED,<br/>shouldRelay, not tfInnerBatchTxn)"]):::plain
    PREDROP(["Diverged / needNetworkLedger<br/>(no span — dropped before tx.receive)"]):::drop
    RCV["tx.receive<br/>(peer TMTransaction in)"]:::span
    RCVDROP["tx.receive<br/>tx_status = rejected_inner_batch /<br/>suppressed / dropped_no_sync /<br/>dropped_queue_full"]:::drop
    CHK(["checkTransaction<br/>(JtTransaction worker, no span)"]):::plain
    PRELAY_IN(["TMTransaction in (no span)"]):::plain

    HTTP -->|processRequest| PROC
    PROC -->|↻ each batch request → doCommand| CMD
    WSU -. "each inbound frame → onWSMessage" .-> WSM
    WSM -->|doCommand| CMD
    GRPC --> GH
    CMD -.->|submit / submit_multisigned| SUBMIT
    SUBMIT -->|processTransaction| TXP
    TXP -.->|relay applied / queued tx| RELAYOUT

    PRELAY_IN -.->|tracking == Diverged / needNetworkLedger| PREDROP
    PRELAY_IN -->|else| RCV
    RCV -.->|inner-batch / dup / age>4min / JtTransaction full| RCVDROP
    RCV -->|addJob JtTransaction| CHK
    CHK -->|processTransaction, trusted=peer| TXP

    RELAYOUT -. "tx.process ⇢ tx.receive (span_id over TMTransaction)" .-> RCV

Ingress branches (all evidence in code):

• onHandoff: WS upgrade vs peer bundle vs status page vs legacy HTTP (ServerHandler.cpp:227).
• doSubmit: tx_blob present → submit signed blob; absent → server sign-and-submit (Submit.cpp:49).
• tx.process: local RPC → doTransactionSync; peer → doTransactionAsync (JtBatch) (NetworkOPs.cpp:1434).
• Pre-span peer drops (no tx.receive created): Diverged (PeerImp.cpp:1299) / needNetworkLedger (1302), before the span at ~1320.
• Post-span peer drops (span exists, tx_status set, no job enqueued): tfInnerBatchTxn (1348), HashRouter dup/BAD (1361), dropped_no_sync when validated-ledger age > 4 min (1416), dropped_queue_full when JtTransaction jobs > maxTransactions (1421).
• Relay fan-out: an accepted/queued tx.process relays to N peers via Overlay::relay, gated on applied || (non-FULL local) || terQUEUED, HashRouter shouldRelay, and not tfInnerBatchTxn; the span context is injected here (NetworkOPs.cpp:1797).

Inbound consensus messages take a two-stage handler — a fresh-root peer.*.receive span created first (kConsumer, always), then a consensus.*.receive span (only if not dropped) that carries the sender's context — before enqueuing a checkPropose / checkValidation worker job. The drop points are asymmetric: proposals drop entirely before consensus.proposal.receive, while validations can drop both before and after consensus.validation.receive.

flowchart TB
    classDef span fill:#0e7490,stroke:#155e75,color:#fff;
    classDef plain fill:#334155,stroke:#0f172a,color:#fff;
    classDef drop fill:#7f1d1d,stroke:#ef4444,color:#fff;

    PPR["peer.proposal.receive<br/>(freshRoot, always)"]:::span
    PPRDROP(["no consensus.proposal.receive<br/>(untrusted+relay-off / dup /<br/>untrusted+Diverged / untrusted+loaded)"]):::drop
    CPR["consensus.proposal.receive<br/>(carries sender ctx)"]:::span
    CP(["checkPropose worker<br/>(no span)"]):::plain
    SIGP(["sig-fail: charge, drop<br/>(no relay, no span)"]):::drop
    PTP(["processTrustedProposal → peerProposal<br/>(no span)"]):::plain

    PVR["peer.validation.receive<br/>(freshRoot, always)"]:::span
    pvrDrop1(["no consensus.validation.receive<br/>(!isCurrent / relay-off / dup)"]):::drop
    CVR["consensus.validation.receive<br/>(carries sender ctx)"]:::span
    cvrDrop2(["dropped after span<br/>(untrusted+Diverged /<br/>untrusted+loaded → no job/relay)"]):::drop
    CV(["checkValidation worker<br/>(no span)"]):::plain
    SIGV(["!isValid: charge, drop<br/>(no span)"]):::drop
    RV(["recvValidation → handleNewValidation<br/>(no span)"]):::plain
    RELAY(["Overlay::relay fan-out to N peers<br/>(no span)"]):::plain

    PPR -.->|4 drop conditions| PPRDROP
    PPR -->|else| CPR
    CPR -->|addJob JtProposalT/Ut| CP
    CP -.->|!checkSign| SIGP
    CP -->|isTrusted| PTP
    CP -.->|if relay| RELAY

    PVR -.->|3 drop conditions| pvrDrop1
    PVR -->|else| CVR
    CVR -.->|untrusted+Diverged / loaded| cvrDrop2
    CVR -->|addJob JtValidationT/Ut| CV
    CV -.->|!isValid| SIGV
    CV -->|recvValidation| RV
    CV -.->|if relay / cluster| RELAY

Consensus-message drop evidence:

• Both peer.proposal.receive and peer.validation.receive are freshRoot spans created at the top of onMessage (PeerImp.cpp:1766, 2389) — so they exist even for dropped messages.
• Proposal drops (all before consensus.proposal.receive at 1868): untrusted+relay-off (1807), duplicate (1832), untrusted+Diverged (1840), untrusted+loaded (1846).
• Validation drops (asymmetric around consensus.validation.receive at 2476): before — !isCurrent (2426), relay-off (2445), duplicate (2468); after — untrusted+Diverged (2489), untrusted+loaded (2506).
• Worker sig-fail drops (charged kFeeInvalidSignature, suppress processing and relay): checkPropose !checkSign (PeerImp.cpp:3105), checkValidation !isValid (3149).

Shared transaction apply pipeline

The apply pipeline is the single protocol tx-processing chain, expressed in code as one composed call (apply.cpp:118): doApply(preclaim(preflight(), …), …). C++ evaluates inner-to-outer, so preflight runs first, feeds preclaim, which feeds doApply. Each stage inspects the prior stage's TER and no-ops if it already failed.

Four invokers point into this one pipeline; the diagram draws it once.

flowchart TB
    classDef span fill:#047857,stroke:#064e3b,color:#fff;
    classDef plain fill:#334155,stroke:#0f172a,color:#fff;
    classDef inv fill:#1d4ed8,stroke:#1e3a8a,color:#fff;
    classDef drop fill:#7f1d1d,stroke:#ef4444,color:#fff;

    I1(["open-ledger applyOne (no span)<br/>↻ up to 3 passes"]):::inv
    I2["txq.apply_direct<br/>(fee ≥ required)"]:::span
    I3["txq.batch_clear / txq.accept_tx<br/>↻ per queued tx"]:::span
    I4["tx.apply<br/>(consensus set, ↻ each tx × 3 passes)"]:::span

    REPF(["txq path: rules/flags changed?<br/>re-run preflight (no span)"]):::plain
    FREE(["xrpl::apply() (no span)"]):::plain
    PF["tx.preflight<br/>(stateless checks)"]:::span
    PC["tx.preclaim<br/>(ledger-aware checks)"]:::span
    TR["tx.transactor<br/>(mutate stage)"]:::span
    CLS(["classify final TER (no span)"]):::plain
    OK(["Success / erase (no span)"]):::plain
    FAIL(["tef / tem / tel → hard fail, erase"]):::drop
    RETRY(["retriable ter → keep in set (no span)"]):::plain

    I1 --> FREE
    I2 --> FREE
    I3 --> REPF --> FREE
    I4 --> FREE
    FREE --> PF
    PF -->|preflight tesSUCCESS| PC
    PF -. "else → classify (no preclaim/transactor)" .-> CLS
    PC -->|likelyToClaimFee| TR
    PC -. "else → classify (no transactor)" .-> CLS
    TR --> CLS
    CLS --> OK
    CLS --> FAIL
    CLS --> RETRY
    RETRY -. "next pass while pass<3 and changes>0" .-> FREE
    RETRY -. "last pass → drop from set" .-> FAIL

Pipeline gates and retry (evidence):

• preclaim short-circuits if preflight !tesSUCCESS (applySteps.cpp:498); doApply short-circuits if !likelyToClaimFee (applySteps.cpp:532); the transactor mutates only when preclaim is tesSUCCESS (Transactor.cpp:1647).
• Final-TER classification: applied → Success; tef | tem | tel → hard Fail; else → Retry (apply.cpp:226).
• Multi-pass retry: both open-ledger applyOne and consensus tx.apply loop pass < LEDGER_TOTAL_PASSES (= 3); a Retry tx is kept for the next pass, and the final pass converts lingering retriable txs into drops (OpenLedger.h:237, BuildLedger.cpp:129; LEDGER_TOTAL_PASSES OpenLedger.h:29).
• TxQ re-preflight: the queue path re-runs preflight when the ledger's rules/flags changed since enqueue (TxQ.cpp:315).
• TxQ cross-ledger retry: a queued tx that fails with a retriable result keeps its slot with --retriesRemaining (kRetriesAllowed = 10) and is re-applied at a later ledger close; on retriesRemaining ≤ 0 or tef|tem it is dropped with an account retryPenalty (TxQ.cpp:1528).
• TxQ::apply outcome fork: preflight-reject / applied_direct / batch_clear / queued (terQUEUED) / reject (TxQ.cpp:762).

tx.apply is set-level, consensus-only. It wraps the retry-pass loop over the agreed set during buildLedger and exists on no other invoker. It is not a per-transaction span, and TxQ / open-ledger apply create no tx.apply.

Consensus round

beginConsensus → startRound starts the round (consensus.round, Open phase). The heartbeat timer drives Consensus::timerEntry each pass; the round stays in Establish across many heartbeats until the outcome is decided.

A single ledger can take many rounds to settle. Two nested multi-round mechanisms (see docs/consensus.md):

1. Avalanche rounds inside one Establish phase — each timerEntry runs phaseEstablish again (establishCounter_++) and raises the inclusion threshold 50% → 65% → 70% → 95% as the round ages (ConsensusParms.h:145). checkConsensus returning No keeps the node in Establish and loops; a round cannot even Expire before a minimum of avalancheCutoffs.size() × avMinRounds = 4 × 2 = 8 passes (Consensus.h:1937).
2. Retry across consensus rounds — a round can end MovedOn / Expired, meaning the network settled a different ledger. The node still builds a ledger, but the next round's checkLedger detects the wrong prior, switches to WrongLedger / SwitchedLedger mode, acquires the correct ledger, and re-deliberates. A ledger is only truly settled once trusted validations reach quorum (ledger.validate); the alternate is abandoned.

flowchart TB
    classDef span fill:#b45309,stroke:#7c2d12,color:#fff;
    classDef plain fill:#334155,stroke:#0f172a,color:#fff;

    BEGIN(["beginConsensus → startRound<br/>(Proposing OR Observing; no span)"]):::plain
    ROUND["consensus.round<br/>(one attempt at next ledger)"]:::span
    OPENS["consensus.phase.open<br/>(collect txs; buffer peer<br/>proposals / gotTxSet)"]:::span
    HB(["heartbeat → timerEntry<br/>(every LEDGER_MIN_CLOSE; no span)"]):::plain
    CKL(["checkLedger<br/>(correct prior ledger? no span)"]):::plain
    WRONG(["handleWrongLedger → leaveConsensus (no span):<br/>if Proposing send BOW-OUT,<br/>mode → Observing; acquire ledger"]):::plain
    MODE["consensus.mode_change<br/>(mode transition)"]:::span
    POPEN(["phaseOpen: shouldCloseLedger? (no span)"]):::plain
    CLOSE["consensus.ledger_close<br/>(close open ledger, seed disputes)"]:::span
    pSend["consensus.proposal.send<br/>(broadcast our position)"]:::span
    PEST["consensus.establish<br/>(phaseEstablish; avalanche round ↻<br/>threshold 50→65→70→95%)"]:::span
    UPOS["consensus.update_positions<br/>(add/drop disputed txs; child of establish)"]:::span
    acqTx(["acquireTxSet → gotTxSet<br/>(async peer tx set; no span)"]):::plain
    PAUSE(["shouldPause?<br/>(wait on laggards; no span)"]):::plain
    CHECK["consensus.check<br/>(checkConsensus; child of establish)"]:::span
    CTC(["haveCloseTimeConsensus?<br/>(else agree-to-disagree +1s; no span)"]):::plain
    ACCEPT["consensus.accept<br/>(round complete)"]:::span

    BEGIN --> ROUND --> OPENS
    HB -->|under mutex| CKL
    CKL -.->|wrong prior| WRONG
    WRONG --> MODE
    WRONG -. "recovered → re-enter Open (playbackProposals)" .-> OPENS
    WRONG -. "still missing → keep deliberating, defer to peers" .-> HB
    CKL -->|prior OK| POPEN
    HB -.->|phase==Open| POPEN
    HB -.->|phase==Establish| PEST
    POPEN -.->|shouldClose| CLOSE
    CLOSE -.->|mode==Proposing| pSend
    PEST --> UPOS
    UPOS -.->|position changed && Proposing| pSend
    UPOS -.->|disagreeing peer position| acqTx
    acqTx -. "gotTxSet ↻ → new disputes" .-> UPOS
    PEST --> PAUSE
    PAUSE -. "pausing → wait (loop)" .-> HB
    PAUSE -->|ready| CHECK
    CHECK -. "No / Expired < 8 passes → next avalanche round" .-> HB
    CHECK --> CTC
    CTC -. "no CT consensus → loop" .-> HB
    CTC -.->|Yes / MovedOn / Expired ≥ 8| ACCEPT
    ROUND -.->|mode set at start| MODE
    ACCEPT -. "endConsensus → next round ↻ (until a ledger validates)" .-> BEGIN

Consensus loops and branches (evidence):

• consensus.establish is the parent of update_positions and check: phaseEstablish creates the establish span (startEstablishTracing), and both child spans parent to its captured context (Consensus.h:2099, 1628, 1837).
• Avalanche-convergence loop (rounds within one ledger): repeated heartbeat → timerEntry → phaseEstablish bumps establishCounter_ and raises the inclusion threshold each pass; checkConsensus = No stays in Establish (NetworkOPs.cpp:1214; Consensus.h:1467; thresholds ConsensusParms.h:145).
• Retry-across-rounds loop (many rounds per settled ledger): MovedOn / Expired accepts a non-preferred ledger; the next round's checkLedger finds the wrong prior and recovers before re-deliberating (Consensus.h:1193); round-to-round via endConsensus → beginConsensus (NetworkOPs.cpp:2315).
• Two extra establish loop-backs before accept: shouldPause (laggard backpressure) and !haveCloseTimeConsensus_ (TX consensus but not close-time) each return and re-loop, distinct from checkConsensus == No (Consensus.h:1496, 1499); close time can "agree to disagree" at prior close + 1s (docs/consensus.md:163).
• acquireTxSet / gotTxSet loop: a disagreeing peer position triggers an async acquireTxSet; the later gotTxSet regenerates disputes and can extend the establish phase (Consensus.h:931).
• Bow-out / mode change: handleWrongLedger → leaveConsensus sends a bow-out proposal and demotes Proposing → Observing for the rest of the round (Consensus.h:1976); startRound begins in Proposing or Observing (docs/consensus.md:176).
• Buffered Open-phase inputs: peerProposal / gotTxSet arriving during Open are stored, then seeded as disputes at closeLedger (createDisputes); playbackProposals replays them at startRound / handleWrongLedger (docs/consensus.md:244; Consensus.h:816).
• Outcome fork after checkConsensus: No (loop) / Yes (onAccept) / MovedOn / Expired (Consensus.h:1515).
• Expired guard: a round cannot leave on Expired before avalancheCutoffs.size() × avMinRounds (= 8) passes — below that, Expired loops like No (Consensus.h:1937).
• The deterministic-vs-random trace-strategy branch at round start (RCLConsensus.cpp:1291) sets only the trace ID — it has zero protocol effect.

Accept, build, and finalize the ledger

onAccept enqueues a JtAccept job; doAccept runs on that worker (consensus.accept.apply). It builds the ledger (running the apply pipeline over the agreed set), cleans the queue, stores the ledger, optionally broadcasts a validation, and rebuilds the open ledger. A built ledger is not final — it is promoted to ledger.validate only when trusted validations reach quorum, an async, validation-driven path re-entered per incoming trusted validation; a built ledger that loses is abandoned.

flowchart TB
    classDef span fill:#6d28d9,stroke:#4c1d95,color:#fff;
    classDef plain fill:#334155,stroke:#0f172a,color:#fff;
    classDef drop fill:#7f1d1d,stroke:#ef4444,color:#fff;

    onAcc["consensus.accept<br/>(round complete)"]:::span
    APPLY["consensus.accept.apply<br/>(JtAccept worker)"]:::span
    BLCL(["buildLCL: replay data? (no span)"]):::plain
    BUILD["ledger.build<br/>(normal: apply agreed set)"]:::span
    RPLY["ledger.build<br/>(replay: TapNone, no tx.apply child)"]:::span
    TXAP["tx.apply<br/>(↻ each tx × up to 3 passes)"]:::span
    CLEAN["txq.cleanup<br/>(expire queue entries)"]:::span
    STORE["ledger.store<br/>(built, NOT yet final)"]:::span
    vSend["consensus.validation.send<br/>(broadcast our validation)"]:::span
    CACC(["consensusBuilt → checkAccept<br/>(quorum gate; no span)"]):::plain
    NEWVAL(["inbound trusted validation<br/>→ handleNewValidation → checkAccept<br/>(async, per validation; no span)"]):::plain
    VAL["ledger.validate<br/>(promote highest-seq ledger ≥ quorum)"]:::span
    LOSE(["built ledger loses:<br/>never promoted → abandoned"]):::drop
    OACC(["OpenLedger::accept<br/>(rebuild open ledger; no span)"]):::plain
    tqAcc["txq.accept<br/>(↻ drain queued txs)"]:::span
    swlStd(["switchLCL standalone:<br/>setFullLedger + tryAdvance (no span)"]):::plain
    swlNet(["switchLCL networked:<br/>checkAccept (no span)"]):::plain
    END(["endConsensus → next round ↻ (no span)"]):::plain

    onAcc -.->|addJob JtAccept| APPLY
    APPLY --> BLCL
    BLCL -->|normal path| BUILD --> TXAP
    BLCL -. "replay path" .-> RPLY
    APPLY --> CLEAN
    APPLY --> STORE
    APPLY -. "validating && isCompatible && !fail && canValidateSeq" .-> vSend
    APPLY --> CACC
    NEWVAL --> CACC
    CACC -.->|highest-seq trusted ledger ≥ quorum| VAL
    CACC -.->|tvc < quorum → no promotion| LOSE
    APPLY --> OACC
    OACC -->|TxQ::accept callback| tqAcc
    OACC --> swlStd
    OACC --> swlNet
    swlStd -. "marks full-validated (no ledger.validate span)" .-> END
    swlNet --> CACC
    onAcc --> END

• Order inside doAccept: buildLCL (build → tx.apply, then txq.cleanup, then ledger.store) → optional validate → consensusBuilt/checkAccept → OpenLedger::accept (rebuilds the open ledger; txq.accept runs in its callback) → switchLCL promotes the built ledger to the new LCL (RCLConsensus.cpp:812 then 833).
• buildLCL replay branch: if releaseReplay() has data, buildLedger replays the stored set with TapNone — it still emits ledger.build (via buildLedgerImpl) but applies txns directly with no tx.apply child and no 3-pass loop; else the normal consensus-set path runs tx.apply over 3 passes (RCLConsensus.cpp:929; BuildLedger.cpp:252).
• processClosedLedger (txq.cleanup) runs after build, before store (RCLConsensus.cpp:950 vs 953).
• ledger.validate is async + lossy: checkAccept is re-entered per incoming trusted validation (handleNewValidation → checkAccept, RCLValidations.cpp:193); it promotes the highest-seq trusted ledger whose valCount > neededValidations (LedgerMaster.cpp:1180), which may be a different ledger than the one this node built. Below quorum (tvc < minVal) it returns early with no promotion — a built ledger that loses is abandoned (LedgerMaster.cpp:980; docs/consensus.md:50). The ledger.validate span is emitted only inside checkAccept (LedgerMaster.cpp:987).
• validation-send guard: broadcast only if validating_ && isCompatible && !consensusFail && canValidateSeq(seq) — silently suppressed for incompatible ledgers or an already-validated seq (RCLConsensus.cpp:730).
• switchLCL: standalone → setFullLedger + tryAdvance — marks the ledger full-validated without emitting ledger.validate (that span lives only in checkAccept); networked → checkAccept (shared async quorum gate) (LedgerMaster.cpp:442).

Side flows: pathfinding and ledger acquire

Pathfinding — an RPC one-shot (path_find / ripple_path_find) plus an async recompute that fires on every ledger close for all active subscriptions, and also garbage-collects dead subscriptions:

flowchart TB
    classDef span fill:#0e7490,stroke:#155e75,color:#fff;
    classDef plain fill:#334155,stroke:#0f172a,color:#fff;
    classDef drop fill:#7f1d1d,stroke:#ef4444,color:#fff;

    REQ["pathfind.request<br/>(path_find / ripple_path_find)"]:::span
    CREATE(["subscribe: makePathRequest →<br/>persistent subscription (no span)"]):::plain
    COMP["pathfind.compute<br/>(doUpdate, one pass)"]:::span
    DISC["pathfind.discover<br/>(findPaths)"]:::span
    PFDR(["Pathfinder + RippleCalc<br/>↻ per source asset (no span)"]):::plain
    UP(["updatePaths (JtUpdatePf, every close; no span)"]):::plain
    UALL["pathfind.update_all<br/>(recompute all active)"]:::span
    DEAD(["dead subscriber → doAborting +<br/>remove_if erase"]):::drop

    REQ -.->|subcommand create| CREATE
    REQ -->|doUpdate| COMP -->|findPaths| DISC --> PFDR
    UP -->|once per close| UALL
    UALL -->|↻ each active request| COMP
    UALL -. "new request arrived → extra pass ↻" .-> UALL
    UALL -.->|dead / aborted| DEAD

Ledger acquire — a flow outside the close flow that fetches a missing or correct-prior ledger from peers, retries per peer/timer, and finishes with a reason-dependent store; checkAccept + tryAdvance run on any completed acquire. ledger.acquire is usually a trace root, but not reliably so — see the parenting known issues:

flowchart TB
    classDef span fill:#0e7490,stroke:#155e75,color:#fff;
    classDef plain fill:#334155,stroke:#0f172a,color:#fff;
    classDef drop fill:#7f1d1d,stroke:#ef4444,color:#fff;

    HIST(["tryAdvance → doAdvance → fetchForHistory<br/>(Reason::HISTORY; no span)"]):::plain
    NEED(["checkAccept / handleNewValidation /<br/>consensus wrong-ledger (no span)"]):::plain
    INB(["InboundLedgers::acquire (no span)"]):::plain
    ACQ["ledger.acquire<br/>(InboundLedger::init)"]:::span
    TRIG(["trigger / addPeers / onTimer<br/>↻ per peer / chunk (no span)"]):::plain
    FAILED(["timeouts > 6 → failed_ →<br/>logFailure (NO store, NO checkAccept)"]):::drop
    DONE(["done() — complete && !failed (no span)"]):::plain
    ONF(["onLedgerFetched (no span)<br/>(HISTORY: no store)"]):::plain
    STORE["ledger.store<br/>(GENERIC / CONSENSUS)"]:::span
    CACC(["checkAccept + tryAdvance (no span)<br/>(↻ may publish/advance many ledgers)"]):::plain

    NEED -.->|GENERIC / CONSENSUS| INB --> ACQ
    HIST -.->|HISTORY| INB
    ACQ -.->|not complete| TRIG
    TRIG -. "retry ↻" .-> TRIG
    TRIG -.->|timeout cap| FAILED
    TRIG -.->|complete| DONE
    DONE -.->|reason == HISTORY| ONF
    DONE -.->|GENERIC / CONSENSUS| STORE
    DONE --> CACC
    CACC -. "advanceWork ↻ → further HISTORY acquire" .-> HIST

Side-flow evidence:

• Pathfind subscription lifecycle: path_find create inserts a persistent subscription (makePathRequest); update_all re-runs each active request every close, removes dead subscribers (doAborting + remove_if erase), and takes an extra pass when a new request arrived mid-run (PathRequestManager.cpp:103, 169, 181).

• Acquire outcome fork: timeouts_ > kLedgerTimeoutRetriesMax (= 6) sets failed_ → terminal logFailure, no store/checkAccept (InboundLedger.cpp:402). A third path never reaches done() at all: the destructor marks any acquisition that is still neither complete_ nor failed_ as outcome=aborted (InboundLedgers.cpp:393 sweep eviction; InboundLedger.cpp:224 abort branch). Give-up fires at roughly 18s, not 21s: init() enters the retry loop through queueJob() with no preceding setTimer(), so the first invokeOnTimer() runs immediately with progress_ still false and takes timeouts_ to 1 at t≈0. The test needs timeouts_ > 6 — the seventh invocation — and only six 3s intervals separate the seventh from the first, so 6 x 3s = 18s. A live aborted rate does not by itself mean acquisitions are stalling. Three unrelated paths produce it:

  • Sweep eviction — the only cause that implies staleness, and it fires a minute after anything last asked for this ledger, not a minute after the last byte arrived.
  • Shutdown — InboundLedgers::stop() clears ledgers_ wholesale, so every clean stop aborts every acquisition still in flight.
  • clearFailures() — also clears ledgers_, and is reachable at runtime from the fetch_info admin RPC (clear: true → NetworkOPsImp::clearLedgerFetch()), so an operator can produce aborts on a perfectly healthy node.

  The 18s-vs-60s gap does not settle it either: while the acquisition lane sits at its job limit the timer body never runs, so timeouts_ cannot advance and the give-up path is disarmed exactly when aborts are likeliest — see The deferral/timeout pair. Rule out shutdown and clearFailures() first, then read a sustained aborted rate against acquire_sweep_evictions.

• done() reason branch (store side only): HISTORY → onLedgerFetched, no storeLedger; else → storeLedger. But checkAccept + tryAdvance run for any complete_ && !failed_ acquire regardless of reason (InboundLedger.cpp:537 store switch; 552 reason-independent checkAccept/tryAdvance on the AcqDone job).

• tryAdvance multi-ledger loop: doAdvance runs do { … } while (advanceWork_), publishing a range of ledgers and recursively triggering further HISTORY acquire (LedgerMaster.cpp:1905).

Where telemetry parenting differs from protocol flow

The graph above is protocol control flow. The OpenTelemetry span parent links are built differently and, in several places, do not represent a real call edge. Read a trace with these in mind:

Telemetry does this	Real protocol flow
tx.process is a hashSpan root from txID — an independent trace root (TxTracing.h:63).	The real edge is the synchronous doSubmit → processTransaction call; it is not a child of rpc.command.submit.
tx.preflight / tx.preclaim / tx.transactor share one txID-derived trace ID.	That shared ID is a correlation trick, not a call edge. The real order is the composed apply() at apply.cpp:118. They are not children of tx.process or tx.apply. Because nothing else nests under it either, tx.apply is always a leaf — the stage spans for the transactions it applied sit in the txID-keyed trace, not beneath it.
consensus.round uses a deterministic trace ID from the previous ledger hash.	This makes all validators share one trace ID (a cross-node shared root), not a per-node parent. The real round-to-round edge is endConsensus → beginConsensus.
consensus.accept (main thread) and consensus.accept.apply (JtAccept worker) are wired via a captured context.	The real edge is the queued JtAccept job, a thread hand-off (RCLConsensus.cpp:483).
pathfind.update_all parents nothing from the original pathfind.request.	The causal link is the ledger-close job on JtUpdatePf, not span nesting.
ledger.acquire and its downstream ledger.store / ledger.validate.	Reached via the AcqDone job, not parent inheritance. All three are non-scoped SpanGuard::span spans, so none of them parents the others; each takes whatever ambient span its own caller happens to have active. See the ledger.* known issue below.
peer.*.receive (fresh kConsumer root) and consensus.*.receive on the same message.	Two sequential stages of one synchronous handler, not parent/child; on a duplicate/untrusted drop the consensus.*.receive is never created.
Receive spans adopt the sender's trace_id + span_id as a genuine cross-node parent.	Deliberate: the receive span becomes a child of a different node's span (a cross-node context marker, not an in-process edge). tx.receive is asymmetric — it borrows only the sender's span_id and re-derives its own trace_id from txID.

Known telemetry artifacts (from live audits, memory otel-span-hierarchy-audit): an RPC entry span's scope can leak across a reused coroutine worker, and the hashSpan roots (tx.*) — along with ledger.acquire / ledger.store / ledger.validate whenever they do come out parentless — can surface in Tempo as dangling "root span not yet received". These are exporter/parenting artifacts, not real control-flow parents.

Three further divergences are known issues in the code, not deliberate design. Unlike the rows above, these produce a parent that is simply wrong, and all three are pending a code fix:

• grpc.* and pathfind.update_all do not open a fresh root. All four RPC entry points create their span with freshRoot, so a reused coroutine worker cannot leak a stale ambient parent into them (ServerHandler.cpp:473, 640). grpc.<MethodName> (GRPCServer.cpp:173) and pathfind.update_all (PathRequestManager.cpp:91) use the plain constructor instead, so either can be adopted by whatever span happened to be active on the worker that picked the job up. A gRPC call appearing beneath an unrelated transaction's trace is this bug, not a real call edge.

• ledger.acquire / ledger.store / ledger.validate are not reliably roots either. All three use SpanGuard::span (InboundLedger.cpp:113, LedgerMaster.cpp:463, 987), which inherits the ambient span (SpanGuard.cpp:233) rather than freshRoot (245) — the same defect as grpc.* above. Whether they come out as roots depends purely on the caller:

  • Root, as documented. On the JtAdvance / AcqDone job path (LedgerMaster::doAdvance, RCLConsensus::Adaptor::acquireLedger → RCLConsensus.cpp:171) no span is active on the worker, so nothing is inherited. acquireSpan_ itself is a non-scoped SpanGuard, so it never becomes the ambient parent of the ledger.store / ledger.validate that follow it.
  • Mis-parented. InboundLedgers::acquire is also called synchronously from an RPC handler — ledger_request → rpc::getOrAcquireLedger (RPCLedgerHelpers.cpp:483) — which runs inside the scoped rpc.command.<name> span (RPCHandler.cpp:168). There ledger.acquire becomes a child of that RPC command, and when init() is satisfied from the local store the ledger.store / ledger.validate it calls (InboundLedger.cpp:164, 168) land there as siblings. A ledger acquisition nested under an rpc.command.* trace is this bug, not a real call edge.

  ledger.build and tx.apply use the same ambient-parent construct but are safe. ledger.build is a plain ScopedSpanGuard (BuildLedger.cpp:55): its only callers are RCLConsensus::doAccept (RCLConsensus.cpp:935-937) on the JtAccept worker and the replay path (LedgerDeltaAcquire.cpp:208), and every consensus accept span is a non-scoped SpanGuard (RCLConsensus.cpp:598-599), so no ambient span exists to be inherited there. tx.apply (BuildLedger.cpp:123) is reached only synchronously from buildLedgerImpl while ledger.build's scope is live, so its ambient parent is always ledger.build — which is exactly the intended edge.

• consensus.round is not always a root. The consensus_trace_strategy=attribute path has two creation branches; the fallback branch — taken on the first traced round of a run, and whenever consensus tracing is off — sets no parent at all (RCLConsensus.cpp:1310), so that round inherits the ambient context instead of starting a trace. Rounds under the default deterministic strategy are unaffected.

---

Insights and Sample Queries

This section shows what questions you can answer using the span attributes, with example Tempo TraceQL queries.

TraceQL syntax note: span attributes must be referenced with the span. prefix inside {}. Conditions are combined with &&. The | pipeline operator is not supported on this Tempo version.

# General pattern
{name="<span-name>" && span.<attr> = <value> && span.<attr2> != <value2>}

# Duration filter (no prefix needed)
{name="<span-name>" && duration > 500ms}

# Regex match
{name="<span-name>" && span.<attr> =~ "<pattern>.*"}

# Multiple span names
{name = "<span-a>" || name = "<span-b>"}

# Name regex
{name =~ "<pattern>.*" && span.<attr> = <value>}

# Structural: find parent spans that have a matching child/event
{name="<parent>"} >> {event:name="<event-name>"}

Transaction Workflow Analysis

# Find all AMM transactions (AMMDeposit, AMMWithdraw, AMMVote)
{name="tx.process" && span.tx_type =~ "AMM.*"}

# Find a specific AMM operation
{name="tx.process" && span.tx_type = "AMMDeposit"}
{name="tx.process" && span.tx_type = "AMMWithdraw"}
{name="tx.process" && span.tx_type = "AMMVote"}

# Find Payment transactions that failed
{name="tx.process" && span.tx_type = "Payment" && span.ter_result != "tesSUCCESS"}

# Find Payment failures due to path issues
{name="tx.process" && span.tx_type = "Payment" && span.ter_result =~ "tecPATH.*"}

# Compare latency of different transaction types
{name="tx.process" && span.tx_type = "OfferCreate"}
{name="tx.process" && span.tx_type = "Payment"}

# Find high-fee transactions (fee > 1 XRP = 1000000 drops)
{name="tx.process" && span.fee > 1000000}

# Find transactions that were not applied
{name="tx.process" && span.applied = false}

# Find NFTokenMint across tx and txq spans
{name =~ "tx.*|txq.*" && span.tx_type = "NFTokenMint"}

# Find all NFT-related activity
{name =~ "tx.*|txq.*" && span.tx_type =~ "NFToken.*"}

# Find TrustSet transactions (IOU trust lines)
{name="tx.process" && span.tx_type = "TrustSet"}

# Find oracle price updates
{name="tx.process" && span.tx_type = "OracleSet"}

DEX (OfferCreate / OfferCancel)

# All DEX offer creates
{name="tx.process" && span.tx_type = "OfferCreate"}

# Offers killed (ImmediateOrCancel/FillOrKill with no fill)
{name="tx.process" && span.tx_type = "OfferCreate" && span.ter_result = "tecKILLED"}

# Offers that failed due to insufficient funds
{name="tx.process" && span.tx_type = "OfferCreate" && span.ter_result = "tecUNFUNDED_OFFER"}

# Offers failed due to insufficient reserve to place the offer
{name="tx.process" && span.tx_type = "OfferCreate" && span.ter_result = "tecINSUF_RESERVE_OFFER"}

# Offer cancellations
{name="tx.process" && span.tx_type = "OfferCancel"}

# OfferCreate transactions received from peers (cross-node relay)
{name="tx.receive" && span.tx_type = "OfferCreate"}

Apply Pipeline by Stage

# All three stages of one transaction (preflight -> preclaim -> apply)
{name=~"tx.preflight|tx.preclaim|tx.transactor"}

# Transactions that failed at the preclaim stage
{name="tx.preclaim"} | ter_result != "tesSUCCESS"

# Transactions that hard-failed preflight (never reached preclaim/apply)
{name="tx.preflight"} | ter_result != "tesSUCCESS"

PromQL on the span-derived metrics (dashboard: Transaction Overview):

# Per-stage throughput — the funnel preflight >= preclaim >= apply
sum by (stage) (rate(span_calls_total{span_name=~"tx.preflight|tx.preclaim|tx.transactor"}[5m]))

# Per-stage p95 latency
histogram_quantile(0.95, sum by (le, stage) (rate(span_duration_milliseconds_bucket{span_name=~"tx.preflight|tx.preclaim|tx.transactor"}[5m])))

# Per-stage failure rate (ter_result != tesSUCCESS; a failing ter completes the
# span normally, so filter on the attribute, not status_code which only flags exceptions)
sum by (stage) (rate(span_calls_total{span_name=~"tx.preflight|tx.preclaim|tx.transactor", ter_result!~"tesSUCCESS|"}[5m]))

Alerting: a rising tx.preflight / tx.preclaim failure rate points to malformed or stale-sequence submissions (often spam or a misbehaving client); a rising tx.transactor failure rate points to apply-time problems. Alert per stage rather than on a single aggregate so the failing stage is obvious.

Sampling caveat: these stage metrics are span-derived, but head sampling is fixed at 100% and is not configurable — the ratio is a compile-time constant (Telemetry.h:234 static constexpr double samplingRatio = 1.0;) and there is no sampling_ratio config key to set (TelemetryConfig.cpp:139 — "nothing to parse"). So locally these counts are exact, not a sample. Volume reduction is a collector-side tail sampling decision instead, and the only policy shipped is a single 0.5% probabilistic one that lives only in otel-collector-config.grafanacloud.yaml — the base otel-collector-config.yaml has no tail sampling at all, so a stock local stack retains every trace. Where that Cloud policy is in force it applies to the trace-storage branch only; spanmetrics run on a separate branch and still see 100% of spans, so the derived RED metrics stay exact either way.

Transaction Queue Health

# Find transactions rejected from the queue
{name="txq.accept_tx" && span.txq_status = "failed"}

# Find transactions being retried
{name="txq.accept_tx" && span.txq_status = "retried"}

# Find transactions that exhausted retries
{name="txq.accept_tx" && span.txq_status = "retried" && span.retries_remaining = 0}

# Which transaction types get queued most often?
{name="txq.enqueue" && span.tx_type = "Payment"}
{name="txq.enqueue" && span.tx_type = "OfferCreate"}
{name="txq.enqueue" && span.tx_type =~ "NFToken.*"}

# Find ledger closes that applied queued transactions
{name="txq.accept" && span.ledger_changed = true}

RPC Debugging

# Find batch RPC requests
{name="rpc.process" && span.is_batch = true}

# Find large RPC payloads (>100KB)
{name="rpc.http_request" && span.request_payload_size > 100000}

# Find resource-heavy RPC commands (by load_type)
{name =~ "rpc.command.*" && span.load_type = "exceptioned RPC"}

# Find a specific WebSocket command
{name="rpc.ws_message" && span.command = "subscribe"}

# Find server_info calls
{name="rpc.command.server_info"}

# Find slow pathfinding with many source assets
{name="pathfind.discover" && span.pathfind_num_source_assets > 10}

PathFinding Performance

# Find pathfinding for specific currencies
{name="pathfind.compute" && span.pathfind_dest_currency = "USD"}

# Find expensive pathfinding (many source assets to explore)
{name="pathfind.discover" && span.pathfind_num_source_assets > 20}

# Find slow pathfinding requests
{name="pathfind.compute" && duration > 1000ms}

Consensus Health

# Find rounds where consensus timed out (expired)
{name="consensus.accept" && span.consensus_state = "expired"}

# Find rounds where we moved on without full agreement
{name="consensus.accept" && span.consensus_state = "moved_on"}

# Find rounds with many disputes
{name="consensus.accept" && span.disputes_count > 5}

# Find slow consensus rounds (>5s)
{name="consensus.accept" && span.round_time_ms > 5000}

# Find bow-out proposals (node resigned from round)
{name="consensus.proposal.send" && span.is_bow_out = true}

# Correlate validation with its ledger
{name="consensus.validation.send" && span.ledger_hash = "<hash>"}

# Find rounds where validators disagreed on close time
{name="consensus.accept.apply" && span.close_time_correct = false}

# Find both validation send and receive (compare sender vs receiver latency)
{name = "consensus.validation.send" || name = "consensus.validation.receive"}

Cross-Subsystem Correlation

# Follow a transaction from receive through queue to ledger
{name =~ "tx.*|txq.*" && span.tx_type = "Payment" && duration > 500ms}

# Find all NFT-related activity across tx and txq spans
{name =~ "tx.*|txq.*" && span.tx_type =~ "NFToken.*"}

# Find all AMM activity across tx and txq spans
{name =~ "tx.*|txq.*" && span.tx_type =~ "AMM.*"}

# Find cross-node transaction receives (no errors)
{name="tx.receive" && status != error}

Where to Look (Quick Reference)

Question	Span	Key Attributes
"Which tx type is slowest?"	tx.process	span.tx_type + duration
"Why was my tx rejected?"	tx.process	span.ter_result, span.applied
"What AMM operations happened?"	tx.process	span.tx_type =~ "AMM.*"
"What DEX offers failed?"	tx.process	span.tx_type, span.ter_result
"What NFT activity occurred?"	tx.process, txq.enqueue	span.tx_type =~ "NFToken.*"
"Is the TxQ backing up?"	txq.accept	span.queue_size, span.ledger_changed
"Why was my tx dropped from queue?"	txq.accept_tx	span.txq_status, span.ter_code
"Are batch requests a problem?"	rpc.process	span.is_batch, span.batch_size
"Which RPC is expensive?"	rpc.command.*	span.load_type, duration
"Did consensus reach threshold?"	consensus.check	span.consensus_result
"Was consensus outcome normal?"	consensus.accept	span.consensus_state
"Did a validator bow out?"	consensus.proposal.send	span.is_bow_out
"Which ledger was validated?"	consensus.validation.send	span.ledger_hash
"Did close time agreement fail?"	consensus.accept.apply	span.close_time_correct
"What tx work fed a given ledger?"	tx.* / txq.*	span.current_ledger_seq

Correlating a transaction to the ledger it was worked on

The tx.process, tx.receive, txq.enqueue, tx.preclaim, and tx.transactor spans carry current_ledger_seq — the open/in-flight ledger they acted on (not an established ledger). Because these spans are keyed on the transaction id (their own trace) while the ledger/consensus spans are keyed on the ledger, use the attribute to bridge the two id-spaces:

# All transaction-side work recorded against ledger N
{span.current_ledger_seq = N}

# Join to the ledger build/consensus trace for the same ledger
{name="ledger.build" && span.ledger_seq = N}

txq.enqueue and the view-bearing apply stages also carry current_ledger_hash (the current ledger's parent hash), which equals the consensus.round deterministic trace-id seed on the consensus-build path. tx.preflight is stateless and omits both attributes.

---

Cross-Node Trace Propagation

xrpld propagates trace context across nodes via protobuf TraceContext fields embedded in peer-to-peer messages. When Node A sends a transaction, proposal, or validation, it injects its active span's trace/span IDs into the protobuf message. Node B extracts that context on receipt and creates a child span, linking the two nodes into a single distributed trace.

How It Works

Node A (sender)                          Node B (receiver)
+-----------------------------+          +-------------------------------+
| tx.process / consensus.*    |          | PeerImp::onMessage()          |
|   |                         |          |   |                           |
|   v                         |          |   v                           |
| SpanGuard::getTraceBytes()  |          | extract TraceContext from      |
|   |                         |          | protobuf message               |
|   v                         |   send   |   |                           |
| injectSpanContext() --------|--------->|   v                           |
| sets TraceContext fields    |  proto   | txReceiveSpan()               |
| (trace_id, span_id, flags) |  msg     | proposalReceiveSpan()         |
+-----------------------------+          | validationReceiveSpan()       |
                                         |   |                           |
                                         |   v                           |
                                         | child span with parent link   |
                                         +-------------------------------+

Send-Side Injection

Message Type	Injection Point	Mechanism
TMTransaction	NetworkOPs::apply()	Injects tx.process span into relay msg
TMProposeSet	RCLConsensus::propose()	Injects active context into proposal msg
TMValidation	RCLConsensus::validate()	Injects active context into validation msg

Receive-Side Extraction

Message Type	Extraction Point	Helper Function
TMTransaction	PeerImp::onMessage(TMTransaction)	TxTracing::txReceiveSpan()
TMProposeSet	PeerImp::onMessage(TMProposeSet)	ConsensusReceiveTracing::proposalReceiveSpan()
TMValidation	PeerImp::onMessage(TMValidation)	ConsensusReceiveTracing::validationReceiveSpan()

Key Files

File	Role
src/xrpld/telemetry/PropagationHelpers.h	injectSpanContext() — SpanGuard to protobuf
include/xrpl/telemetry/TraceContextPropagator.h	OTel context <-> protobuf conversion primitives
src/xrpld/telemetry/ConsensusReceiveTracing.h	Proposal/validation receive span factories
src/xrpld/telemetry/TxTracing.h	Transaction receive span factory

Backwards Compatibility

Older peers that do not populate TraceContext fields in their messages will simply produce empty trace bytes on the receive side. The extraction helpers detect this and create standalone (root) spans instead of child spans. No errors are logged and no data is lost — the receive span is still created with all its normal attributes, it just lacks a cross-node parent link.

Example Tempo Queries

# Find cross-node transaction traces (tx.receive spans with no errors)
{name="tx.receive" && status != error}

# Find proposals received with cross-node parent context
{} >> {name="consensus.proposal.receive"}

# Trace a transaction across the network by its hash
{name =~ "tx.*" && span.tx_hash = "<hash>"}

# Find all spans in a cross-node consensus trace
{resource.service.name="xrpld" && span.consensus_round_id = "<round_id>"}

# Compare latency between sender and receiver for validations
{name = "consensus.validation.send" || name = "consensus.validation.receive"}

Prometheus Metrics (Spanmetrics)

The OTel Collector's spanmetrics connector automatically derives RED (Rate, Errors, Duration) metrics from every span. No custom metrics code is needed in xrpld.

Generated Metric Names

Prometheus Metric	Type	Description
span_calls_total	Counter	Total span invocations
span_duration_milliseconds_bucket	Histogram	Latency distribution buckets
span_duration_milliseconds_count	Histogram	Latency observation count
span_duration_milliseconds_sum	Histogram	Cumulative latency

Metric Labels

Every metric carries these standard labels:

Label	Source	Example
span_name	Span name	rpc.command.server_info
status_code	Span status	STATUS_CODE_UNSET, STATUS_CODE_ERROR
service_name	Resource attribute	xrpld
span_kind	Span kind	SPAN_KIND_INTERNAL

Additionally, span attributes configured as dimensions in the collector become metric labels. The span attribute keys are already underscore form (the naming convention forbids dots), so the label name matches the attribute name verbatim. Prometheus' dots → underscores sanitization only fires for dotted attribute names (e.g. resource attributes like service.name), which does not apply to these dimensions.

Span Attribute	Metric Label	Applies To
command	command	rpc.command.* spans
rpc_status	rpc_status	rpc.command.* spans
consensus_mode	consensus_mode	consensus.ledger_close spans
local	local	tx.process spans
proposal_trusted	proposal_trusted	peer.proposal.receive spans
validation_trusted	validation_trusted	peer.validation.receive spans

Histogram Buckets

Configured in otel-collector-config.yaml (spanmetrics connector, unit: ms):

1ms, 5ms, 10ms, 25ms, 50ms, 100ms, 250ms, 500ms, 1s, 2s, 3s, 4s, 5s, 10s, 30s

Sub-second boundaries cover RPC/tx/ledger spans; 2s-4s resolve second-scale consensus spans (consensus.round, consensus.establish) that would otherwise pile into one 1s-5s bucket and make histogram_quantile a meaningless interpolation; 10s/30s give the ledger.acquire catch-up tail a measurable home. Boundaries must stay strictly ascending. The native beast::insight histograms (ms-scale RPC/IO timers) keep the original 1ms-5s buckets in Telemetry.cpp — they never exceed 5s, so they need no high-range buckets.

System Metrics (OTel native -- beast::insight)

xrpld has a built-in metrics framework (beast::insight) that exports metrics natively via OTLP to the OTel Collector. These complement the span-derived RED metrics by providing system-level gauges, counters, and timers that don't map to individual trace spans.

Configuration

Add to xrpld.cfg:

[insight]
server=otel
endpoint=http://localhost:4318/v1/metrics
prefix=xrpld

The OTelCollector implementation exports metrics via OTLP/HTTP to the same OTel Collector that receives traces. No separate StatsD receiver is needed.

Fallback: Set server=statsd and address=127.0.0.1:8125 to use the legacy StatsD UDP path. This requires re-enabling the statsd receiver in otel-collector-config.yaml and uncommenting port 8125 in docker-compose.yml.

Metric Reference

Gauges

Prometheus Metric	Source	Description
ledgermaster_validated_ledger_age	LedgerMaster.h:373	Age of validated ledger (seconds)
ledgermaster_published_ledger_age	LedgerMaster.h:374	Age of published ledger (seconds)
state_accounting_{mode}_duration	NetworkOPs.cpp:774	Time in each operating mode (Disconnected/Connected/Syncing/Tracking/Full)
state_accounting_{mode}_transitions	NetworkOPs.cpp:780	Transition count per mode
peer_finder_active_inbound_peers	PeerfinderManager.cpp:214	Active inbound peer connections
peer_finder_active_outbound_peers	PeerfinderManager.cpp:215	Active outbound peer connections
overlay_peer_disconnects	OverlayImpl.h:557	Peer disconnect count
jobq_job_count	JobQueue.cpp:26	Current job queue depth (all types)
jobq_{jobtype}_waiting	JobTypeData.h	Jobs of this type enqueued but not yet running
jobq_{jobtype}_running	JobTypeData.h	Jobs of this type currently executing
jobq_{jobtype}_deferred	JobTypeData.h	Jobs of this type held back because the type's concurrency limit was hit
{category}_bytes_in/out	OverlayImpl.h:535	Overlay traffic bytes per category (57 categories)
{category}_messages_in/out	OverlayImpl.h:535	Overlay traffic messages per category

Note that job_count is exported as jobq_job_count: the JobQueue is constructed with collectorManager_->group("jobq") (Application.cpp:386), GroupImp::makeName() joins prefix and name with a . (Groups.cpp:42), and OTelCollectorImp::formatName() then turns the . into _ and lowercases the whole string (OTelCollector.cpp:855-874). The same mechanism produces the jobq_{jobtype}_* names above and the pre-existing jobq_{jobtype}_milliseconds timing family.

Per-Job-Type Queue Saturation

The three jobq_{jobtype}_{waiting,running,deferred} families expose the per-type counters that JobTypeData already maintained but never exported. {jobtype} is the lowercased JobTypes name, so JtLedgerReq ("ledgerRequest") becomes jobq_ledgerrequest_waiting / _running / _deferred.

They are emitted for every non-special job type — 35 of the 46 declared types. A "special" type is one whose concurrency limit is 0 (JobTypeInfo::special(), JobTypeInfo.h:71-74); the limit logic never applies to it, so its deferred is always zero. The gauge members are declared at JobTypeData.h:78-80 and created at :100-102, next to the existing dequeue/execute events and under the same !info.special() guard (:95). They are published by JobQueue::collect() under the mutex_ that already guards the counters (JobQueue.cpp:66-93) — no new locking.

deferred is the leading indicator. JobQueue::addJob() never rejects work — when a type is at its limit, addRefCountedJob() increments deferred and returns true anyway (JobQueue.cpp:131-142), and finishJob() drains one deferred job per completion (JobQueue.cpp:353-362). Backpressure on a capped type therefore shows up only as latency, after the harm is done. deferred > 0 says the cap is being hit now, before the duration histograms move.

Limits that matter for ledger sync (JobTypes.h:54-77):

Job type	Metric prefix	Limit	Producers
JtPack	jobq_makefetchpack_	1	MakeFetchPack
JtLedgerReq	jobq_ledgerrequest_	3	RcvGetLedger, RcvGetObjByHash
JtLedgerData	jobq_ledgerdata_	3	ProcessLData, GotStaleData, GotFetchPack, AcqDone, InboundLedger
JtUpdatePf	jobq_updatepaths_	1	PthFindNewReq, PthFindOBDB, PthFindNewLed, OB<seq> — see the note below
JtTxnData	jobq_fetchtxndata_	5	TxAcq, ComplAcquire, RcvPeerData

JtUpdatePf has four producers, three of them individually visible. All four run the same updatePaths() work but arrive under different names. Three come through LedgerMaster::newPFWork() (LedgerMaster.cpp:1545), which passes its caller's name straight to addJob: PthFindNewReq (:1512), PthFindOBDB (:1533), and PthFindNewLed (:1984). All three are all-letters, so each is its own handler series. The fourth is "OB" + std::to_string(seq) (OrderBookDBImpl.cpp:84), which contains digits and therefore folds to handler="other" — order-book rebuild traffic is the only one of the four that is not directly attributable. Do not read the whole type as invisible: three of its four producers are named.

Sampling caveat: these are gauges read by the JobQueue::collect() hook, which the beast::insight PeriodicMetricReader drives every 1 s (Telemetry.cpp:441). A deferred spike shorter than the sample interval can be missed entirely. Treat a non-zero reading as real saturation, but do not treat a zero reading as proof that no saturation occurred — cross-check job_queued_us for the same type.

OTel MetricsRegistry Gauges

These gauges are exported via the OTel Metrics SDK PeriodicMetricReader (10s interval), NOT through beast::insight.

Prometheus Metric	Source	Description
server_info{metric="server_state"}	MetricsRegistry.cpp	Operating mode (0=DISCONNECTED .. 4=FULL)
server_info{metric="uptime"}	MetricsRegistry.cpp	Seconds since server start
server_info{metric="peers"}	MetricsRegistry.cpp	Total connected peers
server_info{metric="validated_ledger_seq"}	MetricsRegistry.cpp	Validated ledger sequence number
server_info{metric="ledger_current_index"}	MetricsRegistry.cpp	Current open ledger sequence
server_info{metric="peer_disconnects_resources"}	MetricsRegistry.cpp	Cumulative resource-related peer disconnects
server_info{metric="last_close_proposers"}	MetricsRegistry.cpp	Proposers in last closed round
server_info{metric="last_close_converge_time_ms"}	MetricsRegistry.cpp	Last close convergence time (ms)
server_info{metric="last_close_time"}	MetricsRegistry.cpp	Network close time of last closed ledger (NetClock secs since XRPL epoch). Age = time() - (value + 946684800); close interval = 1/rate(ledgers_closed_total), not a gauge delta
build_info{version="<ver>"}	MetricsRegistry.cpp	Info-style metric (always 1)
complete_ledgers{bound="start|end",index="<N>"}	MetricsRegistry.cpp	Complete ledger range start/end pairs
db_metrics{metric="db_kb_total"}	MetricsRegistry.cpp	Total database size (KB)
db_metrics{metric="db_kb_ledger"}	MetricsRegistry.cpp	Ledger database size (KB)
db_metrics{metric="db_kb_transaction"}	MetricsRegistry.cpp	Transaction database size (KB)
db_metrics{metric="historical_perminute"}	MetricsRegistry.cpp	Historical ledger fetches per minute
cache_metrics{metric="AL_size"}	MetricsRegistry.cpp	AcceptedLedger cache size
nodestore_state{metric="node_reads_duration_us"}	MetricsRegistry.cpp	Cumulative read time (microseconds)
nodestore_state{metric="node_writes_duration_us"}	MetricsRegistry.cpp	Cumulative write time (microseconds)
nodestore_state{metric="read_request_bundle"}	MetricsRegistry.cpp	Read request bundle count
nodestore_state{metric="read_threads_running"}	MetricsRegistry.cpp	Active read threads
nodestore_state{metric="read_threads_total"}	MetricsRegistry.cpp	Total read threads configured
rpc_in_flight_requests	PerfLogImp.cpp	RPC requests currently executing (UpDownCounter)

Sync Diagnosis Signals

More label values on the same nodestore_state gauge. They exist to separate the two different reasons a node is slow to reach full — see Slow to reach full. The nudb_* group is published only when the writable backend is NuDB; a memory or RocksDB backend omits those four label values rather than reporting them as zero.

Prometheus Metric	Source	Description
nodestore_state{metric="read_mean_us"}	MetricsRegistry.cpp	Mean time per backend read (microseconds)
nodestore_state{metric="write_mean_us"}	MetricsRegistry.cpp	Mean time per backend write (microseconds)
nodestore_state{metric="nudb_writers_in_flight"}	MetricsRegistry.cpp	Threads inside a NuDB insert right now
nodestore_state{metric="nudb_writer_depth_x100"}	MetricsRegistry.cpp	Mean queue depth at the NuDB insert mutex, ×100
nodestore_state{metric="nudb_insert_mean_us"}	MetricsRegistry.cpp	Mean NuDB insert time, queueing included (microseconds)
nodestore_state{metric="nudb_insert_max_us"}	MetricsRegistry.cpp	Slowest single NuDB insert seen (microseconds)
nodestore_state{metric="acquire_deferrals"}	MetricsRegistry.cpp	Timer jobs skipped because the lane was full, all lanes
nodestore_state{metric="acquire_timeouts"}	MetricsRegistry.cpp	Timer bodies that ran and advanced retry, all lanes
nodestore_state{metric="acquire_ledger_deferrals"}	MetricsRegistry.cpp	Deferrals from ledger acquisition alone
nodestore_state{metric="acquire_ledger_timeouts"}	MetricsRegistry.cpp	Timeouts from ledger acquisition alone
nodestore_state{metric="acquire_give_ups"}	MetricsRegistry.cpp	Acquisitions that exhausted their retry budget
nodestore_state{metric="acquire_aborts"}	MetricsRegistry.cpp	Acquisitions destroyed before finishing
nodestore_state{metric="acquire_aborts_partial"}	MetricsRegistry.cpp	Subset of aborts that discarded partly built maps
nodestore_state{metric="acquire_completions"}	MetricsRegistry.cpp	Acquisitions that finished successfully
nodestore_state{metric="acquire_sweep_evictions"}	MetricsRegistry.cpp	Acquisitions evicted by the 1-minute sweep

nudb_writer_depth_x100 is fixed-point: divide by 100 to read it. The depth sits just above 1.0 even under load, so an integer gauge would truncate the whole signal away. It is depthSum / depthSamples, both accumulated when an insert enters the critical section, so an insert still in flight is part of the mean.

acquire_deferrals and acquire_timeouts sum every TimeoutCounter subclass — inbound ledgers, transaction sets and the three ledger-replay tasks — because both are recorded in that shared base. They answer "is any lane deferring", not "is ledger acquisition deferring". Use acquire_ledger_deferrals and acquire_ledger_timeouts for the ledger-acquisition diagnosis; see The deferral/timeout pair.

Counters

Prometheus Metric	Source	Description
rpc_requests	ServerHandler.cpp:108	Total RPC request count
ledger_fetches	InboundLedgers.cpp:44	Ledger fetch request count
ledger_history_mismatch	LedgerHistory.cpp:16	Ledger hash mismatch count
warn	Logic.h:33	Resource manager warning count
drop	Logic.h:34	Resource manager drop count

Histograms

Prometheus Metric	Source	Description
rpc_time	ServerHandler.cpp:110	RPC response time (ms)
rpc_size	ServerHandler.cpp:109	RPC response size (bytes)
ios_latency	Application.cpp:438	I/O service loop latency (ms)
pathfind_fast	PathRequests.h:23	Fast pathfinding duration (ms)
pathfind_full	PathRequests.h:24	Full pathfinding duration (ms)

Job Instruments

These five come from the PerfLog job hooks, not from beast::insight, so they are exported by the MetricsRegistry meter. job_queued_us and job_running_us have explicit microsecond bucket views registered (addMicrosecondHistogramView() calls at MetricsRegistry.cpp:310-311; the helper itself is at :197) spanning 100 µs to 60 s; without those the SDK default buckets stop at 10 ms and every quantile saturates.

Prometheus Metric	Kind	Labels	Description
job_queued_total	Counter	job_type, handler	Jobs enqueued
job_started_total	Counter	job_type, handler	Jobs dequeued and started
job_finished_total	Counter	job_type, handler	Jobs run to completion
job_queued_us	Histogram	job_type, handler	Time spent waiting in the queue (µs)
job_running_us	Histogram	job_type, handler	Time spent executing (µs)

The handler Label

job_type names the queue a job ran on, not the code that submitted it. Several job types have more than one producer, so job_type alone cannot attribute a latency spike. The clearest case: RcvGetLedger (PeerImp.cpp:1566) and RcvGetObjByHash (PeerImp.cpp:2603) both submit to JtLedgerReq, so both report as job_type="ledgerRequest". JtLedgerData has five producers, JtUpdatePf has four, and JtAdvance has four.

The handler label carries the name string passed to addJob() — the specific call site. It resolves all of those at once, not just the GetObject path.

The value is sanitized, not raw. A raw job name would be unbounded, because two production job names embed a ledger sequence number:

• "Pub" + std::to_string(ledger->seq()) (LedgerPersistence.cpp:84)
• "OB" + std::to_string(ledger->seq() % 1000000000) (OrderBookDBImpl.cpp:84)

A raw label would mint a new Prometheus series for every ledger — unbounded growth at ~1 series every 3-5 s, forever. MetricsRegistry::sanitiseHandler() (declared inline in src/xrpld/telemetry/MetricsRegistry.h) therefore applies one rule:

• Keep the name when it is non-empty and every character is an ASCII letter.
• Otherwise return the constant "other". An empty name, a digit, a hyphen, or any punctuation all fall here.

Both dynamic names always contain digits, so both collapse to "other". Because the rule is a pure function of compile-time string literals, the label domain is fixed at build time and cannot grow at runtime. That is a stronger guarantee than an allowlist, which would silently mislabel any job added later; this rule degrades to "other" instead.

Five static job names also fall into "other". Sweeping every addJob / addRefCountedJob / postCoro / newPFWork / TimeoutCounter::jobName call site outside src/test finds 48 distinct static name literals. 43 are all-letters and pass through; five are not, and two more are built at runtime from a ledger sequence. The domain is therefore 44 values (43 names plus "other"):

Name	Why it is not all-letters	Call site
GetConsL1	digit	RCLConsensus.cpp:169
GetConsL2	digit	RCLValidations.cpp:135
gRPC-Client	hyphen	GRPCServer.cpp:156
RPC-Client	hyphen	ServerHandler.cpp:332
WS-Client	hyphen	ServerHandler.cpp:376

The consequence worth remembering: handler="other" is a mixed bucket, not a residual. It holds the two per-ledger dynamic names and those five static ones, so GetConsL1 and GetConsL2 — two different JtAdvance producers — are not separable, and neither are the three RPC client-session names. Do not read a handler breakdown as exhaustive. The GetObject path is unaffected: RcvGetObjByHash and RcvGetLedger are all-letters and pass through as distinct series.

The count is "at the time of writing" — it is a property of the source, not of the rule. Re-derive it from the call sites after adding a job rather than trusting this number.

GetObject Request Metrics

Five instruments on the TMGetObjectByHash query path, recorded via the XRPL_METRIC_* macros at their call sites in PeerImp.cpp. Label cardinality is fixed and tiny: two result values, two reason values.

Prometheus Metric	Kind	Labels	Description
getobject_lookup_us	Histogram	none	Time inside the NodeStore fetch loop (µs)
getobject_request_objects	Histogram	none	Objects requested per request
getobject_lookups_total	Counter	result = hit | miss	NodeStore hit/miss volume
getobject_rejected_total	Counter	reason = oversize | malformed_ledgerhash	Requests refused before any NodeStore access
getobject_charge	Histogram	none	Dynamic component of the differential resource charge

Aggregation choices worth knowing when reading these:

• getobject_lookup_us times the whole fetch loop once (processGetObjectByHash(), PeerImp.cpp:2713-2742 — the iteration cap is set at :2713 and the loop ends at :2742), not each iteration. The loop can run up to kHardMaxReplyNodes = 12288 times (Tuning.h:30); timing each fetchNodeObject() would cost more than the lookups. It needs an addMicrosecondHistogramView() entry for the same reason the job histograms do — a 12288-lookup loop routinely exceeds 10 ms, so without the view the metric saturates exactly when it matters.
• getobject_lookups_total is incremented once per request with the batch totals, not once per object. A 12288-iteration loop incrementing per object would be a measurable hot-path cost for no extra information.
• getobject_request_objects records packet.objects_size() — the requested count, which is what the charge bands price on, not the count actually found.
• getobject_charge records only the dynamic part returned by computeGetObjectByHashFee() (PeerImp.cpp:3658-3681), applied at PeerImp.cpp:2757-2758 just after the loop. The admission-time base charge is a constant (kFeeModerateBurdenPeer, PeerImp.cpp:2634) and is already implied.
• getobject_rejected_total counts the two early returns in onMessage(TMGetObjectByHash): the malformed-ledgerhash check (PeerImp.cpp:2569, counter at :2573) and the oversize gate (PeerImp.cpp:2585, counter at :2591). Both fire before the job is enqueued, so a rejected request contributes to no other GetObject metric.

On a healthy local network getobject_rejected_total reads zero — no honest peer sends an oversized request. Verify its panel with a synthetic oversized request; do not assume it works because the query parses.

Adding a New Metric

Use the call-site macros in src/xrpld/telemetry/MetricMacros.h -- no MetricsRegistry.h/.cpp edit is needed for any of these:

Need	Macro
Monotonic tally (never decreases)	XRPL_METRIC_COUNTER_INC / _ADD [+ _LABELED]
Running total that can decrease	XRPL_METRIC_UPDOWN_ADD [+ _LABELED]
Distribution of values (latency, size)	XRPL_METRIC_HISTOGRAM_RECORD [+ _LABELED]
Last-value snapshot (not a distribution)	XRPL_METRIC_GAUGE_RECORD [+ _LABELED] -- requires an ABI v2 opentelemetry-cpp build; this repo currently builds ABI v1, so use the observable-gauge row below instead
Value your own code already tracks, sampled on a timer	XRPL_METRIC_OBSERVABLE_GAUGE_REGISTER / _COUNTER_REGISTER / _UPDOWN_REGISTER

#include <xrpld/telemetry/MetricMacros.h>

// Monotonic counter:
XRPL_METRIC_COUNTER_INC(app_, "my_new_thing_total", "Description of what this counts");

// Value that can go up and down, e.g. in-flight work (no _total suffix -- that
// is reserved for monotonic counters; an UpDownCounter is a current value):
XRPL_METRIC_UPDOWN_ADD(app_, "my_in_flight_requests", "Currently executing", 1);   // on start
XRPL_METRIC_UPDOWN_ADD(app_, "my_in_flight_requests", "Currently executing", -1);  // on finish

// Sampled from your own state, on the OTel export timer (register ONCE, in init code):
XRPL_METRIC_OBSERVABLE_GAUGE_REGISTER(app_, "my_thing_size", "Current size",
    [this] { return static_cast<int64_t>(myThing_.size()); });

Counters use a _total suffix by convention. A histogram whose values can exceed ~10,000 units (e.g. a microsecond duration beyond 10ms) still needs one line added to addMicrosecondHistogramView() in MetricsRegistry.cpp -- the only case that still touches a central file. There is no way to read a metric's current value back from application code -- OTel's API is write-only by design; keep your own state if your logic needs to both record and read a running value (see the Doxygen header in MetricMacros.h for the full explanation).

Deployment Tiers

Multiple xrpld instances can send telemetry to per-tier collectors that all forward to one Grafana stack. Four resource attributes segregate the data so one dashboard set serves every deployment:

Dimension	Attribute	Set by	Example values
Node	service.instance.id	xrpld cfg	alice-laptop, ci-runner-7
Service	service.name	xrpld cfg	xrpld, xrpld-validator
Network	xrpl.network.type	xrpld node	mainnet, testnet, devnet, perf
Environment	deployment.environment	collector	local, test, ci, prod
Work Item	xrpl.work.item	perf-iac	RIPD-7455 (empty outside perf runs)
Branch	xrpl.branch	perf-iac	baseline:<ref>:<commit>, test:<ref>:<commit>
Node Role	xrpl.node.role	perf-iac	validator, peer

Dashboards expose these as the template variables $node, $service_name, $xrpl_network_type, $deployment_environment, $xrpl_work_item, $xrpl_branch, and $xrpl_node_role (each variable name matches its Prometheus label). Select them top-down — work item → branch → node role → node for a perf comparison run, or environment → network → service → node for general use. Selecting All matches every value, including series lacking the label, so mixed old/new data never disappears.

The last three ($xrpl_work_item, $xrpl_branch, $xrpl_node_role) are populated only during perf-iac comparison runs, which stamp them as resource attributes from their own alloy pipeline. Outside those runs the labels are absent; leaving the filters on All keeps every dashboard rendering normally.

Who owns which attribute

• Node and service come from xrpld config (service_instance_id, service_name). Unique per process.
• Network is a property of the chain the node joined; the node derives it from [network_id] and stamps xrpl.network.type on all three signals.
• Environment is a property of where the collector runs; each collector serves one environment and stamps it.

The upsert vs insert rule

The collector's resource/tier processor uses two actions on purpose:

• deployment.environment → upsert (overwrite). The collector is the environment, so it is authoritative.
• xrpl.network.type → insert (fill only if absent). The node knows its real network, so the collector must not overwrite it — insert only supplies a value when the source did not (e.g. an older xrpld build). This is what lets a local node connected to mainnet report network=mainnet, not the collector's default.

Configuring a collector for a tier

Each tier runs its own collector. Set the two values in the resource/tier processor of the collector config (otel-collector-config.yaml for local backends, otel-collector-config.grafanacloud.yaml for Grafana Cloud):

processors:
  resource/tier:
    attributes:
      - key: deployment.environment
        value: <tier> # local | test | ci | prod
        action: upsert
      - key: xrpl.network.type
        value: <network> # mainnet | testnet | devnet (fallback only)
        action: insert

Suggested per-tier values:

Collector	deployment.environment	xrpl.network.type (fallback)
Developer laptop	local	devnet
Test machines	test	testnet
CI runs	ci	testnet
Production observer	prod	mainnet

The xrpl.network.type value is only a fallback: when the node stamps its own network (all current builds do), the node's value wins. Set it to the network the collector most commonly serves.

How the tier labels reach metrics

Resource attributes do not become Prometheus labels automatically. Two collector settings make it work, both already enabled:

• prometheus.resource_to_telemetry_conversion: enabled: true promotes resource attributes to metric labels on the local scrape surface.
• spanmetrics.resource_metrics_key_attributes lists the tier attributes so span-derived series stay grouped per node and tier.

Traces and logs carry resource attributes natively; Grafana Cloud ingests all three signals' attributes over OTLP directly.

Grafana Dashboards

Fifteen dashboards are pre-provisioned in docker/telemetry/grafana/dashboards/. Fourteen are Prometheus-backed; log-derived-insights is the only Loki/LogQL board and is documented last, together with the LogQL-specific traps it exposed.

Nine of the fifteen have a reference section. Eight are in this chapter (rpc-performance, transaction-overview, consensus-health, ledger-operations, peer-network, node-health, network-traffic, rpc-pathfinding); the ninth, log-derived-insights, is documented under Log-Trace Correlation. The remaining six — fee-market, job-queue, ledger-data-sync, overlay-traffic-detail, peer-quality, and validator-health — are provisioned but not yet documented here. Their panel descriptions carry the same six-heading reference format, so open the panel info icon in Grafana until a section is written.

RPC Performance (rpc-performance)

Panel	Type	PromQL	Labels Used
RPC Request Rate by Command	timeseries	sum by (command) (rate(span_calls_total{span_name=~"rpc.command.*"}[5m]))	command
RPC Latency p95 by Command	timeseries	histogram_quantile(0.95, sum by (le, command) (rate(span_duration_milliseconds_bucket{span_name=~"rpc.command.*"}[5m])))	command
RPC Error Rate	bargauge	Error spans / total spans × 100, grouped by command	command, status_code
RPC Latency Heatmap	heatmap	sum(increase(span_duration_milliseconds_bucket{span_name=~"rpc.command.*"}[5m])) by (le)	le (bucket boundaries)
Overall RPC Throughput	timeseries	rpc.http_request + rpc.process rate	—
RPC Success vs Error	timeseries	by status_code (UNSET vs ERROR)	status_code
Top Commands by Volume	bargauge	topk(10, ...) by command	command
WebSocket Message Rate	stat	rpc.ws_message rate	—

Transaction Overview (transaction-overview)

Panel	Type	PromQL	Labels Used
Transaction Processing Rate	timeseries	rate(span_calls_total{span_name="tx.process"}[5m]) and tx.receive	span_name
Transaction Processing Latency	timeseries	histogram_quantile(0.95 / 0.50, ... {span_name="tx.process"})	—
Transaction Path Distribution	piechart	sum by (local) (increase(span_calls_total{span_name="tx.process"}[5m]))	local
Transaction Receive vs Suppressed	timeseries	rate(span_calls_total{span_name="tx.receive"}[5m])	—
TX Processing Duration Heatmap	heatmap	tx.process histogram buckets	le
TX Apply Duration per Ledger	timeseries	p95/p50 of tx.apply	—
Peer TX Receive Rate	timeseries	tx.receive rate	—
TX Apply Failed Rate	stat	rate(span_calls_total{span_name="tx.transactor",stage="apply",ter_result!~"tesSUCCESS|"})	stage, ter_result
TxQ Accept: Applied Ratio per Node	state-timeline	applied / (applied+failed) of span_calls_total{span_name="txq.accept_tx"} per node	txq_status, service_instance_id

Consensus Health (consensus-health)

Panel	Type	PromQL	Labels Used
Consensus Round Duration	timeseries	histogram_quantile(0.95 / 0.50, ... {span_name="consensus.accept"})	—
Consensus Proposals Sent Rate	timeseries	rate(span_calls_total{span_name="consensus.proposal.send"}[5m])	—
Ledger Close Duration	timeseries	histogram_quantile(0.95, ... {span_name="consensus.round"}) (full round, not consensus.ledger_close which is only the sub-ms onClose prologue)	consensus_mode
Validation Send Rate	stat	rate(span_calls_total{span_name="consensus.validation.send"}[5m])	—
Ledger Apply Duration	timeseries	histogram_quantile(0.95 / 0.50, ... {span_name="consensus.accept.apply"})	—
Close Time Agreement	timeseries	rate(span_calls_total{span_name="consensus.accept.apply"}[5m])	—
Consensus Mode Over Time	timeseries	consensus.ledger_close by consensus_mode	consensus_mode
Accept vs Close Rate	timeseries	consensus.accept vs consensus.ledger_close rate	—
Validation vs Close Rate	timeseries	consensus.validation.send vs consensus.ledger_close	—
Accept Duration Heatmap	heatmap	consensus.accept histogram buckets	le

Ledger Operations (ledger-operations)

Panel	Type	PromQL	Labels Used
Ledger Build Rate	stat	ledger.build call rate	—
Ledger Build Duration	timeseries	p95/p50 of ledger.build	—
Ledger Validation Rate	stat	ledger.validate call rate	—
Build Duration Heatmap	heatmap	ledger.build histogram buckets	le
TX Apply Duration	timeseries	p95/p50 of tx.apply	—
TX Apply Rate	timeseries	tx.apply call rate	—
Ledger Store Rate	stat	ledger.store call rate	—
Build vs Close Duration	timeseries	p95 ledger.build vs consensus.round (full round, not consensus.ledger_close which is only the sub-ms onClose prologue)	—
Ledger Close Interval & Age	timeseries	Interval: 1/rate(ledgers_closed_total); Age: time() - (server_info{metric="last_close_time"} + 946684800)	—

Peer Network (peer-network)

Requires trace_peer=1 in the [telemetry] config section.

Panel	Type	PromQL	Labels Used
Proposal Receive Rate	timeseries	peer.proposal.receive rate	—
Validation Receive Rate	timeseries	peer.validation.receive rate	—
Proposals Trusted vs Untrusted	piechart	increase() counts in the selected window, split by proposal_trusted	proposal_trusted
Validations Trusted vs Untrusted	piechart	increase() counts in the selected window, split by validation_trusted	validation_trusted

Node Health -- System Metrics (node-health)

Panel	Type	PromQL	Labels Used
Validated Ledger Age	stat	ledgermaster_validated_ledger_age	—
Published Ledger Age	stat	ledgermaster_published_ledger_age	—
Operating Mode (Time Share)	timeseries	rate(state_accounting_X_duration) / sum(rate(all modes))	—
Operating Mode Transitions	timeseries	state_accounting_*_transitions	—
I/O Latency	timeseries	histogram_quantile(0.95, ios_latency_milliseconds_bucket)	—
Job Queue Depth	timeseries	jobq_job_count	—
Ledger Fetch Rate	stat	rate(ledger_fetches[5m])	—
Ledger History Mismatches	stat	rate(ledger_history_mismatch_total[5m])	—
Key Jobs Execution Time	timeseries	histogram_quantile($quantile, sum by (le) (rate(job_running_us_bucket{job_type="acceptLedger"}[$__rate_interval]))) (+ 10 more key jobs)	job_type
Key Jobs Dequeue Wait Time	timeseries	histogram_quantile($quantile, sum by (le) (rate(job_queued_us_bucket{job_type="acceptLedger"}[$__rate_interval]))) (+ 10 more)	job_type
FullBelowCache Size	timeseries	node_family_full_below_cache_size	—
FullBelowCache Hit Rate	gauge	node_family_full_below_cache_hit_rate	—
Ledger Publish Gap	stat	Published_Ledger_Age - Validated_Ledger_Age	—
State Duration Rate (Full vs Tracking)	timeseries	rate(state_accounting_full_duration[5m]) / 1000000	—
All Jobs Execution Time (Detail)	timeseries	histogram_quantile($quantile, sum by (le, job_type) (rate(job_running_us_bucket[$__rate_interval])))	job_type
All Jobs Dequeue Wait (Detail)	timeseries	histogram_quantile($quantile, sum by (le, job_type) (rate(job_queued_us_bucket[$__rate_interval])))	job_type
Server State	stat	server_info{metric="server_state"}	metric
Uptime	stat	server_info{metric="uptime"}	metric
Peer Count	stat	server_info{metric="peers"}	metric
Validated Ledger Seq -- Lag Behind Network Tip	stat	max by (xrpl_network_type) (server_info{metric="validated_ledger_seq"}) - on(xrpl_network_type) group_right() server_info{metric="validated_ledger_seq"}	metric
Validated Ledger Seq -- Convergence (Max - Min, per network)	stat	max by (xrpl_network_type) (server_info{metric="validated_ledger_seq"}) - min by (xrpl_network_type) (server_info{metric="validated_ledger_seq"})	metric
Build Version	stat	build_info	version
Complete Ledger Ranges	table	complete_ledgers	bound, index
Database Sizes	timeseries	db_metrics{metric=~"db_kb_.*"}	metric
Historical Fetch Rate	stat	db_metrics{metric="historical_perminute"}	metric

The four job panels read the MetricsRegistry histograms fed by the PerfLog job hooks, not the jobq_* ones. $quantile is a dashboard template variable holding a fraction (0.95), fed straight into histogram_quantile(). There is no quantile label on any xrpld series — that was a StatsD-era summary convention, and a selector like {quantile="$quantile"} matches nothing and reports no error. The job queue exposes two parallel families: job_running_us / job_queued_us (MetricsRegistry instruments, labelled by job_type and handler, microseconds — what these panels use; MetricsRegistry.cpp:94-95, 363-366, recorded from the PerfLog job hooks at PerfLogImp.cpp:432) and jobq_<jobtype>[_q]_milliseconds (beast::insight, one instrument per job type, milliseconds — JobTypeData.h:97). Both are live; prefer the labelled job_*_us pair so one query covers every job type.

Network Traffic -- System Metrics (network-traffic)

Panel	Type	PromQL	Labels Used
Active Peers	timeseries	peer_finder_active_*_peers	—
Peer Disconnects	timeseries	increase(overlay_peer_disconnects[$__rate_interval])	—
Total Network Bytes	timeseries	rate(total_bytes_in/out[$__rate_interval])	—
Total Network Messages	timeseries	rate(total_messages_in/out[$__rate_interval])	—
Transaction Traffic	timeseries	rate(transactions_messages_in/out[$__rate_interval])	—
Proposal Traffic	timeseries	rate(proposals_messages_in/out[$__rate_interval])	—
Validation Traffic	timeseries	rate(validations_messages_in/out[$__rate_interval])	—
Traffic by Category	bargauge	topk(10, label_replace(sum by (service_instance_id)(rate(<metric>[$__rate_interval])),"__name__","<metric>","","") or …)	—
Duplicate Traffic (Wasted Bandwidth)	timeseries	rate(*_duplicate_bytes_in/out[$__rate_interval])	—
All Traffic Categories (Detail)	timeseries	topk(15, label_replace(sum by (service_instance_id)(rate(<metric>[$__rate_interval])),"__name__","<metric>","","") or …)	—

Why the per-category panels enumerate each metric. A bare rate({__name__=~".*_bytes_in"}[…]) fails on Mimir/Cloud with "vector cannot contain metrics with the same labelset": rate() drops the __name__ label, so the many matched counters collapse to identical labelsets. Wrapping in sum by (__name__, …) does not help (the inner vector is rejected before the outer sum). The working form enumerates each *_bytes_in metric and re-attaches its name with label_replace(..., "__name__", "<metric>", "", ""), so the existing {{__name__}} legend and the per-series display-name overrides keep working.

RPC & Pathfinding -- System Metrics (rpc-pathfinding)

Panel	Type	PromQL	Labels Used
RPC Request Rate	stat	rate(rpc_requests[5m])	—
RPC Response Time	timeseries	histogram_quantile(0.95, rpc_time_milliseconds_bucket)	—
RPC Response Size	timeseries	histogram_quantile(0.95, rpc_size_milliseconds_bucket)	—
RPC Response Time Heatmap	heatmap	rpc_time_milliseconds_bucket	—
Pathfinding Fast Duration	timeseries	histogram_quantile(0.95, pathfind_fast_milliseconds_bucket)	—
Pathfinding Full Duration	timeseries	histogram_quantile(0.95, pathfind_full_milliseconds_bucket)	—
Resource Warnings Rate	stat	rate(warn_total[$__rate_interval])	—
Resource Drops Rate	stat	rate(drop_total[$__rate_interval])	—

The _milliseconds suffix comes from the exporter, not from xrpld. These histograms are created with unit "ms" (OTelCollector.cpp:615), so the Prometheus exporter appends the unit to the family name — rpc_time becomes rpc_time_milliseconds_bucket. Querying the bare rpc_time_bucket, ios_latency_bucket or pathfind_fast_bucket returns no data and no error. Known issue: rpc_size counts bytes but shares the same "ms" histogram constructor, so it is exported as rpc_size_milliseconds_bucket — the suffix is wrong, the name is nonetheless the one to query.

Span → Metric → Dashboard Summary

Span Name	Prometheus Metric Filter	Grafana Dashboard
rpc.http_request	{span_name="rpc.http_request"}	RPC Performance (Overall Throughput)
rpc.ws_upgrade	{span_name="rpc.ws_upgrade"}	-- (available but not paneled)
rpc.ws_message	{span_name="rpc.ws_message"}	RPC Performance (WebSocket Rate)
rpc.process	{span_name="rpc.process"}	RPC Performance (Overall Throughput)
rpc.command.*	{span_name=~"rpc.command.*"}	RPC Performance (Rate, Latency, Error, Top)
tx.process	{span_name="tx.process"}	Transaction Overview (Rate, Latency, Heatmap)
tx.receive	{span_name="tx.receive"}	Transaction Overview (Rate, Receive)
tx.apply	{span_name="tx.apply"}	Transaction Overview + Ledger Ops (Apply)
txq.enqueue	{span_name="txq.enqueue"}	-- (available but not paneled)
txq.apply_direct	{span_name="txq.apply_direct"}	-- (available but not paneled)
txq.batch_clear	{span_name="txq.batch_clear"}	-- (available but not paneled)
txq.accept	{span_name="txq.accept"}	-- (available but not paneled)
txq.accept_tx	{span_name="txq.accept_tx"}	-- (available but not paneled)
txq.cleanup	{span_name="txq.cleanup"}	-- (available but not paneled)
consensus.round	{span_name="consensus.round"}	-- (available but not paneled)
consensus.phase.open	{span_name="consensus.phase.open"}	-- (available but not paneled)
consensus.establish	{span_name="consensus.establish"}	-- (available but not paneled)
consensus.update_positions	{span_name="consensus.update_positions"}	-- (available but not paneled)
consensus.check	{span_name="consensus.check"}	-- (available but not paneled)
consensus.accept	{span_name="consensus.accept"}	Consensus Health (Duration, Rate, Heatmap)
consensus.proposal.send	{span_name="consensus.proposal.send"}	Consensus Health (Proposals Rate)
consensus.ledger_close	{span_name="consensus.ledger_close"}	Consensus Health (Close, Mode)
consensus.validation.send	{span_name="consensus.validation.send"}	Consensus Health (Validation Rate)
consensus.accept.apply	{span_name="consensus.accept.apply"}	Consensus Health (Apply Duration, Close Time)
consensus.mode_change	{span_name="consensus.mode_change"}	-- (available but not paneled)
consensus.proposal.receive	{span_name="consensus.proposal.receive"}	-- (available but not paneled)
consensus.validation.receive	{span_name="consensus.validation.receive"}	-- (available but not paneled)
ledger.build	{span_name="ledger.build"}	Ledger Ops (Build Rate, Duration, Heatmap)
ledger.validate	{span_name="ledger.validate"}	Ledger Ops (Validation Rate)
ledger.store	{span_name="ledger.store"}	Ledger Ops (Store Rate)
ledger.acquire	{span_name="ledger.acquire"}	-- (available but not paneled)
peer.proposal.receive	{span_name="peer.proposal.receive"}	Peer Network (Rate, Trusted/Untrusted)
peer.validation.receive	{span_name="peer.validation.receive"}	Peer Network (Rate, Trusted/Untrusted)

Alerting

xrpld provisions thirteen Grafana alert rules on the health-critical metrics, so a stock stack alerts out of the box with no UI setup. Rules are provisioned from docker/telemetry/grafana/provisioning/alerting/ and load automatically when the Grafana container starts. They appear under Alerting → Alert rules, folder xrpld.

All rules ship isPaused: true. Thresholds are tuned against a small dev/devnet population, so every rule is deactivated on arrival — compare it against your own baseline, then unpause. The key is camelCase: is_paused is silently ignored by the provisioning loader (no error, no warning) and leaves the rule live. Note the sibling field notification_settings is snake_case.

Alert catalogue

All rules evaluate every minute against the Prometheus datasource and aggregate by (service_instance_id) so each node alerts on its own. Every expr selects {service_name="xrpld"} — the same Prometheus may also host a legacy statsd fleet exporting some of these names (state_accounting_* in particular) with no xrpld resource attributes, and without the selector those series get summed in. Alerts fire only after the condition holds for the for dwell time.

Alert	Severity	Fires when	For
LedgerHistoryMismatch	critical	increase(ledger_history_mismatch_total[15m]) > 0	2m
LedgerCloseStalled	critical	rate(ledgers_closed_total) ≈ 0	3m
ValidatedLedgerStale	critical	ledgermaster_validated_ledger_age > 60s	5m
ValidationsMissed	warning	validator miss ratio > 0.1	15m
ValidationsNotChecked	warning	rate(validations_checked_total) ≈ 0	5m
JobQueueTxOverflow	warning	increase(jq_trans_overflow_total[15m]) > 0	2m
JobQueueLatencyHigh	warning	p99 job_queued_us > 1s	5m
NodeStoreIOLatencyHigh	warning	p95 ios_latency_milliseconds > 1s	10m
NodeStateFlapping	warning	> 3 re-entries into FULL per hour	15m
NodeNotFull	warning	server_state < 4 (FULL)	15m
ManifestJobQueueConvoy	warning	jobq_manifest_waiting > 3	10m
ManifestFloodInbound	warning	rate(overhead_manifest_bytes_in) > 512 KiB/s	10m
PeerResourceDisconnects	warning	> 5 resource-driven peer disconnects per 30m	5m

Two expression idioms recur and are load-bearing — do not "simplify" them away:

• Sparse counters use increase(...[15m]) with a short for, not rate(...[5m]) with for: 5m. A single increment keeps rate[5m] nonzero for only ~4 minutes of dwell, so a 5-minute for can never be satisfied and the rule silently never fires for the one-off events it exists to catch.
• "Node stopped doing X" rules synthesise an explicit zero via or (0 * max_over_time(...[1h])), because sum by() returns rows only for still-reporting nodes: one dead node's row simply disappears from the result, so noDataState never triggers unless every node vanishes at once.

Consensus / ledger health

LedgerHistoryMismatch — The node closed a ledger whose history diverges from the validated network chain. Likely causes: corrupted local state, a bug, or a node that fell out of sync and rebuilt incorrectly. Investigate the node's ledger acquisition logs; a healthy node never mismatches.

Query trap — sum(ledger_history_mismatch_total) double-counts. One mismatch increments two instruments inside the same handleMismatch() call: the legacy beast::insight counter, which carries no reason label (LedgerHistory.cpp:323), and the MetricsRegistry counter, which does (LedgerHistory.cpp:331). Both normalise to the same Prometheus family, so an unfiltered sum() or increase() reports exactly twice the real mismatch count. Aggregate over the labelled series only — sum by (reason) (...), or sum(ledger_history_mismatch_total{reason!=""}) — and halve any historical figure taken from the unfiltered form. The alert rule is unaffected: it only tests > 0. This is a known issue; the duplicate producer awaits a code fix.

LedgerCloseStalled — No ledgers closed for 3 minutes. A healthy node closes one every ~3-5s. Likely causes: lost peer connectivity, consensus stall, or the process is hung. This rule also fires on NoData — if the series disappears the node is likely down. Check peer count and process health first.

ValidatedLedgerStale — The validated ledger has fallen more than 60s behind. This is the clearest single "is this node healthy" signal on XRPL: it is the symptom nearly every consensus or sync failure eventually produces, so it is often the first thing to check and the last thing to clear. Measured over 7 days: p50 2s, p95 4s, p99 5s on every node.

The < 1209600 clause in this rule's expression is required — do not remove it. When a node holds no validated ledger at all, LedgerMaster::getValidatedLedgerAge() returns weeks{2} (1 209 600 s) as a sentinel, not a measurement. Without the clause the rule reads that as "14 days stale" and fires on every node during startup — measured, it produced sustained firing on all nine nodes over a six-day window, healthy ones included. A node genuinely stuck without a validated ledger is caught by LedgerCloseStalled and NodeNotFull instead.

Validator health

ValidationsMissed — This validator's validations are not agreeing with the validated ledger. Sustained misses risk removal from UNLs. Check clock sync, peer connectivity, and whether the node is keeping up with ledger close.

Why this is a ratio gated on validations_sent_total, not rate(validation_missed_total) > 0: ValidationTracker classifies a ledger as a miss whenever weValidated && networkValidated is not both true. A node that does not validate never sets weValidated, so every reconciled ledger counts as a miss and the raw rate is permanently nonzero — the measured miss ratio is exactly 1.0 on non-validating nodes. No threshold can separate "not a validator" from "validator disagreeing", so the rule gates on validations_sent_total > 0 to exclude non-validators entirely, and then measures the ratio among nodes that genuinely do validate.

ValidationsNotChecked — The node has stopped checking incoming validations from peers. Likely causes: overlay/peer disconnection or a stalled validation pipeline. Fires on NoData as well.

Job queue / resource health

JobQueueTxOverflow — The transaction job queue is full and transactions are being dropped. The node is shedding load it cannot process. Check CPU, the JobQueueLatencyHigh alert, and offered load.

JobQueueLatencyHigh — p99 queue wait exceeds 1 second, i.e. jobs back up before running. The node is saturated. Correlate with CPU and the Job Queue dashboard.

NodeStoreIOLatencyHigh — p95 node-store IO latency exceeds 1s. Sustained store latency is the usual upstream cause of state flapping and sync stalls, so this often fires alongside NodeStateFlapping and explains it. Check disk utilisation and whether the node store sits on a slow volume — moving it to a local NVMe has previously cut time-to-full by more than 3x. Measured p99-of-p95 is 37-49ms on healthy nodes and 488-566ms on nodes that are actively flapping.

Node operating state

NodeStateFlapping — The node is oscillating full → syncing/connected → full instead of holding sync. Measured: a flapping node re-enters full 4-6 times per hour sustained, while a healthy node manages 0-1, so the > 3 threshold sits between the two populations with roughly a 3x margin.

The rule counts state_accounting_full_transitions, which counts transitions into full and is exported as a cumulative gauge — increase() is therefore correct, and its counter-reset correction turns a process restart into a small positive delta rather than a false spike. state_changes_total cannot be used here: it carries no from/to labels, so it cannot tell a flap from a normal startup walk.

The uptime > 3600 gate is load-bearing. Every node walks disconnected → connected → syncing → tracking → full once at boot; without the gate, every restart pages. The trade-off is deliberate: flapping confined to the first hour after boot is not alerted.

Investigate in this order: NodeStoreIOLatencyHigh (most common cause), peer connectivity, then clock sync.

NodeNotFull — The node has been below FULL for 15m (0=disconnected, 1=connected, 2=syncing, 3=tracking, 4=full). This is deliberately a separate rule from NodeStateFlapping: a node that drops to syncing and stays there produces no further full-transitions, so the flapping counter by definition cannot catch it.

Overlay / manifests

ManifestJobQueueConvoy — Manifest jobs are backing up in the job queue. Peers send TMManifests dumps up to ~57MB (just under kMaximumMessageSize, see overlay/Message.h), and JtManifest is registered with maxLimit (core/JobTypes.h), so every peer's dump runs concurrently and they convoy on ManifestCache::mutex_; OverlayImpl::onManifests also re-verifies the blob a second time on Accept. Measured effect: each RcvManifests job took 16-18s and the entire 8-worker pool was occupied.

This is the most reliable manifest-flood signal because jobq_manifest_waiting is 0 at the 99.9th percentile on every node over 24h — any sustained backlog is a genuine outlier rather than normal variance.

ManifestFloodInbound — Inbound manifest byte-rate exceeds 512 KiB/s (524288 B/s — the rule's literal params: [524288]). Catches the wire-level cause (a peer shipping oversized dumps) even when the job pool absorbs it without a visible backlog. Measured over 7 days: healthy p95 0.2-0.5 kB/s and p99 1.0-1.8 kB/s, against peaks up to 2.7 MB/s during real storms — so the threshold sits ~280x above healthy p99 and ~5x below the peaks.

An earlier revision used 50 kB/s, justified from a 24-hour window. Over a full week that produced ~41 sustained 5-minute firings across six healthy nodes, i.e. routine paging. Prefer a 7-day sample when tuning any threshold here; 24 hours is too short to expose weekly variation.

Both manifest rules deliberately suppress startup. The manifest storm at boot is measured normal behaviour, so ManifestFloodInbound carries an uptime > 1800 gate and ManifestJobQueueConvoy relies on a 10m dwell that the startup burst does not outlast. A flood confined to the first 30 minutes after boot will therefore not alert.

PeerResourceDisconnects — The node dropped more than 5 peers in 30m for exceeding resource budgets. Sustained disconnects starve the node of peers and precede sync loss.

Tuning thresholds

Thresholds live in docker/telemetry/grafana/provisioning/alerting/rules.yaml as the params array of each rule's C (threshold) node. Common tunables:

• JobQueueLatencyHigh — params: [1000000] is 1 000 000 µs (1s). Lower it for latency-sensitive deployments.
• LedgerCloseStalled / ValidationsNotChecked — use lt with a tiny epsilon (0.001) rather than 0, so floating-point rate noise near zero does not suppress the alert.

Edit the file and restart the Grafana container to reload:

docker compose -f docker/telemetry/docker-compose.yml restart grafana

Sending alerts somewhere real

Two contact points are provisioned in docker/telemetry/grafana/provisioning/alerting/contactpoints.yaml:

Contact point	Receivers	Gets
xrpld-default	Slack	warning-severity alerts
xrpld-critical	Slack + email	critical-severity alerts

The severity split lives in docker/telemetry/grafana/provisioning/alerting/policies.yaml: the root route sends everything to xrpld-default, and a child route matching severity = critical overrides to xrpld-critical. So a critical alert goes to Slack and email; a warning goes to Slack only. Both group by alertname + service_instance_id; critical alerts re-page hourly vs the 4h default.

Configure delivery (no secrets in git)

The Slack webhook and email address are not hard-coded. contactpoints.yaml ships deliberately unroutable placeholders — an https://hooks.slack.invalid/… host and an …@xrpld.invalid address — which keep provisioning valid so the stack boots with zero configuration while alerts route nowhere.

To enable delivery, edit those two values in place with a real webhook and address, and do not commit the result.

cp docker/telemetry/.env.alerting.example docker/telemetry/.env.alerting
# edit .env.alerting — gitignored; holds the SMTP relay settings
$EDITOR docker/telemetry/grafana/provisioning/alerting/contactpoints.yaml
docker compose -f docker/telemetry/docker-compose.yml up -d grafana

• Slack — replace the placeholder url: with an incoming-webhook URL. Drives both tiers.
• Email — replace the placeholder addresses: (comma- or semicolon-separated) and point the GF_SMTP_* vars in .env.alerting at a real relay with GF_SMTP_ENABLED=true. Grafana can only send mail once SMTP is configured.

Three traps worth knowing before you edit this file:

• Do not substitute ${SLACK_WEBHOOK_URL} / ${ALERT_EMAIL_TO} here. Grafana expands ${VAR} but does not support ${VAR:-default}, so an unset variable expands to empty, fails validation, and Grafana exits 1 — taking the whole telemetry stack down, not just alerting. A blank variable does not "disable that path"; it breaks startup.
• Never empty a receivers: list to disable a tier. A contact point with no receivers ceases to exist, the policy tree then references a missing receiver, and Grafana refuses to boot. Point the route at a contact point that still exists instead.
• File provisioning is upsert-only. Deleting a receiver from the YAML does not remove it from an instance that already booted with it — the old receiver keeps delivering. Removal needs an explicit deleteContactPoints: block listing the uid (a commented example sits at the bottom of contactpoints.yaml).

To add a third destination (PagerDuty, Opsgenie, a custom webhook), add a receiver to the relevant contact point.

Panel screenshots on alerts need a matching render token

Alert notifications that carry a panel image are rendered by the renderer sidecar, not by Grafana itself. Grafana 13 enables the renderAuthJWT feature toggle by default, so the renderer rejects any request whose token is missing or still the - default — notifications then arrive with no image.

docker-compose.yml feeds both sides from one variable, so they cannot drift: GF_RENDERING_RENDERER_TOKEN on the grafana service and AUTH_TOKEN on the renderer service both read ${GF_RENDERING_RENDERER_TOKEN}, defaulting to a local development value. Override it in the environment to use your own:

GF_RENDERING_RENDERER_TOKEN=$(openssl rand -hex 16) \
    docker compose -f docker/telemetry/docker-compose.yml up -d grafana renderer

If images stop appearing, check that the two containers agree — a token set on only one side fails exactly this way:

docker compose -f docker/telemetry/docker-compose.yml exec grafana \
    printenv GF_RENDERING_RENDERER_TOKEN
docker compose -f docker/telemetry/docker-compose.yml exec renderer \
    printenv AUTH_TOKEN

Deploying alerts to Grafana Cloud

Grafana Cloud has no provisioning filesystem, so these apiVersion: 1 files cannot be loaded there. Cloud deployment goes through the Grafana alerting REST API, driven from the same tracked rules.yaml — it stays the single source of truth, so local and Cloud cannot drift.

Each rule needs three Cloud-specific transforms on the way out:

Field in rules.yaml	Cloud form
local prometheus datasource uid	the Cloud datasource uid
folder: name	an existing folderUID
interval (duration string)	integer seconds

Then, in order:

1. Dry-run first — render what would be sent and review it before writing anything to the Cloud stack.
2. Create the rules paused, so nothing can fire on a threshold that has not been reviewed against this fleet.
3. Read back the deployed rules and verify they are what was sent.

Land the rules with delivery disabled while no recipient has been chosen, and activate them only once the thresholds have been checked against the target fleet's baseline.

Credentials come from .env.grafanaserviceapi (gitignored, a service-account token with alert.rules:write); the recipient address comes from ALERT_EMAIL_TO in .env.alerting. Neither is ever written to a tracked file.

The Cloud notification policy tree must not be pushed. There is exactly one policy tree per org and the PUT endpoint replaces it wholesale. On a shared stack the root receiver and its sibling routes belong to other teams, so pushing an xrpld-shaped tree would silently re-route their alerts. The uploader therefore never touches the tree; instead each rule carries notification_settings.receiver, which routes that rule directly to the xrpld contact point and bypasses the tree entirely. Verify with a before/after hash of GET /api/v1/provisioning/policies.

Verifying alert provisioning loaded

After the stack is up:

# All thirteen rules present, and is each one paused?
curl -s http://localhost:3000/api/v1/provisioning/alert-rules |
    jq -r '.[] | "\(.title)\tpaused=\(.isPaused)"'

# Contact points present?
curl -s http://localhost:3000/api/v1/provisioning/contact-points | jq '.[].name'

Check paused=true explicitly rather than assuming it: a mis-spelled is_paused is dropped without any error and the rule provisions live.

Grafana logs a provisioning error and skips the file if the YAML is malformed:

docker compose -f docker/telemetry/docker-compose.yml logs grafana | grep -i alerting

A malformed expression fails differently and more quietly — the rule loads but every evaluation errors. After a threshold or expr change, confirm each rule's query still returns data:

# Should print a numeric value per node, and no empty results
curl -sG http://localhost:9090/api/v1/query \
    --data-urlencode 'query=sum by (service_instance_id) (rate(ledgers_closed_total{service_name="xrpld"}[5m]))' |
    jq '.data.result | length'

Log-Trace Correlation

When xrpld is built with telemetry=ON, log lines emitted within an active, sampled OpenTelemetry span automatically include trace_id and span_id fields:

2024-Jan-15 10:30:45.123456 UTC LedgerMaster:NFO trace_id=abc123def456789012345678abcdef01 span_id=0123456789abcdef Validated ledger 42

This enables bidirectional navigation between logs and traces in Grafana:

• Tempo -> Loki: Click "Logs for this trace" on any trace in Grafana Tempo to see all log lines from that trace.
• Loki -> Tempo: Click the TraceID derived field link on any log line containing trace_id= to jump to the full trace in Tempo.

Log Ingestion Pipeline

Log files are ingested by the OTel Collector's filelog receiver, which tails debug.log files and parses them with a regex that extracts timestamp, partition, severity, trace_id, span_id, and message fields. Parsed entries are exported to Grafana Loki.

The receiver tails /var/log/xrpld/*/debug.log inside the collector container. docker-compose bind-mounts the host log root there; the source defaults to the repo-relative docker/telemetry/data/logs, which the telemetry configs write to (data/logs/<network>/debug.log) and which needs no root. To tail logs from elsewhere, set XRPLD_LOG_DIR before docker compose up (the integration test does this to point at its own workdir). The single trailing * matches one per-network or per-node subdirectory.

Each file is read from the beginning, because the receiver's own default (end) would skip anything a node wrote before the collector's first poll and would never read a log that has stopped being written to. Read offsets are held in memory by default, so a restarted collector re-reads the files it already ingested. The developer stack avoids that by layering otel-collector-filestorage.yaml as a second --config, which adds a file_storage extension that keeps the offsets on a named volume; a one-shot init service prepares that volume, because the collector runs as a non-root user and a fresh Docker volume is owned by root. Ephemeral stacks such as the workload validation harness create a fresh log directory per run, so they have nothing to resume from and deliberately omit the overlay.

LogQL Query Examples

The OTel Collector emits logs to Loki with service_name="xrpld" (not job="xrpld").

For log-derived panels built on these queries, see the Log-Derived Insights dashboard and its LogQL trap list — partition, severity, and xrpl_network_type are structured metadata, not stream labels, so they must be filtered with | after the selector and cannot be discovered by label_values().

# Find all logs for a specific trace
{service_name="xrpld"} |= "trace_id=abc123def456789012345678abcdef01"

# Error logs with trace context (log lines with ERR severity that have a trace_id).
# Use the severity field, not `|= "ERR"`: a line filter also matches the literal
# "ERR" anywhere in the message body (measured: 4 DBG lines per 6h on devnet).
{service_name="xrpld"} | severity = `ERR` | trace_id != ""

# All logs from a specific partition that were emitted during a span.
# Prefer the structured-metadata filter over a line match: `|= "LedgerMaster"`
# also matches the substring anywhere in the message body.
{service_name="xrpld"} | partition = `LedgerMaster` | trace_id != ""

# Logs from a specific subsystem during a span (e.g. LedgerConsensus)
{service_name="xrpld"} | partition = `LedgerConsensus` | trace_id != ""

# Logs from the last hour containing trace context. `partition`, `severity`, and
# `trace_id` are already parsed into structured metadata by the collector's
# filelog receiver, so re-extracting them with regexp is unnecessary work.
{service_name="xrpld"} | trace_id != ""

# Count of traced vs untraced log lines
sum(count_over_time({service_name="xrpld"} | trace_id != "" [5m]))
sum(count_over_time({service_name="xrpld"} | trace_id = "" [5m]))

Verifying Log Correlation

1. Start the observability stack and xrpld with telemetry enabled.
2. Send an RPC request: curl http://localhost:5005 -d '{"method":"server_info"}'
3. Check the debug.log for trace_id= entries: grep trace_id= /path/to/debug.log
4. Open Grafana at http://localhost:3000 -> Explore -> Loki and search for {service_name="xrpld"} | trace_id != "".
5. Click the TraceID link to navigate to the corresponding trace in Tempo.

Log-Derived Insights (log-derived-insights)

The only Loki/LogQL dashboard. It surfaces detail that no metric or span records, by parsing debug.log text. 41 panels in 10 rows: 8 stat, 18 timeseries, 2 table, 1 state-timeline, 1 logs, 1 text, across 35 queries.

REQUIRES DEBUG LOGS for most rows. xrpld's default threshold is Info (Severity thresh = Severity::Info, app/main/Main.cpp). Rows tagged [DBG] read DBG-severity lines that a default node never writes, so those panels are empty on an unmodified node — and an empty panel means not collecting, not no problem. Rows tagged [DEFAULT OK] work as shipped.

Enable per partition rather than globally (Resource alone emits ~329k lines/6h):

log_level ManifestCache debug
log_level Resource debug
log_level InboundLedger debug
log_level Peer debug
log_level PeerFinder debug

Those five cover every [DBG] row. The [MIXED] stat row additionally reads LedgerConsensus and LoadMonitor, both of which already emit at the default level, so its error/consensus/breach/sync panels populate without any change — only its manifest, fee, and fetch-waste panels need debug enabled.

Row	Gate	Key panels
Worst Offenders — Node Ranking	[MIXED]	8 stat panels ranking nodes by error volume, attack-like input, total fee charged, manifest rejections, consensus problems, job latency breaches, sync instability, and ledger fetch waste
Node Operating State Transitions	[DEFAULT OK]	Transition rate and state timeline from STATE-> (NetworkOPsImp::setMode, info)
Log Volume & Severity Mix	[DEFAULT OK]	Line rate by severity; top-N partitions by rate
Manifests — Disposition & Producers	[DBG]	Disposition rate; accept-vs-reject; top-N master keys
Resource Fee Charges — Load Attribution	[DBG]	Charge rate by reason; fee-weighted load; top-N peers by IP and public key
Ledger Acquisition Efficiency	[DBG]	Duplicate ratio; good vs duplicate vs timeout
Peer Lifecycle & Disconnects	[DBG]	Disconnect reason breakdown (Closed / Ping Timeout / Connect Timeout / Connection Refused); handshake and accept rate
Consensus Phase & Mode	[DEFAULT OK]	Phase transitions; operating-mode proxy; quorum and trusted-set size
Slow Job Latency Breaches	[DEFAULT OK]	Run p99, wait p99, breach rate by job (LoadMonitor, >500ms only)
Error & Warning Stream	[DEFAULT OK]	WRN/ERR/FTL rate by partition; live log tail

Filters: $service_name, $deployment_environment, $node, $xrpl_network_type, $severity, plus log-derived $consensus_phase, $consensus_mode, $manifest_action, $charge_reason, and $topn.

LogQL traps this dashboard exposed

Eleven mistakes that fail silently — each cost a debugging cycle, so check them before adding any LogQL panel.

1. partition is structured metadata, not a stream label. {service_name="xrpld", partition="ManifestCache"} returns zero rows with no error. Correct form: {service_name="xrpld"} | partition = \ManifestCache`. Stream labels are only service_name, service_instance_id, deployment_environment. Everything else — partition, severity, xrpl_network_type, message, trace_id` — is structured metadata.

2. label_values() cannot see structured metadata. A query-type template variable over xrpl_network_type, severity, or partition returns an empty dropdown; only true stream labels populate. Use a custom variable with enumerated values instead. This is why filters appeared blank.

3. A target with no datasource uid resolves to the DEFAULT datasource. The Prometheus dashboards use {"type": "prometheus"} with no uid and work only because Prometheus is the default. A Loki target written the same way sends LogQL to Prometheus and returns nothing. Always pin {"type": "loki", "uid": "${DS_LOKI}"}.

4. $__rate_interval is Prometheus-only — Loki panels must use [$__auto]. Grafana does not substitute $__rate_interval for a Loki target, so Loki receives the literal string and fails with parse error: not a valid duration string: "$__rate_interval", which surfaces as "No data". The other 14 dashboards all use $__rate_interval because they are Prometheus-backed; do not align LogQL panels to that convention.

5. Loki caps a query at 2000 series. Any per-key or per-IP aggregation must be wrapped in topk(N, ...) or it fails with HTTP 400. A true distinct-key count over a large key space is therefore not possible in a panel.

6. Loki tables need labelsToFields plus reduce. Loki attaches labels to the Value field instead of returning columns, so a table panel renders bare Time/Value without labelsToFields. Grafana also runs a Loki table target as a range query even when instant: true is set, producing one row per series per timestamp — visible as the same key repeated many times. Use reduce(lastNotNull, labelsToFields) then organize, and note the value column is then named Last *, which any field override must match.

7. Title Case legends need label_format, not value mappings. A label-driven legend renders the raw log value (full, moderate peer request). Grafana value mappings do not help — they map the metric value, not label text in displayName. Rewrite the label in the query: | label_format state=\{{if eq .state "full"}}Full{{else}}{{.state}}{{end}}``.

8. unwrap must be the last pipeline stage. Any label filter or label_format placed after | unwrap <field> makes the query invalid and it returns zero frames.

9. Non-matching lines yield an empty label. A line in the selected partition that does not match the panel's regexp still passes through with an empty extracted label, which renders as a blank legend entry. Guard with | <label> != \`` before aggregating.

10. Stat panels need [$__range], not [$__auto]. Grafana runs a Loki stat target as a range query even with instant: true, so lastNotNull reads only the final bucket — a window total shows as a single-bucket count. Aggregate over $__range and reduce with max.

11. A regexp anchored on the log prefix silently drops most matches. The Peer Disconnect Rate By Reason panel anchored its capture on \] , which only matches a reason emitted immediately after the [NNN] peer-id prefix. PeerImp does not log that way: PeerImp::fail emits [NNN] <name> failed: <reason> (src/xrpld/overlay/detail/PeerImp.cpp:645) and the clean teardown emits close: Closed (:635). Only ConnectAttempt::fail, which logs the bare reason (src/xrpld/overlay/detail/ConnectAttempt.cpp:136), ever matched — so the panel's Timeout series was connect-attempt timeouts only, Ping Timeout (PeerImp.cpp:762) was invisible, and PeerImp's own Closed was uncounted. The panel now matches all three prefixes ((?:\] |failed: |close: )) and distinguishes Ping Timeout from Connect Timeout. When adding a log-derived panel, enumerate every producer of the string being captured rather than sampling one.

Also worth knowing: the Grafana Cloud image renderer cannot query Loki in this stack. A minimal probe dashboard with a hardcoded datasource uid, a literal expression and no template variables still rendered "No data", while the identical expression returned 241 points through /api/ds/query. Verify LogQL panels with /api/ds/query per target, not with panel-image rendering.

Troubleshooting

No traces appearing in Tempo

1. Check xrpld logs for Telemetry starting message
2. Verify enabled=1 in the [telemetry] config section
3. Test collector connectivity: curl -v http://localhost:4318/v1/traces
4. Check collector logs: docker compose -f docker/telemetry/docker-compose.yml logs otel-collector
5. Verify Tempo is receiving data: open Grafana → Explore → select Tempo datasource → search by service.name = xrpld
6. Check Tempo logs: docker compose -f docker/telemetry/docker-compose.yml logs tempo

No system metrics in Prometheus

1. Check xrpld logs for OTelCollector starting message
2. Verify server=otel in the [insight] config section
3. Verify the endpoint in [insight] points to the OTLP/HTTP port (default: http://localhost:4318/v1/metrics)
4. Check that the otlp receiver is in the metrics pipeline receivers in otel-collector-config.yaml
5. Query Prometheus directly: curl 'http://localhost:9090/api/v1/query?query=jobq_job_count'

Server info gauge shows server_state=0

This is normal during startup. The server starts in DISCONNECTED mode (0) and progresses through CONNECTED (1), SYNCING (2), TRACKING (3), to FULL (4). Wait for the node to sync with the network.

Database metrics showing zero

The getKBUsed*() methods require SQLite databases to exist. If running with --standalone or before the first ledger is stored, these will be zero.

Slow TMGetObjectByHash service

Use this when a peer reports slow object fetches, or when job_queued_us / job_running_us for job_type="ledgerRequest" rises. The goal is to name the cause, not to confirm the slowness.

End-to-end handler time splits into three additive parts, and each has its own signal:

flowchart LR
    A["`**Request arrives**
    onMessage
    TMGetObjectByHash`"] --> B["`**1. Queue wait**
    job_queued_us
    handler=RcvGetObjByHash`"]
    B --> C["`**2. NodeStore lookup**
    getobject_lookup_us`"]
    C --> D["`**3. Everything else**
    protobuf, serialization,
    reply construction`"]
    D --> E["`**Reply sent**`"]

    C -.-> F["`job_running_us
    handler=RcvGetObjByHash
    covers steps 2 + 3`"]
    D -.-> F

    style A fill:#1f4e79,color:#ffffff
    style B fill:#7b3f00,color:#ffffff
    style C fill:#2d5016,color:#ffffff
    style D fill:#4a148c,color:#ffffff
    style E fill:#1f4e79,color:#ffffff
    style F fill:#37474f,color:#ffffff

Reading the diagram

• Step 1 is time the job sat in JtLedgerReq before a worker picked it up. Only job_queued_us measures it.
• Steps 2 and 3 both happen inside the worker, so job_running_us covers them together. getobject_lookup_us isolates step 2 alone.
• Step 3 is therefore not measured directly. Derive it: job_running_us − getobject_lookup_us. That subtraction is what makes the set able to name a cause instead of just reporting a duration.

Procedure — work through these in order. The first row that matches is the answer.

Scope every query to one node. All snippets below carry service_instance_id=~"$node". On a shared Grafana stack an unscoped selector aggregates across every node and branch reporting to it, so another node's saturation would be attributed to this one. Substitute the node's public key for $node when querying Prometheus directly rather than from a dashboard.

1. Split queue wait from run time. Compare the two p99s for the handler:

   histogram_quantile(0.99, sum by (le) (rate(job_queued_us_bucket{handler="RcvGetObjByHash", service_instance_id=~"$node"}[5m])))
   

   histogram_quantile(0.99, sum by (le) (rate(job_running_us_bucket{handler="RcvGetObjByHash", service_instance_id=~"$node"}[5m])))
   

2. If run time is the larger term, split it against the fetch loop:

   histogram_quantile(0.99, sum by (le) (rate(getobject_lookup_us_bucket{service_instance_id=~"$node"}[5m])))
   

3. Match the outcome below.

Observation	Root cause and next step
job_queued_us{handler="RcvGetObjByHash"} high, job_running_us normal	Queue contention — the work is cheap, the wait is not. Confirm with jobq_ledgerrequest_deferred{service_instance_id=~"$node"} > 0, then compare job_queued_us{handler="RcvGetLedger"}: if it is also high, both producers are starved by the limit of 3, not by each other.
job_running_us high and within ~10% of getobject_lookup_us	NodeStore is the bottleneck — nearly all run time is in the fetch loop. Check rate(getobject_lookups_total{result="miss"}[5m]) and the existing NuDB / nodestore_state panels. A miss-heavy mix means real disk seeks.
job_running_us high but getobject_lookup_us low	Cost is outside the fetch loop — protobuf, serialization, or reply construction. Storage is fine. Look at reply size: a large getobject_request_objects with a high hit rate means big replies to build and send.
getobject_request_objects p99 large	Peers are sending big batches — the work is real, not a regression. Nothing is broken; the node is being asked to do more. Decide whether to accept the load or price it higher.
rate(getobject_rejected_total{reason="oversize"}[5m]) rising	Non-conforming traffic — requests above kHardMaxReplyNodes are being refused before any NodeStore access. Check getobject_charge to confirm the pricing escalates for the requests that are accepted.
rate(getobject_rejected_total{reason="malformed_ledgerhash"}[5m]) rising	Malformed requests — a peer is sending a ledgerhash that is not 32 bytes. Refused at the gate; no queue or storage cost incurred.
All GetObject metrics normal, jobq_*_deferred high on another type	This path is exonerated — the slowness is elsewhere. Find the saturated type with topk(5, {__name__=~"jobq_.*_deferred", service_instance_id=~"$node"} > 0) and investigate that producer instead.

Row 2 says "within ~10%", not "equal", deliberately: job_running_us also covers the charge computation (PeerImp.cpp:2757) and the reply send() that follow the loop, so it is always the larger of the two. Treat a small residual as normal and only a large one as a signal — that is what row 3 is for.

The last row matters as much as the others: the set can rule this path out, which a slowness-only metric cannot.

Caveats

• handler collapses to "other" for any job name that is not all ASCII letters, so a handler breakdown is not exhaustive — see The handler Label. This does not affect RcvGetObjByHash or RcvGetLedger; both pass through as distinct series.
• jobq_*_deferred is sampled at 1 s. A zero reading does not prove no saturation occurred — see the sampling caveat under Per-Job-Type Queue Saturation.
• The two families compared in this procedure are exported on different cadences by separate meter providers: the jobq_* gauges every 1 s (Telemetry.cpp global provider) and job_queued_us / job_running_us / getobject_* every 10 s (the private MetricsRegistry provider). A short spike can therefore appear in one family a step before the other. When correlating them, widen the window rather than reading a single scrape, and do not conclude the two disagree from one interval's difference.
• A request rejected at either gate contributes to no other GetObject metric, so a rejection spike will not show up as latency. Check the rejection counters before concluding that traffic is normal.

Slow to reach full

Use this when a node takes far longer than expected to sync. Two completely different bottlenecks look identical from outside: in both, the ledgerData job lane sits pinned at its concurrency cap of 3 with jobs waiting behind it.

Lane occupancy on its own distinguishes nothing. It is true in both cases, so it is never a diagnosis. Two tuning experiments were spent before that was known — do not repeat them. Read the storage-side signals below instead.

The two modes and the signal that separates them:

flowchart TB
    L["`**ledgerData lane at cap 3**
    jobs waiting behind it
    TRUE IN BOTH MODES
    diagnoses nothing`"]

    L --> W["`**Mode W — write-bound**
    fresh or empty store`"]
    L --> R["`**Mode R — cold-read-bound**
    populated store, cold pages`"]

    W --> W1["`reads cheap and always miss
    data comes from peers`"]
    W1 --> W2["`cost is on the WRITE side
    one global mutex per insert
    so inserts queue`"]
    W2 --> W3["`**Look at:** writer depth
    above ~1.2, insert mean
    well above service time`"]

    R --> R1["`no write contention
    writer depth 1.00`"]
    R1 --> R2["`reads pay disk latency
    on every fetch —
    found but cold`"]
    R2 --> R3["`**Look at:** read mean
    several times a warm read,
    then the found rate to split
    cold-but-held from real misses`"]

    style L fill:#7b3f00,color:#ffffff
    style W fill:#1f4e79,color:#ffffff
    style R fill:#4a148c,color:#ffffff
    style W1 fill:#37474f,color:#ffffff
    style W2 fill:#37474f,color:#ffffff
    style W3 fill:#2d5016,color:#ffffff
    style R1 fill:#37474f,color:#ffffff
    style R2 fill:#37474f,color:#ffffff
    style R3 fill:#2d5016,color:#ffffff

node_reads_hit is a found count, not a cache-hit rate

This is the most misleading signal on the board, so read it first. fetchHitCount_ is incremented whenever the fetch returned an object (src/libxrpl/nodestore/Database.cpp:246-255) — not when a cache served it. So node_reads_hit / node_reads_total is the fraction of fetches that found something, and it can read ~100% while every one of those fetches went to disk.

A ~100% "hit rate" at over 100 µs per read is therefore not a contradiction. It is the cold-read signature: the data is on disk, found every time, and paid for every time.

An incident report supplied to this project describes a devnet client-handler node that had not reached full after roughly 25 minutes while reading at 112.7 µs per fetch, against an otherwise-identical peer that reached full in 4.4 minutes at 4.95 µs per fetch. Both reported a found rate of ~99.98%. Those figures come from that report, not from a run on our own hosts. Note that the found rate is identical on the healthy peer and the stalled one, which is exactly why the found rate is never a trigger on its own — see the decision rule below. Read this incident alongside Honest limits of this diagnosis before concluding that cold reads caused the 25 minutes; on our own hardware they did not produce anything like it.

A node configured with online_delete runs DatabaseRotatingImp, which has no NodeObject cache at all (0 cache_ references in src/libxrpl/nodestore/DatabaseRotatingImp.cpp versus 11 in DatabaseNodeImp.cpp), so every fetch reaches the backend. That is why the cold-read mode shows up on exactly those nodes.

The decision rule

Procedure — all of it assumes the ledgerData lane is already at its cap; if it is not, this procedure does not apply.

Answer two questions, in this order. Neither one alone is a diagnosis.

Question 1 — are reads expensive? Read read_mean_us.

• Under ~10 µs: reads are cheap. Both a clean store and a healthy populated store land here.
• Over ~20 µs: reads are expensive.
• Between the two: break the tie on the tail. Take max_over_time of read_mean_us across the window. Over ~100 µs counts as expensive; otherwise treat it as cheap. There is no read-max gauge — nudb_insert_max_us is a write signal and does not answer this.

Question 2 — is the write path queueing? Read nudb_writer_depth_x100 / 100. Above ~1.2 is queueing at the insert mutex. Depth of 1.00 is not. On a non-NuDB backend the series is absent, so this question has no answer and only question 1 applies.

Then read the answer off the pair:

Reads	Write path	Root cause and next step
Cheap	Depth ~1.00	Not a storage bottleneck. Nothing is queueing at either side. Look outside storage: work is either not arriving from peers or being discarded before it lands. Check acquire_ledger_deferrals against acquire_ledger_timeouts, and the sweep counter.
Cheap	Depth over ~1.2	Serialized write path. The queue is at the backend's insert mutex, not at the device. Tuning the disk will not help; see the note below. This is the clean-store mode.
Expensive	Depth ~1.00	Split on the found rate. At 50% or above, cold reads on data the node already has — the walk pays disk latency on objects it holds. Below 50%, genuinely disk-bound on real misses, and storage hardware is the right thing to change.
Expensive	Depth over ~1.2	Both paths queueing. Rarer, and neither fix on its own will be enough. Treat the larger of the two costs as the lead.

Why the rule is shaped this way. Three points about the thresholds, each learned from a dataset that an earlier version of this table got wrong:

• Read cost is a relative judgement, so the band has a floor and a ceiling, not one cut. A cold read on our box measured 31.8 µs mean; a cold read on the devnet node in the reported incident measured 112.7 µs. A single "over 100 µs" cut would call our own cold-read run healthy. Cheap and expensive are set at ~10 µs and ~20 µs with the peak breaking ties in between, because what matters is whether reads cost several times a warm read, not whether they cross one absolute number.
• The found rate is a splitter, never a trigger. A high found rate on its own is the normal, healthy state of a populated store — the healthy peer in the reported incident read 99.98% found at 4.95 µs per read and was fine. Only ask the found rate once reads are already known to be expensive; then it separates "slow on data we have" from "slow because we are missing".
• Writer depth answers before the found rate. Depth above 1 means the queue is at the insert mutex, which no amount of read-side tuning addresses. It is also the only signal that is unambiguous on a clean store, where the found rate is near zero and read cost is uninformative.

Completions and sweeps do not pick the row. They say how severe the read case is once the row is picked: a populated store with cold pages can still finish (our reference run reached full in 260 s with 8 completions) or can fail to finish at all, which is what the reported incident describes. Same cause, different severity.

Queries, scoped to one node as everywhere else in this runbook:

# Are acquisitions finishing? (per minute)
increase(nodestore_state{metric="acquire_completions", service_instance_id=~"$node"}[1m])

# Question 1 -- read cost, in microseconds per read
nodestore_state{metric="read_mean_us", service_instance_id=~"$node"}

# The tail, for the 10-20 us tie-break. No read-max gauge exists, so take the
# highest value the mean reached over the window.
max_over_time(nodestore_state{metric="read_mean_us", service_instance_id=~"$node"}[15m])

# Question 2 -- write-side queueing. Depth is fixed-point, divide by 100.
nodestore_state{metric="nudb_writer_depth_x100", service_instance_id=~"$node"} / 100

# The insert time that goes with that depth
nodestore_state{metric="nudb_insert_mean_us", service_instance_id=~"$node"}

# The found rate, which splits the expensive-read row only. Do not read it on
# its own: a high found rate is normal and healthy on a populated store.
  rate(nodestore_state{metric="node_reads_hit", service_instance_id=~"$node"}[5m])
/ rate(nodestore_state{metric="node_reads_total", service_instance_id=~"$node"}[5m])

# The deferral/timeout pair, scoped to ledger acquisition. Use these two, not
# the all-lane acquire_deferrals / acquire_timeouts -- see the pair section below.
increase(nodestore_state{metric="acquire_ledger_deferrals", service_instance_id=~"$node"}[5m])
increase(nodestore_state{metric="acquire_ledger_timeouts", service_instance_id=~"$node"}[5m])

Measured reference points

Provenance. The two columns below are our own measurements: node2 on the AWS dev box, build e3c2f8279a, 2026-07-27/28, same host and same binary for both runs, differing only in the state of the store. Use them as the shape to compare against, not as thresholds. The read figures below come from the read_mean_us gauge, the only read-latency signal exported; the "highest sample" row is the largest value that gauge reached over the run, not a read-latency percentile. The third dataset in this section — the 25-minute devnet stall and its healthy peer — is not ours; it comes from an incident report supplied to this project and is kept separate for that reason.

Three rows below were measured on a build that got them wrong. Both runs predate the measurement fixes, so read those rows as bounds rather than values:

• Completions were only counted in InboundLedger::done(), so an acquisition satisfied entirely from the local store — init() sets complete_ and returns without ever calling done() — was never counted. Mode W's 0 is therefore not evidence that the node completed nothing; it reached full, which it could not have done without completing acquisitions. The count is now taken at both exits behind an idempotent latch, so on a current build a zero means zero.
• Writer depth was summed at insert entry but divided by a sample count that only advanced at insert exit, so in-flight inserts — the deep, slow ones — contributed depth to the numerator and nothing to the denominator. The mean was biased down, worst exactly when queueing was worst. Mode W's 1.60 is a lower bound on the true depth.
• Queueing per insert is derived from that depth, so its 37 % is a lower bound too. See the derivation below.

Signal	Mode W: clean store	Mode R: populated store, cold pages
Time to full	510 s	260 s
read_mean_us	8.8 µs	31.8 µs
read_mean_us, highest sample	9 µs	223 µs
Found rate	0.00 %	88.3 %
Insert time, mean	20.0 µs	15.9 µs
Writer depth, mean	≥ 1.60 (biased low)	1.00
Queueing per insert	≥ 37 % (derived from ↑)	0 %
Deferrals over run, all lanes	+5441	+1845
Timeouts over run, all lanes	+687	+399
Completions over run	0 (under-counted, see ↑)	8 (under-counted, see ↑)
Sweep evictions	+127	+38

Applying the decision rule: Mode W reads cheap (8.8 µs) with depth 1.60, so it is the serialized write path. Mode R reads expensive (31.8 µs mean, 223 µs peak) with depth 1.00 and a found rate well above 50%, so it is cold reads on data the node holds. The rule reaches both answers without the found rate deciding either mode on its own — and both answers survive the corrected measurements, because a depth-1.60 lower bound is still above the 1.2 threshold and Mode R's 1.00 is a floor that cannot be biased below itself.

The deferral and timeout rows are the all-lane counters, the only ones that existed when these runs were taken. They cannot be attributed to ledger acquisition; the eight-to-one ratio in Mode W is a whole-node figure. Re-measure with acquire_ledger_deferrals / acquire_ledger_timeouts before quoting a ratio as a ledger-acquisition fingerprint.

The reported incident, for contrast — not our measurement. Figures from an incident report supplied to this project. No writer-depth data was captured, so the rule reaches its answer from question 1 alone.

Signal	Stalled client handler	Healthy peer
Time to full	not reached in ~25 min	4.4 min
read_mean_us	112.7 µs	4.95 µs
Found rate	~99.98 %	~99.98 %
Writer depth	not captured	not captured

The healthy peer is the reason the found rate is a splitter and not a trigger: it reported the same ~99.98% as the stalled node and was fine. What separates them is read cost — 4.95 µs is a warm read, 112.7 µs is not.

What healthy looks like: read mean in the single-digit microseconds, writer depth at 1.00, queueing near 0%, and acquire_completions advancing. Any one of a read mean several times a warm read, a writer depth above ~1.2, or completions flat at zero is worth chasing.

Completions flat at zero only means something on a current build. Until the counter was moved to cover both exits, an acquisition served from the local store was never counted, so a build predating that fix could read zero while completing steadily. Check the build before treating a zero on archived data as a symptom.

How the 37% is derived, and why it is a lower bound. It is not measured directly — it comes from the two gauges by Little's Law. With mean queue depth L and mean insert time W, the service time is S = W / L and the queueing component is W − S. Mode W's 20.0 µs at depth 1.60 gives S = 12.5 µs, so 7.5 µs of every insert — 37% — was spent waiting for the mutex rather than writing.

That 37% is not exact: the L it was computed from came from the biased estimator described above, which understated depth. A larger L gives a smaller S and a larger W − S, so the true queueing share of Mode W was at least 37%. Quote it as a floor. Mode R's depth of exactly 1.00 is unaffected — 1.00 is the minimum a depth can be, so no bias can have pushed it there — which is why its 0% stands as measured.

Why the write path serializes. NuDB takes one global mutex per insert (nudb/impl/basic_store.ipp:288). It is a Conan dependency and is not patched here, so this is a property to observe and design around, not a bug to fix locally. nudb_writer_depth_x100 is the queue length at that mutex.

The deferral/timeout pair

Read these two together or not at all. The livelock fingerprint is deferrals rising while timeouts stay flat.

Read the ledger-scoped pair, not the all-lane pair. Both events are recorded in TimeoutCounter, a base class shared by five subclasses — InboundLedger, TransactionAcquire, LedgerReplayTask, LedgerDeltaAcquire and SkipListAcquire — each with its own job limit. acquire_deferrals and acquire_timeouts therefore pool every lane, so a replay lane sitting at its own limit produces the fingerprint shape while ledger acquisition is perfectly healthy. That is a false positive on a headline diagnosis. acquire_ledger_deferrals and acquire_ledger_timeouts count the same two events for the InboundLedger lane only (src/xrpld/app/ledger/detail/TimeoutCounter.h, isLedgerAcquisition()), and they are the pair this procedure means. The all-lane totals remain useful for one question only: whether any lane is deferring.

A deferral happens when the acquisition timer job finds its lane's job count at or above the acquisition's own limit — 5 for InboundLedger (src/xrpld/app/ledger/detail/InboundLedger.cpp:86), compared against getJobCountTotal() in TimeoutCounter::queueJob() (src/xrpld/app/ledger/detail/TimeoutCounter.cpp:62-64). That is not the same as the ledgerData lane's concurrency cap of 3 (include/xrpl/core/JobTypes.h:63): the gate counts running plus queued, so it fires at 3 running plus 2 queued. The timer is re-armed but its body does not run, so the retry counter never advances and the 6-timeout give-up becomes unreachable — the give-up path is disarmed and the acquisition can never end on its own. Neither counter alone shows this: deferrals rising looks like ordinary backpressure, and timeouts flat looks like health. Only the divergence is diagnostic. See the counter documentation in src/xrpld/app/ledger/AcquireStats.h.

Two more pairs from the same family:

• acquire_sweep_evictions rising while acquire_completions stays at zero → partial work is being discarded and redone. The sweep drops any acquisition idle for more than one minute (src/xrpld/app/ledger/detail/InboundLedgers.cpp:400), taking whatever it had built with it.
• acquire_aborts_partial rising → the expensive form of an abort, where partly built maps were thrown away. acquire_aborts alone does not separate the cheap case from this one.

Honest limits of this diagnosis

• Our populated-store run was twice as fast, not slower — 260 s against 510 s, despite reads being roughly 4× more expensive. Reusing local data beats fetching from peers even when every read is cold. Slow cold reads therefore do not on their own explain the ~25-minute stall in the reported devnet incident. The decision rule identifies the mode correctly in both cases; it does not claim that the mode alone accounts for that duration.
• Something compounds it there, and we have not confirmed what. The most likely candidate is a much larger store, where the walk takes long enough that the 1-minute sweep destroys partial work faster than it can complete — which is why the sweep and completion counters are in the table. This is an unconfirmed hypothesis. Treat it as the next thing to test, not as the answer. It was also partly suggested by Mode W's zero completions, which we now know was a counting defect rather than a stalled node, so the hypothesis has lost one of its supports and needs re-testing on a current build before it is pursued.
• Three of the numbers above were measured with instruments that were since corrected: completions (missed local-store hits), writer depth (mean biased low), and the deferral/timeout pair (pooled across five job lanes). The modes and the decision rule are unaffected — each survives the correction, as noted where it appears — but no figure in the reference table should be quoted as an exact measurement without re-running on a build that has all three fixes.
• The nudb_* label values are absent entirely on a non-NuDB writable backend. Absent is not zero — a missing series means "not applicable", so a panel showing a gap there is correct behaviour.
• write_load and nudb_writers_in_flight are the same number on NuDB. Both read the same atomic: NuDBBackend::getWriteLoad() returns concurrentWriters (src/libxrpl/nodestore/backend/NuDBFactory.cpp:355-361), which is also what WriteStats::concurrentWriters reports. Their agreement confirms nothing — it is one signal plotted twice. On RocksDB write_load is a genuinely different quantity, the larger of the recorded load and the pending batch size (src/libxrpl/nodestore/BatchWriter.cpp:47-53), so it is a batch-queue length rather than a thread count.
• stored_object_bytes is not the size of the store on disk. It reports the cumulative object-payload bytes this process has written — the same value as node_written_bytes, from the same accessor — so it excludes keys, padding and the log, and it resets with the process. A ratio of the two is a constant 1.0 and measures nothing. This label value was called nudb_bytes in earlier revisions; it comes from node_store::Database rather than the NuDB backend, so it is not part of the nudb_* family above and reads the same on RocksDB.
• These gauges are sampled on the MetricsRegistry reader's 10 s cadence, while the jobq_* lane gauges are sampled at 1 s by a different provider. Widen the window when correlating them rather than reading a single scrape; see the caveat under Slow TMGetObjectByHash service.
• read_mean_us and write_mean_us are omitted rather than reported as zero when nothing has been read or written yet, so an idle node legitimately shows no series.

Existing panels that already carry part of this picture, on the Ledger Data & Sync dashboard: NuDB Read Latency, NuDB Read Found Ratio, NuDB Read Pressure, and Job Queue Backlog and Deferred by Type for the lane occupancy that this procedure tells you to distrust on its own.

The pair this procedure asks for is on Ledger Acquire Deferrals vs Timeouts (Ledger Lane Only), in the Sync Bottleneck Discrimination row. The adjacent Acquire Deferrals vs Timeouts (All Lanes) plots the pooled totals; read it only to see whether some other acquisition lane is also under pressure.

NuDB Read Found Ratio plots node_reads_hit / node_reads_total. That is the found rate, not a cache hit ratio: the underlying counter increments whenever a fetch returned an object, and a node with online_delete has no object cache at all. A rate near 1.0 alongside an expensive read time is the cold-read signature, not a sign the cache is working.

High memory usage

• Reduce trace volume with collector-side tail sampling (xrpld head sampling is fixed at 1.0 and is not configurable)
• Reduce max_queue_size and batch_size
• Disable high-volume trace categories: trace_peer=0

Collector connection failures

• Verify endpoint URL matches collector address
• Check firewall rules for ports 4317/4318
• If using TLS, verify certificate path with tls_ca_cert

No trace_id in log output

• Verify xrpld was built with telemetry=ON (the XRPL_ENABLE_TELEMETRY preprocessor flag)
• Verify enabled=1 in the [telemetry] config section
• Log lines only contain trace_id/span_id when emitted inside an active span — background logs outside of RPC/consensus/transaction processing will not have trace context
• Check that the specific trace category is enabled (e.g., trace_rpc=1)

No logs in Loki

• Verify the log file mount in docker-compose.yml points to the correct xrpld log directory (default source docker/telemetry/data/logs, or the XRPLD_LOG_DIR override) and that xrpld actually writes debug.log there
• Check OTel Collector logs for filelog receiver errors: docker compose logs otel-collector
• Verify Loki is running: curl http://localhost:3100/ready
• Check the filelog receiver glob /var/log/xrpld/*/debug.log matches your log layout — the log file must sit one subdirectory below the mount root

Performance Tuning

Scenario	Recommendation
Production mainnet	trace_peer=0; reduce volume via collector tail sampling
Testnet/devnet	Full tracing (head sampling fixed at 1.0)
Debugging specific issue	Full tracing (head sampling fixed at 1.0)
High-throughput node	Increase batch_size=1024, max_queue_size=4096

Disabling Telemetry

Set enabled=0 in the [telemetry] config section (runtime disable, no rebuild), or compile telemetry out:

conan install .. --output-folder . --build missing -o telemetry=False --settings build_type=Release
cmake -DCMAKE_TOOLCHAIN_FILE:FILEPATH=build/generators/conan_toolchain.cmake -DCMAKE_BUILD_TYPE=Release -Dtelemetry=OFF ..

Pass the flag explicitly rather than omitting it — an omitted flag resolves to whatever the build's current default is. That default is ON on the telemetry branches so CI compiles the instrumented paths, and OFF once the feature is merged; -Dtelemetry=OFF is correct either way. -DXRPL_ENABLE_TELEMETRY=OFF does not work: that name is only a compile definition added when telemetry is ON, not a CMake option, so telemetry stays compiled in and CMake only lists it under Manually-specified variables were not used by the project.

When telemetry is compiled out, all trace macros expand to no-ops with zero overhead.

Validating Telemetry Stack

After deploying telemetry, use the workload tools in docker/telemetry/workload/ to validate the full stack end-to-end.

Quick Validation

# Run the full validation suite (starts cluster, generates load, validates):
docker/telemetry/workload/run-full-validation.sh --xrpld .build/xrpld

# Check the report:
cat /tmp/xrpld-validation/reports/validation-report.json | jq '.summary'

# Tear the stack and the node processes down:
docker/telemetry/workload/run-full-validation.sh --cleanup

Harness options (run-full-validation.sh):

Flag	Default	Effect
--xrpld PATH	.build/xrpld	Binary to run. Also settable via the XRPLD env var.
--nodes NUM	5	Size of the local validator cluster.
--profile NAME	full-validation	Load profile from workload-profiles.json (full-validation, quick-smoke, stress). This is the only thing that sets load shape.
--skip-loki	off	Skip the log-trace correlation checks. CI always passes this.
--skip-regression	off	Skip timing capture and the baseline comparison. Local exploration only.
--with-benchmark	off	Also run benchmark.sh (telemetry-off vs telemetry-on overhead) after validation.
--cleanup	—	Tear everything down and exit.

--rpc-rate, --rpc-duration, --tx-tps and --tx-duration are accepted by the parser but never read — they predate profiles and have no effect. Use --profile, or add a profile to workload-profiles.json.

Exit codes: 0 all checks and the regression gate passed; 1 a validation check failed or the gate detected a regression; 2 infrastructure error (stack or cluster did not come up, or timing capture failed).

What Gets Validated

The counts are not hard-coded in the validator — it iterates the inventory files, so those files are authoritative. The figures below are the inventory as it stands today.

Category	Checks	Description
Spans	Every required entry in expected_spans.json — 41 span types at the time of writing: 26 required, 15 marked "optional": true	Span name found in Tempo carrying its required_attributes, plus the declared parent-child relationships. An "optional": true entry that does not fire is recorded as a skip, not a failure — it needs traffic the harness may not generate (HTTP/JSON-RPC client, gRPC client, missing-ledger fetch, mode transitions).
Metrics	Every entry in every asserted category of expected_metrics.json — 58 metrics in 23 categories at the time of writing	SpanMetrics, beast::insight gauges/counters exported over OTLP, and the MetricsRegistry OTLP metrics. Each must have > 0 Prometheus series; none are optional. The separate not_asserted group lists metrics deliberately left out of the gate because they are workload-gated or defect-gated; it has no metrics key, so the validator skips it.
Logs	2 checks	trace_id/span_id present in Loki, and a Tempo trace id resolves in Loki. Skipped in CI, which runs --skip-loki.
Parity	10 checks	6 span attributes the external-parity dashboard panels read, plus 4 metric value-sanity bounds.
Dashboards	Every uid in expected_metrics.json under grafana_dashboards.uids — currently all 15 provisioned dashboards	Each listed dashboard loads and reports a panel count. This is a provisioning check only: it does not execute the panels' queries, so a dashboard can pass while individual panels render empty. log-derived-insights is Loki-backed, so under --skip-loki only its provisioning is meaningfully covered.

Running Individual Tools

# RPC load only:
python3 docker/telemetry/workload/rpc_load_generator.py \
    --endpoints ws://localhost:6006 --rate 50 --duration 120

# Transaction mix only:
python3 docker/telemetry/workload/tx_submitter.py \
    --endpoint ws://localhost:6006 --tps 5 --duration 120

# Validation only (assumes load already ran):
python3 docker/telemetry/workload/validate_telemetry.py \
    --report /tmp/report.json

Interpreting Failures

• Span failures: Check that the relevant trace category is enabled in [telemetry] config (e.g., trace_rpc=1).
• Metric failures: Verify the OTel Collector is running and Prometheus is scraping port 8889.
• Dashboard failures: Ensure Grafana provisioning is mounted correctly.

run-full-validation.sh brings the stack up with docker compose -f docker/telemetry/docker-compose.workload.yaml, so a bare docker compose logs from the repository root finds no project. Pass the same compose file:

docker compose -f docker/telemetry/docker-compose.workload.yaml logs otel-collector
docker compose -f docker/telemetry/docker-compose.workload.yaml logs grafana
docker compose -f docker/telemetry/docker-compose.workload.yaml ps

Regression Gate and CI

The validation checks answer "is the telemetry there?". A second, independent gate answers "did xrpld get slower?" — it is the part of this harness that can fail CI on a performance change, so it is worth understanding before you push.

It runs as step 6 of run-full-validation.sh, after validation, and is skipped only with --skip-regression:

flowchart TB
    classDef stage fill:#1d4ed8,stroke:#1e3a8a,color:#fff;
    classDef data  fill:#047857,stroke:#064e3b,color:#fff;
    classDef gate  fill:#b45309,stroke:#7c2d12,color:#fff;
    classDef out   fill:#334155,stroke:#0f172a,color:#fff;

    PROM[("Prometheus<br/>localhost:9090")]:::data
    MET["regression-metrics.json<br/>(spans + job_queue groups)"]:::data
    CAP["capture_timings.py<br/>--window REGRESSION_WINDOW"]:::stage
    TIM["reports/timings.json<br/>(key to value + unit)"]:::data
    BASE["baselines/baseline-timings.json<br/>(committed)"]:::data
    THR["regression-thresholds.json<br/>(pct AND abs bounds)"]:::data
    CMP["compare_to_baseline.py"]:::stage
    PH{"baseline is a placeholder<br/>or has no metrics?"}:::gate
    PASTE["Print paste-me JSON<br/>exit 0 — gate does NOT run"]:::out
    DIFF["Diff per metric<br/>regression = over BOTH bounds"]:::gate
    REP["reports/regression-report.json<br/>exit 1 on any regression"]:::out

    MET --> CAP
    PROM --> CAP --> TIM --> CMP
    BASE --> CMP
    THR --> CMP
    CMP --> PH
    PH -->|yes| PASTE
    PH -->|no| DIFF --> REP

Key properties:

• A metric regresses only when it exceeds BOTH the percentage and the absolute bound. The AND is deliberate: SpanMetrics latency histograms use explicit buckets, so a quantile sitting near a low bucket boundary can jump a whole bucket (1 ms to 5 ms) with no real change. Bounds live in regression-thresholds.json — defaults per category and quantile, with per-metric overrides (e.g. span.consensus.ledger_close is held to 5%).
• A metric with no configured threshold is captured but never gates. It is reported with a note instead. Today only span.* and job.* keys have thresholds; rpc.* is not produced and would not gate if it were (see docker/telemetry/workload/baselines/README.md).
• A metric missing from the current run is not a regression. summary.missing_in_current in regression-report.json is a count; the identities are the metrics[] entries whose note is "not captured in current run".
• REGRESSION_WINDOW (env var, default 3m) is the window handed to Prometheus rate() during capture. Keep it close to the workload duration — a longer window dilutes a short-lived regression. BASELINE_FILE, THRESHOLDS_FILE and METRICS_FILE are also env-overridable.

# Validation without the gate (fast local loop):
docker/telemetry/workload/run-full-validation.sh --xrpld .build/xrpld \
    --profile quick-smoke --skip-loki --skip-regression

# Narrow the rate window to a short profile:
REGRESSION_WINDOW=1m docker/telemetry/workload/run-full-validation.sh \
    --xrpld .build/xrpld --profile quick-smoke

# Inspect the gate's own output:
jq '.summary' /tmp/xrpld-validation/reports/regression-report.json
jq -r '.metrics[] | select(.regressed) | "\(.key) \(.baseline) -> \(.current) \(.unit)"' \
    /tmp/xrpld-validation/reports/regression-report.json

Refreshing the baseline

The baseline is a committed file, and moving it is a reviewed change — that PR review is the audit point for "who moved the performance bar". There is no automatic promotion from develop.

1. Run the Telemetry Validation workflow on the branch. It always captures timings, so timings.json is uploaded as an artifact and the regression summary is written to the run's Step Summary.
2. If the baseline in the checkout is a placeholder ("placeholder": true or an empty metrics object), the Step Summary contains a fenced JSON block under "Paste into baselines/baseline-timings.json", already formatted the way the file expects (sorted keys, 2-space indent, trailing newline).
3. Open a PR replacing the file contents with that block, dropping the placeholder key. For a refresh of an already-populated baseline, take the timings.json artifact instead and justify the delta in the PR description.

Never hand-edit baseline-timings.json — every entry should trace back to a real CI run so its variance characteristics are preserved. Details in docker/telemetry/workload/baselines/README.md.

CI workflow

.github/workflows/telemetry-validation.yml runs three jobs — linux-image-tag (reads the CI image tag from the build matrix so this workflow cannot drift onto a different compiler than the main CI), build-xrpld (self-hosted runner, same container as the main CI, so Conan and ccache hit the shared caches), and validate-telemetry (ubuntu-latest, which has Docker).

• Triggers: workflow_dispatch, and push on pratik/otel-phase*, feature/otel-*, feature/telemetry-* limited to a paths filter covering the workflow file, docker/telemetry/**, and the telemetry sources under include/xrpl/telemetry/**, src/libxrpl/telemetry/** and src/xrpld/telemetry/**. There is no cron schedule.
• Invocation: run-full-validation.sh --xrpld <binary> --skip-loki, so the default full-validation profile is used and the Loki checks are skipped.
• Inputs: only run_benchmark changes behaviour. rpc_rate, rpc_duration, tx_tps and tx_duration are inert, as noted in their descriptions.
• Results: reports are uploaded as the telemetry-validation-reports artifact and node logs as xrpld-node-logs when validation did not succeed. Summaries go to the run's Step Summary; the workflow does not comment on PRs.

Performance Benchmarking

Measure the overhead of the telemetry stack against a baseline:

docker/telemetry/workload/benchmark.sh --xrpld .build/xrpld --duration 300

Benchmark Thresholds

Metric	Target	Description
CPU overhead	< 3%	Average CPU increase across nodes
Memory overhead	< 5MB	Peak RSS increase per node
RPC p99 latency	< 2ms	Additional p99 latency for server_info
Throughput impact	< 5%	Reduction in ledger close rate
Consensus impact	< 1%	Increase in consensus round time

Tuning for Production

If benchmarks exceed thresholds:

1. Reduce trace volume with collector-side tail sampling. There is no sampling_ratio config key — xrpld's head sampling is a compile-time constant fixed at 1.0 (Telemetry.h:234 static constexpr double samplingRatio = 1.0;), and TelemetryConfig.cpp:139 explicitly parses nothing for it. Volume reduction is a collector decision. The only policy shipped is a single 0.5% probabilistic tail_sampling processor in otel-collector-config.grafanacloud.yaml; the base otel-collector-config.yaml has no tail sampling, so a stock local stack keeps every trace. Where the Cloud policy is in force it sits on the trace-storage branch only — spanmetrics runs on a separate branch and still sees 100% of spans, so the derived RED metrics stay exact.
2. Disable peer tracing: trace_peer=0 (highest volume category)
3. Increase batch delay: batch_delay_ms=10000 (less frequent exports)
4. Reduce queue size: max_queue_size=1024 (back-pressure earlier)

See docker/telemetry/workload/README.md for full documentation.