Fix pebble db metrics - #4031
Conversation
ff62126 to
4783f13
Compare
PR SummaryLow Risk Overview MVCC-layer metrics now attach a Pebble-internal scraping is removed from Operators should expect metric shape changes for some former counters (gauges or delta-based counters; Reviewed by Cursor Bugbot for commit 08e13a0. Bugbot is set up for automated code reviews on this repo. Configure here. |
|
The latest Buf updates on your PR. Results from workflow Buf / buf (pull_request).
|
Codecov Report❌ Patch coverage is Additional details and impacted files@@ Coverage Diff @@
## main #4031 +/- ##
==========================================
- Coverage 61.26% 60.26% -1.00%
==========================================
Files 2153 2054 -99
Lines 188367 176593 -11774
==========================================
- Hits 115403 106426 -8977
+ Misses 62224 60411 -1813
+ Partials 10740 9756 -984
Flags with carried forward coverage won't be shown. Click here to find out more.
🚀 New features to boost your workflow:
|
![seidroid[bot]](https://avatars.githubusercontent.com/in/2745135?s=60&v=4){kind=small}
There was a problem hiding this comment.
The diagnosis is right — pebble's Metrics() returns cumulative values, so Add-ing them each scrape double-counts — but the chosen remedy (counter → gauge + Record) introduces new problems: the shared package-level gauges are last-write-wins across the multiple mvcc.Database instances a node opens, the duration metrics now conflict in instrument kind with the identically-named histograms in the sibling pebbledb package, and two //nolint:gosec directives were removed while the offending conversions stayed. The repo already contains the correct fix pattern in pebbledb.PebbleMetrics (addDelta + a db attribute), which resolves all three.
Findings: 3 blocking | 4 non-blocking | 3 posted inline
Blockers
- None at the file/PR level.
- 3 blocking issue(s) flagged inline on specific lines.
Non-blocking
- [suggestion]
docker/monitornode/dashboards/cryptosim-dashboard.jsonstill queries the_totalnames (lines 6163, 8665, 8854, 9138, 13726, 13821). Those panels keep working for the stores served bypebbledb.PebbleMetrics, but the MVCC state-store/receipt-store contribution drops out of them. Keeping counter semantics viaaddDeltaavoids touching the dashboard at all; if the gauge direction is kept intentionally, the dashboard needs updating in this PR. - [suggestion] The
pebble_*names are a public monitoring contract consumed outside this repo. If the counter → gauge rename does land, it is worth calling out in the PR body's breaking-change note which external dashboards/alerts own the_totalseries, since neither implementation emits both names during a transition window. - 2 non-blocking pre-existing issue(s) listed below under pre-existing issues.
Pre-existing issues
- [suggestion]
mvcc's metric instruments (sei-db/db_engine/pebbledb/mvcc/metrics.go) are package-level singletons shared by everymvcc.Databasewith no per-DB attribute, while the already-gauge metrics (sstableCount,memtableCount,walSize,cacheSize) are recorded viaRecord. When two MVCC DBs are open these already overwrite each other. The PR extends the pattern rather than introducing it. - [suggestion]
sei-db/db_engine/pebbledb/mvcc/collectAndRecordMetricsandpebbledb.PebbleMetrics.scrapeare two independent implementations scraping the samepebble.Metricsinto the same instrument names, with the latter a strict superset. The duplication is what allows the two definitions to drift out of sync; there is no test on either side pinning the instrument kinds.
|
|
||
| // Compaction metrics - report raw counts | ||
| otelMetrics.compactionCount.Add(ctx, m.Compact.Count) | ||
| otelMetrics.compactionCount.Record(ctx, m.Compact.Count) |
There was a problem hiding this comment.
[blocker] Record on a shared gauge is last-write-wins, so this does not produce a correct series when more than one mvcc.Database is open.
otelMetrics is a package-level singleton and every mvcc.OpenDB starts its own collectMetricsInBackground (db.go:253). None of these recordings carry a per-DB attribute — only level on the two per-level ones. A node that runs the state store on the pebble MVCC backend and the receipt store on receiptBackendPebble (ledger_db/receipt/receipt_store.go:168) therefore has two collectors writing the same instrument with the same (empty) attribute set every 10s. With Add the exported value was at least the sum of both; with Record the single series alternates between the state store's cumulative total and the receipt store's, giving a sawtooth that is as unusable as the bug being fixed.
The repo already has the shape that solves both halves of this: pebbledb.PebbleMetrics converts pebble's cumulative counters to counter deltas via addDelta (pebble_metrics.go:1189) and attaches a db attribute precisely so multiple DBs get distinct series (pebble_metrics.go:19-25). mvcc already imports that package as pebbledbmetrics (db.go:28) and holds a *pebble.DB in db.storage, so pebbledbmetrics.NewPebbleMetrics(ctx, db.storage, <name>, interval) can replace this function and its duplicate instrument block outright — which also removes the instrument-kind conflict and the _total rename noted on metrics.go.
There was a problem hiding this comment.
indeed. pebble metrics were duplicated mvcc/metrics.go and pebble_metrics.go fixed 2b8b377
| metric.WithUnit("{count}"), | ||
| )), | ||
| compactionDuration: must(meter.Float64Histogram( | ||
| compactionDuration: must(meter.Float64Gauge( |
There was a problem hiding this comment.
[blocker] These names are not private to this package: pebbledb.PebbleMetrics registers the same instrument names on the same seidb_pebble meter, and its doc comment states that alignment is deliberate ("Instrument names match sei-db/db_engine/pebbledb/mvcc for dashboard compatibility", pebble_metrics.go:20). Both are live in one process — flatkv/store.go constructs pebbledb.New for the SC store while the SS store uses mvcc.OpenDB.
Two consequences:
pebble_compaction_durationandpebble_flush_durationare stillFloat64Histogramthere (pebble_metrics.go:306,:429). Registering them asFloat64Gaugehere means the same meter yields two conflicting instrument kinds under one name, and after Prometheus naming both land on the identical family namepebble_{compaction,flush}_duration_secondswith different types. Only one of them can be exposed, so one silently disappears from/metrics.- For
compaction_count/flush_count/*_bytes_*/cache_{hits,misses}the sibling counters keep the_totalsuffix while these gauges drop it, so one logical metric splits into two differently-named series depending on which store produced it.
Switching this file to the addDelta conversion used by the sibling keeps the instrument kinds and exported names identical across both implementations and still fixes the cumulative-double-counting bug.
| otelMetrics.sstableTotalSize.Record(ctx, levelMetrics.TablesSize, metric.WithAttributes(levelAttr)) | ||
| otelMetrics.compactionBytesRead.Add(ctx, int64(levelMetrics.TableBytesIn), metric.WithAttributes(levelAttr)) //nolint:gosec | ||
| otelMetrics.compactionBytesWritten.Add(ctx, int64(levelMetrics.TableBytesCompacted), metric.WithAttributes(levelAttr)) //nolint:gosec | ||
| otelMetrics.compactionBytesRead.Record(ctx, int64(levelMetrics.TableBytesIn), metric.WithAttributes(levelAttr)) |
There was a problem hiding this comment.
[blocker] The //nolint:gosec directives were dropped from these two lines but the uint64 → int64 conversions remain, so gosec G115 will now fire and make lint fails. Note lines 1430 and 1433 keep the directive for the same kind of conversion.
Either restore the directives, or clamp instead — pebbledb.uint64ToInt64Clamped (pebble_metrics.go:1182) is the local precedent, though it is unexported so it would need a small equivalent here.
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
yzang2019
left a comment
yzang2019
left a comment
There was a problem hiding this comment.
LGTM overall after fixing the blocker comments
4783f13 to
2b8b377
Compare
|
@yzang2019 i have slightly changed the scope of the PR based on the bot comments.
|
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.
❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
Reviewed by Cursor Bugbot for commit 08e13a0. Configure here.
| return nil, fmt.Errorf("failed to retrieve latest version: %w", err) | ||
| } | ||
|
|
||
| dbName := filepath.Base(dataDir) |
There was a problem hiding this comment.
DB metric names collide across stores
Medium Severity
dbName is filepath.Base(dataDir), so default store paths (state_store/cosmos/pebbledb, state_store/evm/pebbledb, ledger/receipt/pebbledb) all label series db="pebbledb". Gauges then overwrite each other and counters mix increments, so the new attribute does not separate SeparateEVMSubDBs or other in-process Pebble instances.
Additional Locations (1)
Reviewed by Cursor Bugbot for commit 08e13a0. Configure here.
2 participants
Some pebble db metrics are wrong.
The rootcause is this this
We keep adding
m.BlockCache.Hitsto thecacheHitsmetric (the one shown in the image) BUTm.BlockCache.Hitsis cumulative. Meaning we keep adding the cumulative value over and over, producing a wrong metric.I found this bug when checking
pebble_cache_hits_totalin a benchmark where I was not reading any data (just a read at startup. But this metrics kept increasing forever. Bug fixed (expected) left what we have (bug) right.This PR fixes it by setting the value pebble provides, the cumulative. Instead of adding it over and over.
countertogauge. Which means that the_totalsufix is removed from the following metrics. eg:pebble_compaction_count_totalpebble_compaction_count