Conflict resolution, OpenTelemetryPlan/09-data-collection-reference.md
Known Issues table: kept phase-9's expanded 17-row table rather than the
incoming 4-row version, and re-applied the incoming correction to the one
row it changed. The other three incoming rows are already present in
phase-9's table, and phase-9's version is more current (jobq_job_count,
not job_count).
The rpc_requests row now says [insight] server=otel; server=statsd
reaches a collector with no statsd receiver on this branch. The section 8
Quick Reference and the fallback caveat auto-merged cleanly.
09-data-collection-reference.md contradicted itself. Section 2 presents
server=otel as the recommended transport with StatsD as a fallback, but
the Known Issues table and both Configuration Quick Reference examples
still prescribed server=statsd, which on this branch reaches a collector
with no statsd receiver and an unpublished 8125/udp.
Switch the Known Issues row and the Minimal and Production examples to
server=otel with the OTLP metrics endpoint. Keep the labelled fallback
block, and state what it actually requires: re-adding the statsd
receiver and republishing the port. Also record that StatsDCollector
applies prefix to metric names while OTelCollector does not, so the two
transports do not produce the same series.
The phase-6 copy is left alone; server=statsd is correct there.
check-otel-naming fails Rule D on this branch: ledger-data-sync.json
aggregates by xrpl_work_item, which nothing in-tree emits because
perf-iac's alloy pipeline stamps it, so it cannot be derived from
*SpanNames.h. The sibling perf-iac identities xrpl_branch and
xrpl_node_role are already allowlisted; this one was missed when the
by() clauses were reintroduced.
Add it to EXTERNAL_INFRA_LABELS alongside them. Verified by removing
the entry again, which restores the failure.
Conflict resolutions:
- docker/telemetry/xrpld-telemetry.cfg: relocation conflict. phase-9 had
already moved [insight] to the end of the file with server=otel, so the
incoming block was dropped rather than inserted. Keeping both would have
produced two [insight] sections, which merge last-wins into a single
effective section, silently reviving the bug this branch just fixed.
phase-9's per-branch service_instance_id=xrpld-devnet is preserved.
- OpenTelemetryPlan/06-implementation-phases.md: kept both corrections.
phase-9's "Tempo" is right (no Jaeger anywhere in the stack) and
phase-8's "active, sampled span" is right: Log.cpp:328 injects only
when spanCtx.IsValid() && spanCtx.IsSampled().
- OpenTelemetryPlan/09-data-collection-reference.md and
docs/telemetry-runbook.md: kept phase-9's structured-metadata LogQL.
The collector's filelog regex_parser already extracts partition,
severity, trace_id and span_id, so phase-8's inline regexp forms are
redundant, and a line filter matches the literal text in a message body.
The generated node config carried two [insight] blocks. Duplicate ini
sections do not replace one another: parseIniFile emplaces the section
name (a no-op when it already exists) and appends the lines to the same
vector, then Section::append writes each key with insert_or_assign. The
effective section was therefore server=statsd with the first block's
endpoint and service_instance_id surviving but unused.
CollectorManager selects StatsDCollector for that value, so the nodes
emitted beast::insight metrics over UDP to 8125, which has no receiver
in the collector pipeline and no published port. The script's own check
asserts that 8125 is not listening, and its insight metric assertions
fail on zero series.
Keep only the server=otel block so the config matches what the script
verifies.
This branch removes the collector's StatsD receiver and un-publishes
8125/udp, but xrpld-telemetry.cfg still selected server=statsd, so the
sample config sent beast::insight metrics over UDP to a port nothing
listens on. Phase7_taskList.md:132 lists this switch as required work.
Select server=otel and replace address= with the OTLP metrics endpoint.
Document that endpoint and prefix are informational only, since
OTelCollector records on the global MeterProvider that [telemetry]
configures and formatName() applies no prefix, and note that beast
instruments are not exported yet because the collector is constructed
before the MeterProvider is registered.
Six findings from the review of #6494 survived independent verification.
Each was checked against the branch tip, and where behaviour was in
question, against a live collector and Loki rather than from the
reviewer's claim or from documentation alone.
Plan-doc section numbering. 06-implementation-phases.md used "## 6.9"
twice: for the new Phase 8 section and for the pre-existing Risk
Assessment. Three references already pointed at 6.8.1 and none at 6.9,
and the later phases are numbered 6.8.2 through 6.8.4, so Phase 8
becomes 6.8.1 and the sequence is monotonic. Renumbering to 6.10, as
suggested on the PR, would have collided with Success Metrics.
filelog read position. The receiver relied on the upstream default
start_at=end, which skips everything a node wrote before the first poll
and reads nothing at all from a log that has stopped being written to.
Read from the beginning instead, paired with a file_storage extension so
a restart resumes at the last offset rather than re-ingesting the file.
The collector image runs as 10001:10001 and ships no writable directory,
and a fresh named volume is root-owned, so a one-shot init service
prepares the volume first. It reuses an image the stack already pulls,
adding no new dependency.
Loki log stream label. The job resource attribute did not become a Loki
index label, so the documented {job="xrpld"} queries matched nothing.
Verified against grafana/loki:3.4.2 with its default config: only
service_name and deployment_environment are indexed, and job arrives as
structured metadata, which a stream selector cannot match. Dropped the
attribute and moved the twelve queries this branch introduced to
{service_name="xrpld"}. Three further occurrences in
07-observability-backends.md originate on the phase-1a branch and are
left for a commit there.
Trace ids on unsampled spans. Logs::format emitted trace_id and span_id
whenever the span context was valid. A span dropped by the
ParentBasedSampler still carries its parent's ids, so log lines
advertised traces that were never exported and the log-to-trace link
resolved to nothing. Require the sampled flag as well, and correct the
task list and the documentation that promised the fields unconditionally.
The remaining two findings were refuted. The reported risk of signing
material reaching Loki does not hold: Logs::format already scrubs seven
sensitive fields, and there is a single write path to the log file, so
every JLOG site is covered. The suggestion to add internalLink to the
Loki derived field is not applicable, because that key is not part of
Grafana's schema.
Both the CMake option and the Conan recipe default telemetry on, so the
build worked, but this workflow exists to exercise telemetry and should not
depend on a default it does not control. If that default ever flipped, the
binary would build cleanly and then every span and metric assertion would
fail for a reason no log names.
Stated explicitly, the failure mode inverts: CMakeLists.txt does
find_package(opentelemetry-cpp CONFIG REQUIRED) under this option, so a
dependency graph without it fails at configure time with a clear message.
The shared build-deps action is left alone deliberately — forcing the Conan
option there would change every other workflow that uses it.
The setting was flagged in review as unauthenticated admin access. It is
deliberate, and it matches the sibling stack in docker-compose.yml, which
carries the same two variables and the same published port with its intent
in comments. This copy had none, so the reasoning lived only in a review
thread and was rediscovered as a finding each time the file was looked at.
Viewer would break the harness rather than harden it: the validation suite
drives the Grafana API against this instance to confirm each dashboard
provisions and loads, and the dashboards and datasources come from the
read-only mounts on the same service.
Two defects reported against the harness, both confirmed.
The TPS field was computed with `bc` at scale=2, and bc omits the leading
zero: it prints ".25", not "0.25". A bare ".25" is not valid JSON, and this
was the normal case rather than an edge case — ledgers close every few
seconds, so ledger-advance over elapsed-seconds is well under 1 for any
realistic window. It survived earlier checks because those piped the file
through jq, which accepts the malformed form; Python's json rejects the whole
file. awk's %.2f always pads, so the field is now produced with awk. Audited
the other numeric fields at the same time: CPU average and memory peak
already used awk, and the p99, sample count and consensus mean are integers,
so TPS was the only one affected.
Separately, a failing attribute fetch was reported under the span's own check
name, which had already recorded the trace as found. That produced two
entries for one name, one passing and one failing, inflating the check total
and blaming the trace-existence check for a failure in a later network call.
The fetch now carries its own error handling and reports under
`span.attrs.<span>`, matching where its successful counterpart reports. It
moved into a helper rather than growing `validate_spans`, which was already
well over the line limit.
Three consecutive validation runs timed out at Step 3 with nodes stuck at
"unreachable", and the reason was not recoverable from the logs. The node
logs showed the failing nodes stopping at an identical point, immediately
after JobQueue initialisation and before the debug log is opened, with no
error text at all. The harness knew each node's pid and never used it, so a
crashed node was indistinguishable from a slow one.
The readiness loop now checks whether each node process is still alive and
fails as soon as one is not, instead of waiting out the remaining window and
burying the cause under two minutes of progress output. Liveness is not a
bare `kill -0`: an exited-but-unreaped child keeps its pid, so a zombie
answers `kill -0` and reads as alive for the whole window, which is exactly
how a crashed node came to look like a slow one.
On failure each stopped node reports its wait status and the tail of its
stdout. The status is the discriminator that was missing: 137 for a SIGKILL,
139 for a segfault, 134 for an abort, anything below 128 for a deliberate
exit. stdout is printed inline rather than left to the artifact upload,
because a node that dies before its debug log opens writes nothing else and
a cancelled run uploads nothing at all.
This is instrumentation, not a fix. The failure is not attributable to the
recent changes on this branch: the first red run touched only the two Python
files used at Steps 4 and 5, both of which run after this gate, and the same
harness passed 5/5 twice before that.
Run with no arguments the script iterated an empty list, found no
violations and printed "OK: 0 dashboard(s) passed" with exit 0 -- a clean
bill of health for reading no files, indistinguishable from a real pass.
A bare run now defaults to every dashboard beside the script, and a run
that still ends up with nothing to check exits 2 rather than reporting
success. Passing paths explicitly behaves as before.
Single-value panels carry the filtered network type in their title because
they have no legend to put it in; multi-series panels carry it in
legendFormat instead. This gauge was the only single-value panel across the
fifteen dashboards without it, so its reading was ambiguous once more than
one network type was in scope.
A validation run timed out at Step 3 with only 4 of 5 nodes proposing, and
the reason was unrecoverable afterwards. Two gaps caused that.
The node-log artifact collected `node*/debug.log` but not `node*/stdout.log`.
A node that dies before its log sink opens never writes a debug.log at all,
so stdout is the only place its reason survives — and that file is written by
the harness and read by nothing, so it went to the runner and was discarded.
The failing node's log was simply absent from the artifact.
The readiness loop also fetched each node's `server_state` and threw it away,
reporting only a count. "4/5 nodes proposing" says a node is missing but not
which one, so there is nothing to grep for even once the logs are kept. The
timeout now names each node that is not proposing along with the state it
last reported, distinguishing a node that answered with a non-proposing
state from one whose RPC port did not answer at all.
Neither change affects a healthy run: the accumulator resets each attempt and
stays empty while every node is proposing.
The alerting example env file, the contact-point provisioning header and one
runbook line still pointed at a gitignored helper script and at a rollout
phase number, neither of which ships. The contact-point header now states the
policy-tree warning inline rather than deferring to a file the reader cannot
open.
Four files conflicted.
xrpl.proto and TraceContextPropagator.h: phase-9's note is a semantic
superset of phase-8's and already carries no plan-folder pointer, so
phase-9's text is kept and nothing phase-8 said is lost.
node-health.json: the only change phase-8 makes to this file is renaming a
row that phase-9 had already deleted, so phase-9's file stands. Panel count
holds at 59 and both Validated Ledger Seq panels survive.
TESTING.md: resolved per hunk rather than by side. Phase-9's replacement of
the drifted span table and its new Grafana Cloud section are kept, and
phase-8's de-numbered "Test 3: Log-Trace Correlation" heading is taken --
keeping phase-9 wholesale would have reinstated the phase number that
phase-8 removed.
The runbook pointed at a planning document and a helper script that are
outside the shipped tree, and at a dashboard script that no longer exists,
so an operator following it hit three dead ends. The Cloud alert workflow
is now described by what it does to the tracked rules.yaml.
The integration-test sections, alert provisioning headers and a naming-check
test comment were indexed by rollout phase and task numbers defined only in
planning documents outside the shipped tree. Each now names what it covers.
These comments were indexed by task, use-case and limitation numbers that
are defined only in planning documents outside the shipped tree. Nothing
in the repository defined them, so the cross-references resolved nowhere.
Each comment now states what the code does.
Two unrelated edits had crept in. The added `.claude/` line duplicates
`/.claude`, which the same file has already carried since well before this
branch, so it never changed behaviour. Removing the root `.env.*` rule did
change behaviour: contributors lost repo-level cover for any env file
outside docker/telemetry, and the local ignore that masks this on one
machine is not part of the repo.
Restore the rule and drop the duplicate, which leaves the root gitignore
byte-identical to the parent branch.
The scoped rules added under docker/telemetry stay: they keep the three
tracked .env.*.example files visible and stop the generated Grafana Cloud
dashboards being committed, which is telemetry work this branch owns.
Two defects reported against the validation harness. Both premises were
correct, but neither suggested fix was, so the remedies differ.
Dashboard panel count: `len(dashboard["panels"])` treated Grafana row
objects as panels and skipped the panels nested inside collapsed rows, so
every dashboard was over-reported by between 1 and 10 (`log-derived-insights`
read 41 against a true 31). The check also passed unconditionally on HTTP
200, so a dashboard that renders nothing would still pass. `_leaf_panel_count`
now walks row children and the result gates the verdict. Gating on the old
top-level length, as suggested, would not have caught the case it was aimed
at: a dashboard made only of collapsed rows counts its rows and reports a
positive number while rendering nothing.
RPC latency percentiles: `LoadStats.record` appended a latency for every
outcome, including requests that never got a reply, where the value is a
time-to-failure rather than a round trip. A timeout contributed the full
receive timeout, and at the error rate a real run shows this reported p95 and
p99 of 10000 ms where the true figure was 5 ms. `record` now takes an
optional latency and the timeout path passes none. The suggestion to append
only on success was not adopted: a reply carrying `status: error` is a
completed, timely round trip whose latency is a genuine measurement, and
discarding it would throw away real data. `per_command` is now keyed off the
request counts rather than the latency map, so a command whose every request
timed out still appears in the report instead of vanishing from it, and each
entry carries a `latency_samples` count.
The runbook told operators to read a missing outcome as "never went to the
network", which stopped being true once the abort path started setting it,
and the glossary still described outcome as a two-value split.
Document all three values, where each is written, and why peer_count is
absent on the abort path. Several claims were wrong and are corrected:
- Give-up is reached at about 18s, not 21s. There is no setTimer call, so
the first timer runs immediately and the old derivation counted a wait
that does not happen.
- A live aborted rate does not imply stalled acquisitions. A clean
shutdown clears every in-flight acquire, the admin fetch_info clear does
the same, and a full job lane stops timeouts advancing so give-up cannot
fire. The runbook already said the last of these elsewhere.
- A missing outcome does not mean exactly one thing. tryDB can set failed_
and return before done() runs, exporting a span with no outcome at all.
- failed covers unusable ledger data as well as exhausting the retry
limit, so a failed span can carry timeouts=0.
- The aborted lower bound of one minute holds only on the sweep path.
- The sweep measures time since anything last asked for the ledger, not
since data last arrived.
Also fixes the mainnet verification command, which still used the old RPC
port, and drops an inaccurate claim from the config comment about which
ports the workload scripts use.
The mainnet and devnet telemetry configs are host processes sharing one
network namespace and working directory, and they agreed on every port and
both database paths. Running them together left the second unable to bind,
and running them in sequence pointed a mainnet node at a store holding the
other network's ledgers, silently and with no error.
Offset the mainnet ports by ten and name both database paths after the
network, matching what debug_logfile already did.
Also narrow the attack surface this config carried. [port_ws_public] has
no admin key, so every caller on it is a guest, and it was bound to all
interfaces on a node following Mainnet; it is now loopback, and nothing in
the repo connects to that port. [signing_support] is removed: it only
affects non-admin callers, the admin ports already grant signing to local
callers, and upstream deprecates the commands it exposes. Verified against
a running node: signing still succeeds on the admin port and is refused as
notSupported on the public one.
Every aggregation that filtered on xrpl_work_item left it out of its
sum by() grouping, so PromQL dropped the label and the xrpl_ident legend
these panels build from it rendered without the work item. 217 clauses
across the touched boards.
Perf-iac gives each work item its own set of nodes, so service_instance_id
already separates the runs and the visible effect is the legend rather
than merged series. Four clauses are the exception and aggregate across
nodes, where the grouping does real work: Convergence, Lag Behind Network
Tip and Build Version now measure spread, lag and version distribution
within one work item instead of smearing two independent clusters
together. Lag Behind Network Tip also needed its on() join key extended to
match, or two coexisting work items would fail the query outright.
The five heatmap inner by (le) clauses are deliberately untouched: an
extra grouping label would superimpose several distributions.
Also in this change:
- transaction-overview gains per-type-per-stage rate and failure-rate
panels, so all three requested dimensions exist rather than latency
alone, and its stage-rate panel picks up the filters and legend its
siblings already had.
- The acquire-duration panel splits by outcome. An aborted acquisition is
open until the sweep reaps it, so mixing it into one percentile
conflated fetch latency with time-to-abandon.
- Ledger Acquire Rate by Outcome gets the same Title Case value mapping as
its neighbour, and node-health gains the $outcome variable the two
panels now need.
LedgerReq Wait by Handler and NodeStore Read Latency (Bottleneck
Discriminator) filter on xrpl_work_item but left it out of their sum by()
grouping, so PromQL dropped the label and the xrpl_ident legend lost that
segment.
These two targets took their current form on this branch, so they are
fixed here. The other nine in this file predate it and were fixed on
pratik/otel-phase7-native-metrics, which merges forward.
The discriminator panel divides two aggregations; both sides carry the
label so their label sets stay equal for vector matching.
Nine targets across four panels filtered on xrpl_work_item while leaving
it out of their sum by() grouping. PromQL keeps only the labels listed in
by(), so the label was dropped from the result and the xrpl_ident legend
these panels build from it lost that segment.
Perf-iac runs stamp xrpl_work_item and give each work item its own set of
nodes, so service_instance_id already separates the runs; the visible
effect is the legend rather than merged series. Adding the label makes the
work item identifiable, which is the comparison these panels exist for.
NuDB Read Latency and NuDB Read Found Ratio each divide two aggregations.
Both sides get the label so their label sets stay equal and vector
matching still works.
Two comment-only notes.
MetricsRegistry.cpp: record why the app and overlay includes exist and
what they cost. They are what makes loops.txt carry
xrpld.app <-> xrpld.telemetry and xrpld.overlay <-> xrpld.telemetry, where
ordering.txt previously had telemetry strictly below both. The observable
gauges are pull-model, so their callbacks need the concrete types to
sample live state. The cycle is confined to this translation unit: no
telemetry header includes app or overlay, and all of src/xrpld builds
into one target, so there is no header or link cycle. Inverting it needs
a metrics-source interface below overlay, which is left as follow-up.
Also note loops.txt is generated and must never be hand-edited.
docker-compose.yml: state the Compose >= 2.24.0 floor. The grafana
service uses the long-form env_file mapping, which older Compose cannot
parse, and it fails for the whole file rather than that one service. The
long form is required because .env.alerting is gitignored and absent in a
fresh clone, and the short form treats a missing env file as an error.
The test carried both <xrpl/proto/xrpl.pb.h> and the bare <xrpl.pb.h>.
Both resolve to the same generated header, because the proto helper puts
the generated tree and its prefixed subdirectory on the target, so the
second include expanded to nothing behind the header guard.
The bare form arrived from merging two same-day clang-tidy commits that
added the include with different spellings. Keep the prefixed spelling,
which is what the telemetry headers and the upstream phase branches use.
An InboundLedger destroyed while !isDone() recorded recordAbort() in the
metrics but left ledger.acquire carrying only the attributes set at
construction, so an abandoned acquisition was indistinguishable from one
still in flight.
Set outcome=aborted plus timeouts on that path, and add the val::aborted
constant. peer_count is deliberately omitted: reading it goes through
Overlay, and a destructor must not depend on Overlay still existing.
The status stays Unset for an abort, because InboundLedgers::stop()
clears every in-flight acquisition, so a clean shutdown would otherwise
report errors. done() now sets Error when failed_, which is an
unambiguous failure of the operation. Success is left Unset rather than
Ok, per the OpenTelemetry guidance that instrumentation should not
assert Ok.
rpcStart, jobQueue, jobStart and jobFinish each acquired a lock without
braces, so it stayed held to the end of the function and covered the OTel
recording calls this branch added. counters_.jobsMutex and
counters_.methodsMutex are process-wide, so every worker thread starting
or finishing a job serialised on the SDK's work.
That work is not a bare atomic add: each record builds a map-backed
attribute set and takes a spin lock inside the SDK, whose backoff reaches
a millisecond-scale sleep under contention.
Brace the lock plus the state it guards, then record after it releases.
None of the metric calls read lock-protected state, so this is
semantics-preserving, and rpcEnd() in the same file already had this
shape.
The row was titled after the rollout phase its metrics came from, a number
defined only in a planning folder outside the shipped tree. Grafana shows
this title to operators, so it named something no reader could look up.
Two panel descriptions pointed at a rollout phase and task number defined
only in a planning folder outside the shipped tree. The note still names
the file and the change it is waiting on, which is the part a reader can
act on.
The catalog tagged every span with a rollout phase number defined only in
a planning folder outside the shipped tree, so the column meant nothing
to a reader of the repository. The span name and source file identify
each entry.
