analyze-db-logs

Analyze DB Slow-Query Logs

Reads slow-query log files written by lib/database/slow_query_logging.dart and produces an actionable diagnosis grounded in this codebase's stack (Dart, Flutter, Drift, SQLite).

Invocation

The user provides the path to the log directory — typically a gitignored local copy they pulled off a device or simulator. Repository convention is ./logs/ at the repo root (already in .gitignore).

/analyze-db-logs ./logs/                       # most recent date in ./logs/
/analyze-db-logs ./logs/desktop                # platform-scoped subdir
/analyze-db-logs ./logs/mobile 2026-05-02      # specific date
/analyze-db-logs /tmp/from-device              # any absolute path

Treat $ARGUMENTS as <path> [optional date or window]. The first positional must be the path; refuse with a one-line prompt if it is missing — do NOT invent a default location, the user knows where their logs are. Date is optional and defaults to "the most recent file per stem found under that path".

Always print the resolved paths and line counts before analysing, so the user can confirm the right files are loaded.

Log file shapes

The SlowQueryInterceptor.fileReporter writes two daily files per platform under whatever documentsDirectoryPath/logs/ the running app resolved to. The user is responsible for copying them into the path they hand to this skill. Expect:

• slow_queries-YYYY-MM-DD.log — every query above the configured threshold (default 10ms; gated by SlowQueryLoggingGate.isEnabled).
• super_slow_queries-YYYY-MM-DD.log — duplicates of queries above the super-slow threshold (default 200ms), enriched with EXPLAIN QUERY PLAN rows under PLAN: and (when first-call stack capture is on) filtered application stack frames under STACK:.

Subdirectory layout

By convention this repo's ./logs/ contains one subdirectory per platform — typically desktop/ and mobile/ (sometimes ios/, android/). Each subdir holds its own daily files. When the path the user supplies is a directory that contains only subdirectories (no *.log files at its top level), treat each subdirectory as a separate platform-scoped scan and label every finding with the subdir name so the user can tell which device it came from.

If the user supplies a more specific path (e.g. ./logs/mobile/), respect that and don't go up a level. If they supply the parent and both desktop/ and mobile/ are present, analyse both and produce one report section per platform plus a brief cross-platform summary at the end.

Glob discovery rule of thumb:

<path>/{slow_queries,super_slow_queries}-*.log         # path is leaf
<path>/*/{slow_queries,super_slow_queries}-*.log       # path is parent

Pick the most recent date per (subdir, stem) pair unless a date arg is given.

Log line formats

A slow-query line is a single line:

2026-05-02T19:11:48.592 [db.sqlite] select 388.759ms args=0 SELECT * FROM ...

A super-slow entry is the same line plus indented continuation lines:

2026-05-02T19:11:48.592 [db.sqlite] select 388.759ms args=0 SELECT ...
  PLAN: 4|0|SEARCH journal USING INDEX idx_journal_browse (deleted=? AND type=?)
  PLAN: 84|0|USE TEMP B-TREE FOR ORDER BY
  STACK: #10     JournalDb.getAllDashboards (package:lotti/database/database.dart:3056:34)
  STACK: #11     dashboardsProvider.<anonymous closure> (package:lotti/features/dashboards/state/dashboards_page_controller.dart:20:14)

The interceptor filters STACK: lines to drop drift, dart-runtime, riverpod and the slow-query plumbing itself; only package:lotti/... frames remain.

Analysis procedure

1. Resolve the file set. Honor the platform argument; pick the most recent date per stem if no date is given. Tail-load the file (last ~2000 lines is usually enough — the interceptor appends, so the wave you care about is at the end). Print the resolved paths.

2. Bucket entries by query shape. Collapse per-row args; group by the normalized statement. For each bucket capture: count, p50/p95/max elapsed, the unique PLAN: shapes, and the unique STACK: heads (top app-code frame).

3. Diagnose against the known patterns below. For every finding, cite the specific log lines (timestamp + elapsed + statement prefix). Do not generalize — show the data.

4. Recommend the fix. Reference the existing seam (drift query, coalescer, transaction wrapper, index, etc.) and where in the code the change would land. If a fix is speculative, say so.

Known stall patterns to look for

These are battle-tested findings from this codebase. Apply them in order; the first match is usually the dominant cost.

1. Read waves — N queries reporting hundreds of ms but finishing in a tight burst

Signature: a cluster of 10–20+ queries whose elapsed is roughly identical (e.g., all in the 600–700ms band) but whose timestamps span only 10–30ms. Each query's plan is fine; the SQL itself is fast.

Diagnosis: the queries were queued behind something — typically a write transaction holding the writer lock, an ANALYZE on the boot path, or a slow beforeOpen hook. The wall-clock measurement starts when drift accepts the request, so queue wait shows up as "elapsed".

Confirm by:

• Check the elapsed band: a 100ms+ spread across the wave with near-identical finish timestamps (use the leading ISO timestamp, not the elapsed) means the queries unblocked together. The interceptor strips drift / dart-runtime frames so the original STACK: #5 DatabaseConnectionUser.doWhenOpened boilerplate is not visible in the log — infer the gate from the timing pattern, not from a frame name.
• Look at the line just before the wave for a long-running write or transaction.
• Boot waves often correlate with EntitiesCacheService.init firing Future.wait of definitions queries — the surviving STACK: heads for the wave will point at distinct controllers / repositories whose initial fetches all queued together.

Likely fixes:

• Move ANALYZE and other heavy work off the boot path (beforeOpen must return fast).
• Narrow transaction scopes (see pattern 4).
• Raise the read pool size only after confirming isolate-spawn cost is the bottleneck — bumping readPool adds isolate-spawn cost upfront.

2. N+1 reads — a cluster of single-id selects from one call site

Signature: many lines like SELECT * FROM "journal" WHERE "id" = ? AND "deleted" = ?, all from the same STACK: head, all reporting nearly identical elapsed times.

Diagnosis: a Future.wait(ids.map(byId)) or a per-row provider family fanning out single-id reads. Each call queues through the read pool independently.

Recurring offenders fixed in this branch:

• taskLiveDataProvider (FutureProvider.family per task) → solved by JournalDb._coalesceEntityById (microtask-coalesced bulk fetch).
• LinkedAiResponsesController._fetch → switched from Future.wait to journalRepository.getJournalEntitiesByIds(...).
• EditorStateService.init drafts loop → switched to journalEntitiesByIdsUnorderedAllPrivate(idList).

Likely fixes:

• Replace fan-out with a bulk drift query (journalEntitiesByIdsUnorderedAllPrivate, getJournalEntitiesForIdsUnordered, linksForEntryIds, etc.).
• For Riverpod families that genuinely need per-row instances, route through the existing entity-by-id microtask coalescer.

3. Drift MultiExecutor contention — reads inside a transaction run on the writer

Reminder: in drift, anything wrapped in db.transaction(() async { ... }) runs on the write connection — even pure select(...) calls inside the block. The read-pool isolates do NOT pick those up. So a transaction { read; read; …; write; commit } body serialises every read behind every other write that touches the same writer.

Confirm by: look at the STACK: head for a frame inside an apply / upsert / migration path. If the read appears to hit a fast plan but elapsed is high and other writes are visible nearby, the read was forced onto the writer.

Likely fixes:

• Pull pre-read / post-write side effects out of the transaction block.
• Only wrap statements that genuinely need atomicity together.
• Cross-DB writes (e.g., to sync_db, settings_db, agent_db, ai_config_db) cannot be atomic with JournalDb writes anyway — doing them inside a JournalDb.transaction only holds the journal writer lock for unrelated work.

4. Transaction scoping — broad wrappers around unrelated DB writes

Signature: the same wave shape as pattern 1, but the suspected "writer holding the lock" is a sync apply that's logging a SyncJournalEntity-shaped write while also awaiting a _sequenceLogService.recordReceivedEntry (sync_db) or other cross-DB work inside the same transaction { ... }.

Diagnosis: this codebase fixed exactly this in queue_apply_adapter.dart via _writesJournalDb(SyncMessage) — the adapter now wraps in JournalDb.transaction only for payload families that actually write to JournalDb tables (journal entity, entry link, entity definitions, outbox bundle, conservatively the backfill request/response paths). Theming, ai-config, agent entity/link/bundle writes bypass the wrapper because they target other databases.

_persistJournalEntity was also restructured: the pre-read diagnostic journalEntityById, the post-write _sequenceLogService.recordReceivedEntry (sync_db), and the entry-exists check now run outside the narrow journal transaction.

When inspecting new code: any new db.transaction(() async { ... }) that contains an await to a different database, a network call, or a filesystem write is a candidate for narrowing.

5. Wrong-index plans masquerading as slow SQL

Signature: a single query reporting hundreds of ms with a plan that includes USE TEMP B-TREE FOR ORDER BY, SCAN <table>, or an index match where the leading column is not the most selective predicate.

Recurring offenders already fixed in shipped code — verify these still match the current lib/database/database.drift, lib/database/sync_db.dart, and lib/database/database.dart before citing them. Treat the list as historical context, not an asserted current truth:

• task_priority_rank ordering with high-cardinality category IN (...) predicate. The fix shipped as a partial index named (at the time) idx_journal_tasks_status_priority_date; if it is still in database.drift, recommend it as the steady-state path, otherwise treat the symptom as an open issue.
• getBulkLinkedTimeSpans join over linked_entries. The fix shipped as a covering index idx_linked_entries_from_id_hidden_to_id.
• inbound_event_queue stats MIN(enqueued_at) SCAN. The fix shipped as idx_inbound_event_queue_status_enqueued plus idx_inbound_event_queue_status_due_lease.
• claimNextOutboxBatch SCAN from status = pending OR (status = sending AND updated_at < cutoff). The fix shipped as two indexed seeks merged in Dart.

If a query still matches one of these symptom shapes despite the named index existing, suspect stale stats first (recommend ANALYZE) before proposing a new index.

When inspecting new logs: if the plan is suboptimal, recommend running ANALYZE first (planner stats can drift); only after confirming with fresh stats should you propose a new index. Bad plans on a freshly-ANALYZEd DB are real index gaps.

6. WAL / OS-level factors

When patterns 1–5 don't explain a stall, consider:

• WAL checkpoint storms: a write that crosses wal_autocheckpoint (default 1000 pages, ~4MB) triggers a checkpoint that briefly takes a more aggressive lock. Bursty write workloads can stall reads at checkpoint boundaries.
• macOS sandboxed file locks: fcntl(F_FULLFSYNC) and BSD locks on iCloud-backed paths have shown long tails. Less likely on dev but worth flagging.
• Background isolate spawn (createInBackground(readPool: N)) — isolates may spawn on first use; the first boot wave can pay per-isolate setup cost (~50–200ms each). Subsequent app sessions should be much faster — first wave is the worst case.
• PRAGMA foreign_keys = ON runs per-connection; cheap but real.

7. Per-launch repair / housekeeping that shouldn't run every boot

Signature: an entry whose STACK: head points at a beforeOpen, self-heal, or migration helper — but the user just opened the app normally. With readPool: N, drift calls beforeOpen on every connection, so any repair work done there runs 1 + N times per launch.

Lessons from this branch:

• Per-launch ANALYZE was removed — stats persist in sqlite_stat1 and the v42 migration runs ANALYZE once on upgrade.
• The self-heal CREATE INDEX IF NOT EXISTS block was removed entirely — the recovery path it covered targeted an aborted-migration scenario that has not occurred in production.
• beforeOpen is now just PRAGMA foreign_keys = ON.

When you see boot-time housekeeping in a stack, ask: does this need to run on every launch, or once per upgrade?

Output format

Produce a concise report, in this order:

1. Inputs — paths read, line counts, date window.
2. Findings, each with: severity (high/medium/low), pattern matched (numbered above), evidence (specific log lines with timestamps), root-cause hypothesis, and the file/seam where the fix would land.
3. Recommended actions, ordered by impact. Cite specific call sites (lib/...:line) where possible.
4. Things you ruled out (and why), so the user can sanity-check.

Keep the report skimmable. If the log is mostly clean, say so — don't manufacture findings.

Guardrails

• Never propose schema bumps (schemaVersion++) without confirming the v42-style migration shape and asking the user. New indices that don't need a column change can land via migration; per-launch defensive code should not.
• Never propose a per-boot ANALYZE, CREATE INDEX IF NOT EXISTS loop, or other "self-heal on every open" — this codebase explicitly removed those.
• Don't assume an INDEXED BY hint solves a planner problem. The autoindex names (sqlite_autoindex_*_1) are not part of the public SQLite contract; recommend ANALYZE first.
• Match real seams: prefer adding fixes to existing patterns (_PendingEntityByIdWave for entity-by-id coalescing, _PendingLinksWave for to-id link batches, _writesJournalDb for per-payload transaction scoping) over inventing new ones.
• Show your work: every finding should cite specific log lines so the user can verify.