Add two Peer Quality panels reading peer_disconnect_total: Peer Disconnect
Rate, the per-second teardown rate per node, and Peer Disconnects By Reason
& Direction, the per-interval increase split by cause and by which side
opened the connection. Both sit in the existing Disconnects & Connection Mix
row beside Resource Disconnects, which counts only the resource-charge
subset and carries no reason label.
The Ledger Sync Health board already shows the same split as a window
total, so it says how much of each reason but not when. These give the
time-shaped view, letting a reason spike be lined up against a stall.
Add disconnect_reason and disconnect_direction template variables for the
two new label dimensions and wire both queries to them, so the panels
filter on every dimension their series carry.
Update the 09 reference panel column and the _a7_note panel list to name
the panels that now render this counter.
The Peer Disconnect Rate By Reason panel anchored its LogQL capture on
"\] ", which only matches a reason logged immediately after the [NNN]
peer-id prefix. PeerImp does not log that way: PeerImp::fail emits
"[NNN] <name> failed: <reason>" and the clean teardown emits
"close: Closed". Only ConnectAttempt::fail, which logs the bare reason,
ever matched. The panel's Timeout series was therefore connect-attempt
timeouts only, Ping Timeout was invisible, and PeerImp's own Closed was
uncounted.
Match all three prefixes and separate Ping Timeout from Connect Timeout.
Recorded as LogQL trap 11 in the runbook, alongside the other silent
failures this dashboard exposed.
Also document six overlay observability gaps found while auditing what
ping/pong and gossip traffic is actually tracked. All are pre-existing
and none is fixed here: the code fixes belong in develop-owned overlay
files (TrafficCount, OverlayImpl, PeerImp, PeerfinderManager), not on a
telemetry branch, and one of them needs a public signature change.
- 09 section 6: six known issues, each marked NOT IMPLEMENTED with
file:line evidence -- mtCLUSTER counted as unknown (overhead_cluster_*
always zero, 8 panels flatline), squelch_ignored byte counts always
zero, inbound/outbound byte-basis asymmetry plus a stale Total header
comment, ping/endpoints instrumentation absent, peer span coverage,
and PeerFinder exporting 2 of ~17 available readings.
- 02 section 2.3.2: add a Status column to the span catalog. Of 36
catalogued spans, 16 are live, 15 were never built, and 5 shipped
under different names (consensus.phase.establish became
consensus.establish, ledger.close became consensus.ledger_close,
rpc.request split into rpc.http_request/rpc.ws_message, txq.apply
became txq.apply_direct/txq.accept_tx). The catalog is a design
inventory; 09 section 1.1 remains authoritative for what emits.
- Phase9_taskList: tasks 9.14-9.17 tracking the deferred work, with
exit criteria checked only for what is actually done.
- Glossary: new Ping / pong keepalive term distinguishing ping timeout
from connect timeout. Correct the Cluster and Squelch entries, which
described behaviour the metrics cannot show.
The glossary header pointed at tasks/telemetry_terms.py as its
generator. That file is in no branch and nowhere on disk -- tasks/ is
gitignored one directory up -- so the header now states the file is
hand-maintained and gives the entry format.
Gates: check_otel_naming.py passes all 9 rules (Rule D over 555
dashboard queries, Rule E over the runbook); 19 doc anchors verified;
dashboard JSON valid with a one-line diff. No C++ changes.
Bumps the Loki pin this branch introduced, and updates the two docs that
name the version so they do not drift from the compose file:
- OpenTelemetryPlan/09-data-collection-reference.md (log backend section)
- OpenTelemetryPlan/Phase8_taskList.md (compose snippet)
The service runs the config bundled in the image rather than one from this
repo; `-verify-config` reports "config is valid" under 3.7.6, and native OTLP
ingestion at /otlp is unchanged across 3.4 -> 3.7.
The heading still read "Future:" while the section body documents the
validation suite this branch ships and gives commands to run it. Section 6.8.3
in 06-implementation-phases.md already drops its "Future Enhancement" marker
here; this makes 5c consistent with it.
These changes were developed on the phase-10 branch but belong to content this
branch and its upstreams introduced. Carrying them on phase-10 made its PR diff
report churn in files phase-10 does not own, and left each PR claiming a scope
that did not match its contents.
Moved here from phase-10 (identical content, no functional change):
- Dashboards: all 14 existing boards plus the new log-derived-insights board.
- Docs: telemetry-runbook.md (minus the workload/benchmark sections, which
describe phase-10 tooling) and the new telemetry-glossary.md.
- Grafana Cloud + Alloy export path: collector config, compose override, the
two .env examples and alloy/config.alloy.
- Local stack: otel-collector-config.yaml gains sub-millisecond and
second-scale spanmetrics buckets, pins unit=ms, and promotes
close_time_correct; integration-test.sh and TESTING.md follow.
- Node configs: exported_instance -> service_instance_id in comments; the
mainnet sample now logs at warning to bound log volume.
- Metrics code: Telemetry.cpp builds the metrics pipeline in the constructor
via initMetrics() so the global MeterProvider is published before any
subsystem creates a beast::insight instrument, and the histogram view keeps
each instrument's own name instead of collapsing them under one series.
MetricsRegistry gains a last_close_time gauge and skips negative job-queue
durations. OTelCollector drops an unused accessor.
- Naming CI: xrpl_work_item joins EXTERNAL_INFRA_LABELS and Rule E accepts the
dotted perf-iac resource-attribute form. This must travel with the
dashboards and runbook that reference those labels, or the rules fail.
- Doxygen input glob no longer recurses dot-directories.
Sections describing phase-10 tooling stay on phase-10 and keep their
"Future Enhancement" / "Planned, not yet implemented" markers here; phase-10
removes those markers when it lands the tooling.
Three measurement fixes landed with no doc or dashboard change, leaving text
that is now false and one fix unusable from a dashboard.
Deferrals and timeouts are recorded in TimeoutCounter, a base shared by five
subclasses, so the all-lane pair could show the documented livelock
fingerprint while ledger acquisition was healthy. The runbook procedure and
the reference doc now name acquire_ledger_deferrals and
acquire_ledger_timeouts and say why the all-lane pair misleads; a new panel
plots the ledger-scoped pair as rates on one axis, since the divergence is
the signal. The existing panel is retitled All Lanes and points at it.
Writer mean depth is depthSum over depthSamples, not over insertCount, and
the measured 1.60 came from the biased estimator, so it and the 37% queueing
share derived from it are lower bounds rather than values. The reference
table now marks them as such, and the decision rule is shown to survive the
correction rather than depending on the exact figures.
Completions were never counted for acquisitions satisfied from the local
store, so the run that read zero across 510 seconds had in fact reached
full. Every place that treated a zero as a symptom now says it only means
something on a build that has the fix.
Also corrects the sync-diagnosis label-value count from 13 to 15 and a stale
source line range; the instrument count stays 35, because both new values
multiplex onto the existing nodestore_state gauge.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Four findings from a review pass over the PR.
The "Spans & traces" row was empty. Moving the row header down to clear
the back-fill panels was only half the change -- the seven span-derived
panels stayed at their old y, one unit below the native panels, so every
pair overlapped and Grafana parented all fifteen to "Back-fill &
persistence". The panels now sit below the row header, which restores
the split the runbook already describes: eight native panels answer "how
much", seven span-derived ones answer "which". Both rows stay expanded,
so the docs no longer call them collapsed.
metric_constants() excises each namespaced block before the flat
prefix pass. The flat pass classifies by identifier prefix and is meant
for headers that name the role in the identifier because they have no
`namespace metric`/`label`/`lval`; it was running over the whole header,
so a `kLabel`-prefixed constant written inside `namespace metric` landed
in both buckets and an instrument name became a valid label key for Rule
D. Nothing in the tree does that today, which is why it went unnoticed,
and why the guard is a test rather than a fix for an observed failure.
The `site` label now keeps a non-default port and drops userinfo residue
from the host. Omitting the port unconditionally merged two local sites
that differ only by port; printing it unconditionally would have renamed
the existing `https://vl.ripple.com` series. Comparing against the
scheme default distinguishes a configured port from the one the Resource
constructor fills in. parseUrl's host group also permits '@', so a
malformed URI with two of them leaves part of the userinfo in `domain`.
Adversarial validation of the previous commit found one of its two code fixes
was diagnosed wrongly and the other incomplete. Both are corrected here, along
with the layers the first pass missed.
1. The new dial outcome was named for the wrong condition. It was added as
`duplicate` on the belief that PeerFinder had already granted a slot for the
address. It has not: `Logic::onConnected` contains exactly ONE false-returning
path and it is the self-connect check, which logs "Logic dropping as self
connect" (include/xrpl/peerfinder/detail/Logic.h). The duplicate check lives
in `newOutboundSlot`, evaluated before a ConnectAttempt exists, so a real
duplicate can never reach this branch.
That mattered beyond the name: the previous commit told operators the outcome
was benign churn to ignore, when it actually reports a local misconfiguration
-- this node has its own address in [ips_fixed] or behind its advertised
endpoint, and every dial to it is wasted. Renamed to `self_connection`,
reusing the slug `handshake_negotiation_fail_total` already publishes for the
same fault so it reads identically on both signals, and every description
corrected to say so. The fail() string now reads "Self connection" too.
The first pass also missed three enforcement and contract sites: the
ConnectAttempt.h Doxygen state machine (which still mapped the slot branch
onto tls_fail), the LedgerSpanNames unit test (which pinned exactly five
values over a std::array<..., 5> and so left the new member untested), and the
span-derived twin panel plus two reference docs that still published the old
five-value domain.
2. The credential-free site label was incomplete twice over.
- It appended the port, and `Resource::Resource` DEFAULTS that to 443/https
and 80/http when the config omits one. The label would have become
`https://vl.ripple.com:443/` where Grafana Cloud currently holds
`https://vl.ripple.com`, silently renaming the series for every deployment
already scraping this metric. Verified against live label values before and
after; the port is now omitted.
- parseUrl's path group is `(/.*)?`, greedy to end of string, so a query or
fragment lands inside `path`. A list URL authenticated by `?token=...` would
have leaked exactly as userinfo did. The path is now truncated at the first
'?' or '#'.
Also updated the MetricNames.h usage example, which still taught the raw-URI
pattern to the next author, and the 09-doc row that described the label as the
configured URI.
3. Rule J hardening from the same review: `classify_instrument_kind` returns an
`other` sentinel for a non-factory macro, and storing it in the kind set could
render a future conflict as "created as counter and other". The sentinel is
now skipped, keeping it doing what it already did -- matching no shape rule.
Added a second regression test whose input the pre-fix code reported as CLEAN
(gauge-then-histogram on a `_us` name), so the guard is proven by a 0-vs-1
difference and not only by a changed message. Both new tests were run against
a reconstructed last-wins implementation and both fail against it.
Documented the conflict class in the Rule J rows of the checker README and
CONTRIBUTING, which previously described only the suffix conventions.
Verified: naming checker exits 0 with Rule J passing all 40 real names; 140
checker tests pass; 15 dashboards validate; both workload JSON files parse;
clang-tidy over the full compile database reports no finding on any changed line
of ConnectAttempt.cpp or ValidatorSite.cpp; pre-commit passes.
Not verified: not compiled. The label change adds string truncation and the
outcome rename touches a constexpr used across three translation units, so CI's
build remains the first real check on both.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Brings in the phase-10 revert of the nodestore read-latency histogram plus the
nudb_bytes -> stored_object_bytes rename.
Conflicts in MetricsRegistry.{h,cpp} resolved keeping both intents:
- MetricsRegistry.cpp: dropped everything that existed only to serve the
reverted nodestore_read_us histogram -- the addSubMillisecondHistogramView()
helper, its call site, the kSubMillisecondBoundaries array and the
NodeStoreMetricNames.h include. Kept every view this branch registers
(consensus round duration, sweep_malloc_trim_us, dns_resolve_latency_ms,
overlay_dial_latency_ms) and the shared addHistogramView() base helper.
Took the rename at the storage_detail observe() call site.
- MetricsRegistry.h: took phase-10's move of the four nodestore_state observe
helpers and their ObserveFn sink from private to public, while keeping this
branch's enriched Doxygen on observeNodeStoreTotals().
Also corrected the registered-view count in the 09 reference doc: neither side's
arithmetic survives the merge, since this branch adds four views phase-10 never
saw and the revert removes one. Ten views are registered now, not six or seven.
Drops nodestore_read_us and everything added to reach it. read_mean_us already
carries microsecond precision and separated the two sync failure modes cleanly
in live testing -- 8.8 us on a clean store against a 223 us cold-store peak --
so the distribution added no signal that changed a diagnosis.
The cost of getting it was disproportionate. NodeStoreScheduler had no path to
the metrics registry, so its production constructor grew a ServiceRegistry
parameter: a metric addition changing a production signature. That in turn
forced an edit to a pre-existing test, src/test/app/SHAMapStore_test.cpp, whose
only stake in this is that it constructs a scheduler. Worse, the scheduler is
built in Application's member initializer list, long before metricsRegistry_
exists, so the registry could not be captured once and had to be re-resolved on
every fetch -- a lookup on a path that runs millions of times per sync.
The constructor returns to taking JobQueue& alone and SHAMapStore_test.cpp
returns to the single-argument call, leaving that file differing from its
pre-change form only by the NodeStore:: to node_store:: rename it picked up from
develop.
FetchReport::elapsed stays microseconds and onFetch keeps its explicit
duration_cast to milliseconds for addLoadEvents, which takes milliseconds. That
widening was a separate fix and is what makes read latency measurable at all.
kSubMillisecondBoundaries loses its only consumer and regains [[maybe_unused]],
which is the state the commit that introduced it left it in; without the
attribute an unused constant is an error under wextra with werr.
Also removes the ledger-data-sync panel that charted the histogram and the
fetch_type and found template variables, which filtered on labels no metric
emits any more, plus the runbook and reference-doc sections and the two
instrument and view counts that named it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The nudb_bytes label value on the storage_detail gauge named something the
code never measured. It observes Database::getStoreSize(), which returns the
storeSz_ accumulator: the cumulative payload bytes of objects this process
handed to the NodeStore. That is not a NuDB file size. It excludes NuDB's
keys, bucket padding and log, and it resets with the process while the files
on disk do not.
The name caused two concrete errors. It invited sizing the store on disk from
a number that cannot do it, and it invited a write-amplification ratio against
node_written_bytes -- which reads the same accessor at MetricsRegistry.cpp:836,
so that ratio is a constant 1.0 and measures nothing.
The nudb_ prefix was wrong too. storeSz_ is written only by
Database::storeStats(), called from DatabaseNodeImp, DatabaseRotatingImp and
Database itself. No backend code touches it, so the value reads the same on
RocksDB. That distinguishes it from the real nudb_* family
(nudb_writers_in_flight and friends), which come from getWriteStats() and are
absent entirely on a non-NuDB backend.
stored_object_bytes says what the value is and claims nothing about the
filesystem. Docs already described the value correctly; they keep that
explanation and now also record the old name, so a query pinned to it can be
traced. Neither Backend nor Database exposes an on-disk size accessor and none
was added -- no metric reports the store's on-disk size today.
Updates the node-health panel title, description and PromQL, and the four docs
that name the label value.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Docs-only from phase-10: two commits refining the sync-diagnosis decision rule
and nodestore_state metric semantics.
Reviewed for conflicts with this branch's work and found none. Two points worth
noting because they touch documentation this branch also edits:
- node_reads_hit is re-described from "reads served from cache" to "fetches that
found an object (not a cache hit)". That is the accurate reading of the
counter and does not change any query here.
- The cold-read decision rule is retuned (insert mean threshold ~1.2, and the
read-mean/found-rate split for telling cold-but-held from real misses). It
does not contradict the retirement of nodestore_latency on this branch, which
removed a duplicate gauge rather than any nodestore_state series.
git merge-tree reported 0 conflicts before merging.
Two conflicts, both additive-vs-additive; each resolution keeps both sides.
check_otel_naming.py -- phase-10 taught the L6 label extractor to match the
label MAP first and to resolve a key hoisted into a `k...Label` constant,
scanning headers as well as sources. Our side had added the two-regex
first/subsequent literal scan and the `metric_constants(root)[1]` union that
covers the `namespace label` header style.
Kept phase-10's mechanism whole: METRIC_LABEL_MAP + the `(?:^|\{)` key regex
already subsumes what METRIC_LABEL_NEXT did, since matching inside the map body
makes every pair after the first open with a single `{`. So METRIC_LABEL_NEXT is
dropped as genuinely redundant rather than kept as a duplicate scan, and the
reason it existed is folded into METRIC_LABEL's comment. Re-added our
`metric_constants(root)[1]` union on top: LABEL_CONST_DEF only matches
`k`-prefixed identifiers, so it cannot see MetricNames.h's `label::jobType`
style, and without that union Rule D would reject dashboards querying labels
Rule I forced into constants. The two derivations are complementary and both
are now documented as such.
MetricsRegistry.cpp -- both sides added a new sibling view-registration helper
next to addMicrosecondHistogramView, and both added a registration call in
initExporterAndProvider(). Kept all four helpers
(addHistogramView/Microsecond/RoundDuration/SubMillisecond) and every
registration: phase-10's addSubMillisecondHistogramView + kNodeStoreReadUs
alongside our addRoundDurationHistogramView, sweepMallocTrimUs and the two
millisecond dial/resolve ladders.
phase-10's nodestore_read_us histogram does not duplicate our work. The
nodestore_latency gauge that would have overlapped it was retired in c4e434d520
before this merge, and the surviving nodestore_state gauge is complementary
rather than duplicative: both read the same fetch measurement, but the gauge
publishes only a since-boot mean via scaledMean() and cannot yield a
percentile -- the consequence observeNodeStoreTotals' own docs state plainly --
while the histogram buckets each fetch and can. The histogram also splits by
fetch_type and found, which the gauge cannot. phase-10 registered its
explicit-bucket View, so it does not inherit the SDK default ladder.
Each file keeps its own existing naming style: phase-10's k-prefixed constants
are left as-is, ours stay namespaced.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The decision rule for "slow to reach full" keyed on an absolute read-time
threshold and a found-rate threshold that misclassified the very run they
were written to explain: a populated-store run reading 31.8 us at 88.3%
found fell through both cuts and came out as "disk-bound" rather than the
cold-read case it is.
Replace it with two ordered questions -- is the read cost several times a
warm read, and is the write path queueing -- and demote the found rate to a
splitter that only applies once reads are known to be expensive. A high
found rate on its own is the normal state of a populated store, so it can
never be a trigger. The rule now classifies all four reference datasets
correctly, and the runbook shows the rule applied to each so the "confirm
against the reference points" step agrees with the table.
Also in the runbook:
- name the source of the devnet incident figures at the point of use, and
point forward to the caveat from the same paragraph
- state the provenance of the measured columns, and split the incident
figures into their own table marked as not our measurement
- say plainly that the compounding-factor explanation is an unconfirmed
hypothesis
- correct the deferral gate: it fires at the acquisition's own job limit of
5, not at the ledgerData lane cap of 3
- note that no read-max gauge exists, so the tie-break uses max_over_time
of the mean or the read histogram's p99
- split a PromQL block that put two expressions on adjacent lines, which
parses as one invalid expression
In the data-collection reference:
- node_reads_hit counts fetches that found an object, not cache hits
- nudb_bytes is cumulative payload bytes from the same accessor as
node_written_bytes, not on-disk size, so their ratio is a constant 1.0
- write_load and nudb_writers_in_flight are the same atomic on NuDB;
document that and what write_load means on RocksDB
- correct the instrument count to 8 and the view count to 7 after
nodestore_read_us and its view were added
- scope the Phase 9 query examples to one node with the regex form
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two different bottlenecks both present as the ledgerData job lane pinned
at its concurrency cap of 3, so lane occupancy diagnoses neither. One is
write-serialized (NuDB takes one global mutex per insert, so inserts
queue), the other is cold-read-bound on a populated store. Telling them
apart needs the storage-side signals, not the lane.
Adds to docs/telemetry-runbook.md a "Slow to reach full" procedure: a
Mermaid diagram of the two modes, a decision table keyed on whether
acquisitions are completing, the measured reference values from both
runs, and the deferral/timeout pair that fingerprints the disarmed
give-up path. States plainly that node_reads_hit is a found count rather
than a cache-hit rate, which is why a ~100% "hit rate" at 113 us per
read is the cold-read signature and not a contradiction.
Records honestly that the populated-store run was twice as fast despite
slower reads, so cold reads alone do not explain the long incident.
Adds reference rows for the 13 new nodestore_state label values and the
nodestore_read_us histogram to
OpenTelemetryPlan/09-data-collection-reference.md, in the authoritative
Phase 9 OTel SDK section alongside the existing NodeStore I/O table.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
nodestore_latency published six values that nodestore_state already
publishes from the same Database accessors, so the two gauges were
duplicate readings of the same atomics:
write_count -> node_writes getStoreCount()
read_count -> node_reads_total getFetchTotalCount()
write_duration_us -> node_writes_duration_us getStoreDurationUs()
read_duration_us -> node_reads_duration_us getFetchDurationUs()
write_mean_us -> write_mean_us store duration / count
read_mean_us -> read_mean_us fetch duration / count
nodestore_state is kept because its means go through scaledMean(), which
saturates at INT64_MAX instead of wrapping and omits a mean when the
denominator is zero rather than reporting a misleading 0 us.
Removes registerNodeStoreLatencyGauge, its instrument member, the
metric::nodestoreLatency constant and the lval::nodestore_latency label
namespace. The gauge-over-histogram rationale and the "p99 is not
obtainable" consequence are folded into observeNodeStoreTotals' docs.
Retargets the gauge-contract test onto nodestore_state rather than
deleting it: the scaledMean arithmetic is covered by the static_asserts
in tests/libxrpl/telemetry/MetricsRegistry.cpp, but nothing else asserts
that these named series multiplex onto one instrument keyed by `metric`.
The test now calls the production scaledMean instead of a copy of the
division, and its sub-microsecond case asserts scaledMean's actual
behaviour (a genuine mean of 0 on a zero numerator with a non-zero
count), which differs from the retired gauge's extra numerator guard.
Rewrites both ledger-sync-health copies' panel 38/39 queries and drops
the obsolete claim that the write numerator was never written: all three
concrete store paths call recordStoreDuration, so write_mean_us is live
on an ordinary node. The same stale [import_db] caveat is removed from
the runbook, the 09 reference row and the workload validator's note.
The Peer Ledger Supply Window panel drew supply_min_seq, supply_max_seq and
nothing else on one linear axis. Measured on a mainnet node, those sit around
105,890,000 and roughly 300,000 apart, so the 588-ledger tip movement that
shows whether sync is progressing was 0.0006% of the axis and read as a flat
line. unit "none" also printed the sequences unabbreviated and clipped the
legend.
The panel's own "Watch for" text asked the reader to compare supply_min_seq
against this node's validated sequence, but that line was not on the panel at
all, so the comparison meant switching dashboards.
Plot the two distances instead, which is what the panel was always asking
about:
History Headroom = validated_ledger_seq - supply_min_seq
Tip Gap = supply_max_seq - validated_ledger_seq
Zero is now the boundary in both directions: negative headroom is exactly the
"every peer pruned what I still need" case the description warns about, and it
becomes a zero crossing rather than a line-order comparison. Tip Gap gets the
right-hand axis because the two ranges differ by orders of magnitude
(measured: 299999..300001 against -1..1).
Both operands are gated `> 0`. Ungated, differencing the documented
"unknown window" sentinel of 0 yields the whole sequence space: measured
-105854935 for headroom and 105890295 for tip gap during the first ticks,
which destroys the axis for the rest of the window. Gated, the panel stays
blank until the node has a validated ledger and a peer has advertised a
range, which is the honest reading for that state.
Both queries verified against a live mainnet node through the full template
substitution: refId A = 300001 legend "History Headroom [xrpld-mainnet]",
refId B = -1 legend "Tip Gap [xrpld-mainnet]".
Runbook branch-C table, step 11 walkthrough and the 09 reference row follow
the rename and the new reading.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two suspects from the 3.3.0 slowdown investigation had no signal. Both were
already computing the numbers and throwing them away, so this exposes them
rather than adding measurement.
Per-sweep heap trim. The trim runs after every cache sweep, and its cost
scales with resident heap, so it is the leading explanation for a node with
a populated database syncing slower than a fresh one. The report already
carried duration, fault deltas and reclaimed pages, but the whole
measurement sat behind a debug-journal check, so an ordinary node measured
nothing, and the call site discarded the result. The measurement now always
runs and only the log line stays gated. Records trim duration, minor faults
and reclaimed kilobytes. Measured cost of the always-on path is about six
microseconds per sweep against a trim costing milliseconds, at a cadence of
ten to a hundred and twenty seconds.
Honest limit, stated in the runbook: the fault delta spans only the trim
call, so it shows the trim itself faulting but not the faults that follow as
caches refill. The duration is the signal to correlate against sweep-job
queueing.
Rotation writes. Rotation copies archive-served reads forward and re-stores
nodes missing from both backends, both of which compete with sync I/O and
only happen on a populated online_delete database. The copy-forward count
existed but was reset by the rotation's own log line, so a metric reading it
would drop to zero on every swap; a never-reset total sits beside it now.
The re-store count was not measured at all. Rotation duration is
deliberately not recorded: the health throttle sleeps at eight points inside
the sequence and dominates exactly when the node is unhealthy, so the number
would conflate work with waiting.
Nothing added for the other two suspects. Get-object serving is already
covered by the handler label, the lookup histogram and the deferred and
saturation gauges; peer churn by the disconnect-reason counter.
Also replaces nine per-file cspell ignores with one ignoreRegExpList entry
for the telemetry macro names, and picks up the levelization baseline for the
consensus span-name test.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Found by reviewing what each metric actually measures, with attention to the
derived and bucketed ones. All five could report healthy while the node was
not, or the reverse.
- The nodestore latency panel took rate() of a mean. The gauge already
divides duration by count in code, so rating it produced a figure with no
unit, and Prometheus discards a gauge's decreases, so a heavy back-fill
read as roughly zero microseconds per operation. The cumulative duration
totals are now exported alongside the means, and the panel divides the
rate of the total by the rate of the count, which is the latency over the
panel's own window rather than a since-boot average that flattens with
uptime.
- The DNS-resolve and outbound-dial histograms had no explicit buckets, so
they inherited a ladder that stops at ten seconds while the dial timer is
fifteen. Every timed-out dial fell in the overflow bucket and p95 read
exactly ten seconds however bad it got. Both now have a ladder reaching
thirty seconds with fifteen on its own boundary, so a timeout is
distinguishable from merely slow.
- The missing-node counts only cleared when a tree completed, so a
timed-out or failed acquire left its last count latched. Since the gauge
reports the maximum across everything still in the collection, and
eviction waits on a grace period plus the sweep interval, a finished node
reported as stuck for minutes. That inverts the one signal that separates
stuck from slow. Cleared unconditionally on the terminal path instead.
- A disabled quorum published a sentinel so large that, on a timeseries
axis shared with the trusted-key count, it flattened the key line to the
baseline and hid the outage it was meant to mark. The series is now
omitted and a quorum_disabled flag carries the state.
- Two panel descriptions claimed a one-second export cycle. The reader is
configured for ten.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Phase-10 independently instrumented the peer object-fetch path while this
branch instrumented fresh-node sync, so the two overlapped in three places.
Resolved by keeping each side's stronger implementation rather than shipping
both.
Per-job-type waiting/running/deferred existed twice. Phase-10's version
survives: it publishes per-type gauges from JobQueue::collect(), which
snapshots under the queue lock and publishes after releasing it, a
deliberate lock-order fix against the collector's own lock. This branch's
jobq_backlog gauge and the JobQueue::getJobTypeCounts() accessor that fed it
are removed, along with their panels, assertions and reference rows.
jobq_saturation stays: it reports the whole worker pool, which phase-10 has
no equivalent for.
The histogram view helper also existed twice with identical bodies under two
names; one survives, and the microsecond ladder is now the named array
rather than boundaries repeated inline. The job_type label was declared
twice, once as a file-local constant invisible to the naming check; both it
and handler now come from the constants header.
Two things phase-10 adds are complementary, not duplicates, and are kept as
they are: the handler label, which separates the two request kinds that both
report as the same job type, and getobject_rejected_total, which counts
malformed requests where this branch's serve_refused_total counts requests
this node declined to serve.
Also fixes two naming-check failures that pre-date this merge on phase-10.
The check derived label keys only from namespaced constants, so it could not
see the per-subsystem headers' flat k-prefixed style and rejected dashboards
querying labels the code really emits. It now reads both styles, with the
enforcement rules unchanged.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The board and runbook had grown by append across eight work packages, so
they read in the order the work was done rather than the order a node
progresses. This is the coherence pass; it adds no new instrumentation.
- Dashboard: 52 panels regrouped from two rows into nine that follow the
fresh-start sequence — bootstrap, peer supply, sync state, acquire and
SHAMap fetch, job queue, quorum and publish, terminal blockers, then
back-fill and spans collapsed since they answer conditional questions.
Layout only: no title, query or description changed.
- Runbook: the flat step list becomes a decision tree branching on the
observed symptom, with the amendment-block check first because it is
terminal. Each branch names the panels, what healthy and unhealthy look
like, and what to conclude. The existing steps are kept as the detail
bodies.
- Reference table: every signal name re-checked against the code and every
named panel against the board; four stale panel references fixed.
- Validation: every signal is now either asserted or covered by a note
explaining why a five-node local cluster cannot produce it.
Also fixes the write-latency signal, which was inert on a real node: the
store duration was only recorded on the database-import path, while the two
production store implementations did not time themselves, so an ordinary
node reported a write count with no latency. Both now time the backend
write, which is the disk work this signal exists to expose. Without it the
"existing database syncs slower than a fresh one" diagnosis had no primary
signal.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A slow fresh-sync ledger produced spans scattered across threads with no
way to relate them. They now share a trace id derived from the ledger's own
hash, the one value every participating site already holds, so nothing new
is plumbed across threads. This is the pattern the transaction pipeline
already uses for its tx id.
Joined: ledger.validate, ledger.store, and a new
consensus.validation.accept recorded when a trusted validation arrives. In
Tempo, searching one ledger hash returns them together, so an operator can
tell whether the ledger was slow to arrive, slow to be accepted, or slow to
be stored. They are siblings rather than a chain because the accept gate is
entered from three different threads, so no fixed parent order exists.
consensus.validation.accept also records why an arriving validation did or
did not advance the gate, which makes "validations arrive but are all
rejected" visible for the first time.
consensus_round_duration_ms turns the existing round-time span attribute
into a histogram, so a fleet trend needs a metric query rather than raw
trace inspection. An explicit bucket view is required, not optional: the
SDK default tops out at ten seconds while consensus abandons a round at two
minutes, so slow rounds would all fall in one bucket and every quantile
would read exactly ten seconds. Cost is one record per round.
Record layer: the histogram is native and needs no collector change. The
two new bounded attributes are added as span-metric dimensions to both
collector configs. The ledger hash stays out of them, since a per-ledger
dimension mints a series per ledger; it is indexed in Tempo as the join key.
The ledger.acquire span is not joined yet, because that file was being
changed concurrently. It is registered as an optional member of the join
group so nothing fails, and switching it is a one-line follow-up.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The quorum and publish gauges were emitted but never surfaced: no panel, no
harness assertion, no reference entry. Completes those layers.
- Four panels: trusted validations against the quorum target on one axis so
a tally climbing toward quorum is visually distinct from one flat below
it; publish lag; pre-accept shortfall rate; and time to first validated
ledger.
- Both signals are asserted by the workload validator. The shortfall
counter does fire on a healthy cluster, because this node validates and
then immediately re-enters the accept gate before its peers' validations
arrive, so the first evaluation of every round tallies short. The panel
and note say so, and give the fault signature instead: the shortfall rate
outpacing the ledger-close rate while the tally stays flat and nothing
ever reaches first-validated.
- The quorum target is deliberately drawn as its own line rather than as a
headroom stat, so the disabled-quorum sentinel reads as an unreachable
target instead of an unreadable negative number.
Also removes three reference rows that were appended twice when two agents
each documented the same back-fill signals.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Registers the new gauges, renders them, asserts them and documents them, so
each signal reaches an operator rather than stopping at the emit site:
- MetricsRegistry: gauge registration for ledger_quorum_publish,
nodestore_latency, peer_ledger_supply, peerfinder_slot_census and
amendment_block, each guarded by the detached-callbacks check and
tolerant of services that are not ready yet.
- Ledger Sync Health dashboard: panels for the new signals, filtered by
the node template variable like every other board.
- Workload validation: the new series are asserted, so a signal that
regresses to absent fails CI. Signals the local cluster structurally
cannot produce, such as a replay fallback or an amendment block, are
noted rather than asserted, which would fail red on a healthy run.
- Reference, runbook and glossary entries, including the diagnosis order
for a node that has peers and validators but never validates.
- Regenerated levelization baseline: three new one-way edges from the
telemetry and test modules, no new cycles.
Also drops an unused cstddef include from the macro tests, which the
include checker rejects.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Sync-critical job types run at very low concurrency limits (ledgerRequest
and ledgerData allow 3 each), so a node can stall simply because those
jobs are held back behind other work. Nothing exposed that until now:
the existing job metrics are rates and quantiles of jobs that already
moved, or a single queue-wide depth.
- jobq_backlog{metric,job_type}: instantaneous waiting, running and
deferred counts per job type. Deferred is the starvation signal and had
no exposure anywhere; it is set when a type is at its concurrency limit.
- jobq_saturation{metric}: running tasks, worker-thread count and total
waiting, so a slowdown spanning several subsystems can be attributed to
worker-pool exhaustion instead of being diagnosed once per victim.
Both read through two new const accessors on JobQueue that take the
existing mutex once and copy integers, so a single reading is internally
consistent and no per-job cost is added. The job_type label reuses the
same JobTypes name helper the existing job counters use, so the two label
sets join.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Resolved five files. In each case both sides had content worth keeping,
so nothing was taken wholesale:
- MetricsRegistry.cpp: kept the incoming `handler` label on the job
instruments and re-applied this branch's `queuedDurUs >= 0` guard,
which the incoming side does not have.
- telemetry-runbook.md: took the incoming gauge table, which adds the
three per-job-type rows, and re-applied this branch's corrected
`jobq_job_count` name.
- 09-data-collection-reference.md: kept this branch's validation
inventory (newer counts, extra Config File column) and inserted the
incoming call-site and per-job-type gauge rows plus their explanation.
- node-health.json: merged structurally rather than by text. This
branch's panels are authoritative; only the two incoming job-queue
panels were appended, below the existing layout. The
`Validated Ledger Seq` panels added directly in Grafana are preserved.
- job-queue.json, ledger-data-sync.json: panel sets were identical, so
took the incoming side for its `$handler` variable, the handler filter
on existing queries, and the new panels. Verified no panel was lost.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Resolved OpenTelemetryPlan/09-data-collection-reference.md: both sides
document the same new metrics, reached by different routes. The incoming
sections came from the phase-6 edit merged forward; this branch already
documents the same surface in its own layout, with the job lifecycle
metrics and the GetObject request path under the MetricsRegistry
section rather than a separate 2a.
Kept this branch's sections to avoid duplicate entries for every metric.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Resolved OpenTelemetryPlan/09-data-collection-reference.md: this branch
replaced the StatsD-shaped Section 2 with OTel-native naming, so the
incoming StatsD tables and the StatsD per-job timer section do not apply
here. Kept this branch's names throughout.
The incoming per-job-type saturation gauge section is new content and is
retained, renumbered to 2.5 and rewritten for the OTel naming: bare
lowercase `jobq_<jobtype>_{waiting,running,deferred}` derived through
OTelCollector rather than the case-preserving `xrpld_`-prefixed StatsD
form.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Slowness on the peer object-fetch path could be observed but not
attributed. Job duration metrics carry only `job_type`, and both
`RcvGetLedger` and `RcvGetObjByHash` report as `ledgerRequest`, so a
queue-wait spike could not be traced to a handler. Nothing measured
NodeStore cost, request size, or the differential charge.
Latency now decomposes into three additive parts, each separately
measurable:
end-to-end = queue wait + NodeStore lookup + everything else
- `handler` label on job_queued_total/_started_total/_finished_total and
job_queued_us/job_running_us. The value is sanitised: a name passes
through only if non-empty and all ASCII letters, else "other". Two job
names embed a ledger sequence, so a raw label would mint one series
per ledger; the rule bounds the domain at 43 names plus "other".
- getobject_lookup_us, _request_objects, _lookups_total{result},
_rejected_total{reason} and _charge, recorded at their call sites.
All three histograms get explicit bucket views: the SDK default stops
at 10,000, which every one of them exceeds.
- Per-job-type waiting/running/deferred gauges for the 35 non-special
job types. `deferred` is the leading indicator, since addJob never
rejects -- it defers, so backpressure otherwise shows up only as
latency after the fact.
`JobQueue::collect()` snapshots the counters under the queue lock and
publishes gauges after releasing it. Writing them while holding the lock
would invert a lock order against the collector's own lock, which the
collector's flush thread already holds when it calls this hook.
Tests assert exact values, including that the charge is priced on the
requested count rather than the capped iteration count.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Add reference entries for the observability surface introduced on
phase-9: the `handler` label on the job instruments, the five
`getobject_*` request metrics, and the per-job-type queue saturation
gauges.
Names here follow this branch's StatsD pipeline, which preserves case
and carries the `xrpld_` prefix, so they differ from the lowercased
OTel-native names used from phase-7 onward. The sections state where
the implementing code lives, since it is introduced downstream.
Also correct pre-existing entries: `job_count` exports as
`jobq_job_count` via the collector group prefix, the non-special job
type count is 35 (not 36), and `JtLedgerData` has five producers.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Signals that separate a sync that is merely slow from one that will never
finish:
- sync_acquire{missing_state_nodes_max, missing_tx_nodes_max, in_flight,
received_data_depth}: how many SHAMap nodes each in-flight acquire is
still waiting for. getMissingNodes already computed this and the callers
discarded it after a trace log. A count that stays flat means the
acquire is wedged; a shrinking count means it is progressing. Recorded
once per sweep, never inside the per-node walk, and reset when a tree
completes so a finished acquire does not read as stuck forever.
- shamap_cache_hit_rate{treenode}: hit rate of the in-memory tree-node
cache, which sits above the node store, so it is distinct from the
existing NuDB ratio. A cold cache on a fresh node sends every traversal
step to disk.
- sync_acquire_no_progress_total: timer ticks where an acquire made no
progress, previously only logged.
- sync_addnode_total{good,duplicate,invalid}: whether arriving nodes are
useful, duplicated or rejected, so wasted fetch work is visible.
- sync_acquire_source_total{local,network}: whether a ledger was served
from the local store or had to be fetched.
Adds getBad()/getDuplicate() to SHAMapAddNode and an acquireProgress()
accessor on InboundLedgers so the xrpld gauge can read these without
libxrpl depending on telemetry.
ledger_seq is deliberately not a metric label: it is unbounded. Per-ledger
identity stays on the ledger.acquire span; the metrics expose bounded
aggregates instead.
The full-below cache hit rate is not exported: KeyCache updates different
counters than getHitRate() reads, so it would always report zero. That
libxrpl bug is documented rather than papered over.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Five signals that explain why a node is not advancing toward full, none of
which were observable before:
- state_changes_total now carries {from,to} mode labels, emitted at
setMode using the existing strOperatingMode helper. A bare count could
not distinguish a healthy climb from a node flapping between tracking
and connected. Removes the now-unused incrementStateChanges wrapper.
- sync_state{initial_full_duration_us}: time to first reach full, which
StateAccounting already computed but exposed only in server_info.
- sync_state{network_ledger_gate}: whether the node is still refusing to
build ledgers because it has no network ledger.
- sync_state{server_stall_seconds} and server_stall_events_total: how
long the main thread has been unresponsive. LoadManager computed this
and only logged it, so a stall was invisible until the fatal threshold.
The episode rule is a pure function so it can be tested without adding
a test-only mutator to LoadManager.
- sync_state{ledgers_behind}: how far our validated sequence trails the
best sequence any peer advertises, read from already-cached peer ranges
so no extra network traffic is added.
Also fixes the naming checker: it derived only the first label of a
multi-label instrument, so a dashboard querying the second label was
wrongly rejected.
Note: the clang-tidy hook cannot run in this worktree (no build
directory); the remaining pre-commit hooks, the naming check, dashboard
schema and harness syntax all pass.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A freshly started node most often stalls before it ever peers or reaches
quorum, and that whole chain had no telemetry. Adds the six signals that
make it observable:
- dns_resolve_total / dns_resolve_latency_ms: configured-peer hostname
resolution, emitted from OverlayImpl so libxrpl stays independent.
- overlay_connect_total / overlay_dial_latency_ms: outbound dial outcome
by terminal reason, plus dial duration.
- handshake_negotiation_fail_total: protocol and network-id negotiation
rejections, labelled by reason, so a misconfigured network is no longer
indistinguishable from unreachable peers.
- unl_fetch_total and the unl_quorum gauge: validator-list fetch outcome
per site and trusted key count against the required quorum. Without
these a bad validators.txt leaves the node syncing forever with no
signal.
- clock_close_offset_seconds: network close-time offset, which server_info
hides below 60s but which stalls consensus participation.
Panels land in the Bootstrap row of the Ledger Sync Health dashboard, the
metrics are asserted by the workload validator, and both the reference and
the runbook flow describe them.
Levelization baseline regenerated: overlay now includes MetricMacros.h, so
the overlay/telemetry pair is reported one-way instead of bidirectional.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
