Fix review: H-L8-04 and H-L8-06 (hygiene)

From mnema-talk v3.3.13, rendered on this site.

Independent reviewer; I wrote neither fix (I was the phase 2 refuter of both findings). Both branches are based on msch128/hunt-lane8 at b5e4195d6. Each was reviewed in its own temporary worktree, removed afterwards. Nothing was pushed.

A. msch128/fix-h-l8-04 (f23a8569b)

Verdict: MERGE.

The change: setIdle (internal/ws/presence.go:66-71) deletes active[user] when the decrement reaches zero, the same way forget already did (internal/ws/lifecycle.go:345-349). Two tests are flipped from inverse assertions into guards: TestIdleThenOfflineLeavesNoActiveEntry, and TestHubConnectTypeDisconnectCycleReturnsToBaseline, which no longer lists active as a known leak.

Cause removed: yes. The invariant is "active[u] is the number of registered clients of u that are not idle, and there is no entry at zero". It now holds on every path.

  • Every writer and reader of the map, at the fix commit:
    • register (handshake.go:93) increments, so a missing entry becomes 1;
    • forget (lifecycle.go:346-347) decrements only for a client that is not idle and deletes at zero;
    • setIdle (presence.go:69-73) decrements and deletes, or increments;
    • the only production reader is statusLocked (presence.go:27). It reads > 0, so a missing entry and a zero entry mean the same thing there. Nothing used the presence of a zero key. The tests that read the map (limits_test.go:97-119, greeting_order_test.go:31-43) still pass, and greeting_order_test.go:43 already expected the key to be absent.
  • c.idle is written only in setIdle (presence.go:65), under h.mu, and only for a registered client (presence.go:60-62). So the counter cannot go negative, and forget cannot decrement a client that setIdle has already counted down.
  • Lock: the delete is inside the same safego.Locked(&h.mu, ...) closure as before.
  • My own situations, in a scratch test that is not committed: 200 seeds of 60 random steps for one account, mixing new connections, idle, wake, close, and setIdle after close. After every step it checks that the entry equals the number of registered non-idle clients, that there is never an entry at zero, and that the status is offline, online or away to match. The map is empty at the end. Wake after deletion, two idle tabs where one wakes, and a late setIdle on a closed client are all covered. It passed under -race.

What I ran:

  • go test -race ./internal/ws/: ok (62 s).
  • Revert check: with presence.go at f23a8569b^ both flipped tests fail. One reports "hub.active = 301 after the cycle, baseline 1", the other "offline account keeps active[user] = 0".
  • CYCLE_DEEP=1 cycle tests: 20 000 connections, active back to 0; voice cycle 20 000 joins, ok.
  • gofmt and go vet are clean. presence.go has 239 lines. The commit type is fix, with no attribution lines.

Findings: none blocking.

  • NOTE: the fix adds no cost on a hot path; setIdle runs once per idle change.

B. msch128/fix-h-l8-06-hygiene (90fee2502)

Verdict: MERGE AFTER FIXES. Both fixes are small; neither is a defect in the behaviour.

The change:

  • A dropped presence audit event now increments metrics.WSDropped.With("audit_presence") (internal/server/audit_presence.go:42) before the existing warning.
  • TestAuditPresenceNeverBlocksTheHub now asserts the count.
  • doc/moderation.md:517 and doc/operations.md:1518 describe the drop and the label (operations.md also adds the voice_churn label, which was already in use and undocumented).
  • The comment in metrics.go lists the reasons.

What I checked:

  • The drop itself is unchanged. The select / default branch, the queue size of 1024 and the log line are the same; the only addition is one counter increment on the drop path.
  • Cardinality: there is one constant label value, so one new series. Vec is capped at maxSeries anyway (metrics.go:63-64). With costs one read-locked map lookup, and only when an event is dropped.
  • Revert check: with audit_presence.go at 90fee2502^, the unit test fails with "dropped presence events counted = 0, want 1 to 1976". With the fix it passes 3 times in a row. Its upper bound is safe: the writer's context is cancelled before the burst, so at most 1024 events wait and the writer can only take some.
  • Docker integration test TestAuditPresenceBurstLosesJoinsPastTheQueue (-tags=integration): it passes. 3000 joins, 1025 recorded, 1975 lost, writer about 840 events/s in this run. I also added a temporary local line that compared the counter with the loss: counter delta 1975, lost 1975. The metric counts exactly the rows that are missing.
  • "the server log keeps each" join and leave: true in the default configuration, by reading the code.
    • The only Presence(..., true) call (voice_join.go:165, through announceVoiceJoin at voice_join.go:99) comes directly after c.logger().Info("voice join", ...) at voice_join.go:98, on the same path. The join line is also written for a rejoin, which is not audited: the log is a superset.
    • The only Presence(..., false) call (voice_presence.go:97, in endPresence) comes after log.Info("voice leave", ...) under the same wasIn condition.
    • Both lines are at Info level. LOG_LEVEL is empty by default, which means info in production (cmd/server/main.go:335-351). With LOG_LEVEL=warn or error the lines are not written, and how long they are kept is up to the operator's log retention.
  • gofmt, go vet with and without -tags=integration, and scripts/check-english-only.mjs are clean. The commit type is fix, with no attribution lines.

Findings:

  • SHOULD-FIX: doc/moderation.md:517. "the server log keeps each" is true only at the default log level, and "keeps" claims a retention the server does not control. Suggested wording: "the server log writes each one at the default log level (voice join / voice leave)".
  • SHOULD-FIX: internal/server/audit_presence_burst_integration_test.go, header comment. It still says the queue drops "with a log line and no metric", which is false after this commit. Update the comment, and add the check I ran: the counter delta equals burst - last. That turns the new doc sentence ("counts the misses") into a tested property.
  • NOTE: the counter does not count entries lost for other reasons: a failed insert (logged by RecordOrLog, internal/audit/audit.go:104-107) or a failed DM lookup (deliberately not recorded, audit_presence.go writePresence). "counts the misses" in moderation.md means queue misses; precise enough, but "counts the entries dropped from the queue" would be exact.
  • NOTE: the admin Ops panel shows ws_dropped as the total of all reasons (internal/server/system_ops.go:107, WSDropped.Total()), labelled "Events dropped (slow or flooding clients)" (web/src/i18n/en.json:1967) and marked as a warning when it is above 0. An audit queue drop now shows there under a label that blames clients. voice_churn already did this before this branch, so the problem is older than this branch. Widen the label in a follow-up, or keep the audit reason out of the panel total.

Not checked

  • No browser or UI run; the Ops panel finding is from reading the code.
  • No production database rate; the integration test runs on a local throwaway database.

Something wrong on this page? Tell us.