discovery: fix keyspaceState leak for keyspaces missing in localCell

[corbantek]

Description

Fixes a keyspaceState leak in KeyspaceEventWatcher for any keyspace whose SrvKeyspace does not exist in the local cell (common in multi-cell vtgate deployments where some keyspaces are only served from a subset of cells).

The leak grows by one orphan keyspaceState per healthcheck event for the missing keyspace. At fleet healthcheck rates (>1k/sec for an active keyspace) this manifests as steady RSS + CPU growth on vtgate over hours.

Root cause

1. processHealthCheck → getKeyspaceStatus allocates a fresh keyspaceState via newKeyspaceState.
2. newKeyspaceState calls WatchSrvKeyspace. srvtopo.watchEntry.addListener (in go/vt/srvtopo/watch.go) unconditionally appends the callback to entry.listeners and synchronously fires it with the cached NoNode. onSrvKeyspace sets kss.deleted = true and returns false.
3. newKeyspaceState then calls WatchSrvVSchema. onSrvVSchema always returns true, so its listener is never reaped from the listeners slice.
4. getKeyspaceStatus removes the kss from kew.keyspaces (the kss.deleted branch), but the SrvVSchema watcher still holds the orphan reference.
5. The next healthcheck for that keyspace repeats step 2, pinning another orphan. Forever.

Fix

Two pieces in go/vt/discovery/keyspace_events.go:

• Skip the WatchSrvVSchema registration when WatchSrvKeyspace synchronously reported NoNode (i.e. kss.deleted is already true after the call returns). With no onSrvVSchema listener attached, no orphan reference exists.
• Negative cache (missingKeyspaces, 30s TTL). Once a keyspace is confirmed missing in localCell, subsequent healthchecks short-circuit in getKeyspaceStatus before allocating a new keyspaceState. 30s balances suppressing the storm vs. recovering quickly when the keyspace becomes locally served.

Also adds a small isDeleted() accessor mirroring the existing isConsistent() helper from #16683 — both kss.deleted reads in newKeyspaceState and getKeyspaceStatus now take kss.mu rather than racing with the watcher goroutine's writes in onSrvKeyspace.

Regression test

TestKeyspaceEventWatcherMissingKeyspaceCache drives 100 lookups against a fake topo where WatchSrvKeyspace returns NoNode for the target keyspace, asserts:

• WatchSrvKeyspace is invoked exactly once within missingKeyspaceTTL.
• WatchSrvVSchema is never invoked (the original leak vector).
• After expiring the negative cache, the next lookup re-probes exactly once and WatchSrvVSchema still isn't called.

This passes under go test -race.

Related Issue(s)

Fixes #19992

Checklist

• "Backport to:" labels have been added if this change should be back-ported to release branches
• If this change is to be back-ported to previous releases, a justification is included in the PR description
• Tests were added or are not required
• Did the new or modified tests pass consistently locally and on CI?
• Documentation was added or is not required

The fix is request-path-relevant in any multi-cell vtgate deployment, so I'm requesting backport to release-22.0 (the only currently-supported release where I've personally validated the fix). If maintainers want it in older releases too, the patch should apply with no conflicts — the affected functions haven't changed substantively in several releases.

Deployment Notes

No flags, no migrations. Behavior change is limited to a 30s window where a SrvKeyspace that becomes available in localCell will be picked up at the next healthcheck after TTL expiry rather than the very next one. Note this is a fixed 30s window — unlike srv-topo-cache-ttl (default 1s, configurable), the negative cache here is hard-coded. The 30s value was chosen to suppress the per-event allocation storm without delaying recovery materially when a keyspace becomes locally served.

AI Disclosure

The code, tests, and PR/issue text were drafted with Claude Code; I (the author) reviewed each change and the diagnostic chain end-to-end before submitting.

[vitess-bot]

Review Checklist

Hello reviewers! 👋 Please follow this checklist when reviewing this Pull Request.

General

• Ensure that the Pull Request has a descriptive title.
• Ensure there is a link to an issue (except for internal cleanup and flaky test fixes), new features should have an RFC that documents use cases and test cases.

Tests

• Bug fixes should have at least one unit or end-to-end test, enhancement and new features should have a sufficient number of tests.

Documentation

• Apply the release notes (needs details) label if users need to know about this change.
• New features should be documented.
• There should be some code comments as to why things are implemented the way they are.
• There should be a comment at the top of each new or modified test to explain what the test does.

New flags

• Is this flag really necessary?
• Flag names must be clear and intuitive, use dashes (-), and have a clear help text.

If a workflow is added or modified:

• Each item in Jobs should be named in order to mark it as required.
• If the workflow needs to be marked as required, the maintainer team must be notified.

Backward compatibility

• Protobuf changes should be wire-compatible.
• Changes to _vt tables and RPCs need to be backward compatible.
• RPC changes should be compatible with vitess-operator
• If a flag is removed, then it should also be removed from vitess-operator and arewefastyet, if used there.
• vtctl command output order should be stable and awk-able.

[Copilot]

Pull request overview

Fixes a keyspaceState leak in KeyspaceEventWatcher when a keyspace has no SrvKeyspace in localCell (common in multi-cell vtgate deployments), preventing unbounded listener growth and associated RSS/CPU drift.

Changes:

• Avoids registering WatchSrvVSchema when WatchSrvKeyspace synchronously returns NoNode (keyspace missing in local cell).
• Adds a missingKeyspaces negative cache (30s TTL) to short-circuit repeated lookups for missing keyspaces.
• Adds a regression test asserting WatchSrvKeyspace is probed once per TTL and WatchSrvVSchema is never registered for missing keyspaces.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
go/vt/discovery/keyspace_events.go	Skips SrvVSchema watch on synchronous NoNode, adds negative cache + locked isDeleted() accessor to avoid races and repeated allocations.
go/vt/discovery/keyspace_events_test.go	Adds regression test covering missing-keyspace negative caching and ensuring no SrvVSchema watch registration occurs.

[codecov]

Codecov Report

❌ Patch coverage is 96.15385% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 67.98%. Comparing base (70c7a72) to head (098f2b8).
⚠️ Report is 258 commits behind head on main.

Files with missing lines	Patch %	Lines
go/vt/discovery/keyspace_events.go	96.15%	1 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##             main   #19993       +/-   ##
===========================================
- Coverage   69.67%   67.98%    -1.69%     
===========================================
  Files        1614       12     -1602     
  Lines      216793     1846   -214947     
===========================================
- Hits       151044     1255   -149789     
+ Misses      65749      591    -65158     

Flag	Coverage Δ
partial	67.98% <96.15%> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[mattlord]

Thanks, @corbantek !

There's a linter error (modernize) that you will need to address. Beyond that, I only have a tiny nit in that the deployment note says this is equivalent to existing srvtopo cache TTLs, but srv-topo-cache-ttl defaults to 1s and is configurable, while this adds a fixed 30s negative-cache window. The 30s tradeoff seems reasonable, but I’d make that wording a little more explicit.

[corbantek]

Thanks @mattlord, this was a crazy one here at HubSpot when we starting deploying Vitess 22 to our larger environments with 2000+ keyspaces--vtgates were OOMing all over the place 😄

Can we also mark this one for backport to 22+ if possible?

[mattlord]

@corbantek v22 is EOL as of this past week so we won't backport it that far (no more releases anyway). But we can backport it to v23.

[corbantek]

Good to know, we were waiting for v24 to officially release and then we were going to move our upgrade target to v23

[corbantek]

Is there anything waiting on me for this fix to merge?

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

Comments suppressed due to low confidence (1)

go/vt/discovery/keyspace_events.go:660

• The error check here uses the function parameter err instead of the local kerr returned from getMoveTablesStatus, so failures from getMoveTablesStatus can be silently ignored (or logged based on an unrelated incoming watcher error). This should check kerr != nil to reflect the call you just made.

	var kerr error
	if kss.moveTablesState, kerr = kss.getMoveTablesStatus(vs); err != nil {
		log.Error(fmt.Sprintf("onSrvVSchema: keyspace %s failed to get move tables status: %v", kss.keyspace, kerr))
	}

[mhamza15]

LGTM! I've approved, but I also left a comment for a change that wasn't introduced in your PR but could be good to have anyway.

[mhamza15]

Also I think you'll need to merge in main to get this job to pass: https://github.com/vitessio/vitess/actions/runs/25869133670/job/76028121188?pr=19993

[mattlord]

We should check kerr from getMoveTablesStatus in go/vt/discovery/keyspace_events.go:657-660. onSrvVSchema stores the getMoveTablesStatus error in kerr, but checks the watcher callback parameter err instead. Normal SrvVSchema updates have err == nil, so topo/shard lookup failures can be silently treated as a successful {MoveTablesNone, Unknown} update, masking MoveTables state during a topo blip. Please check kerr != nil and avoid updating moveTablesState from the error result.

We should reap deleted listeners even when SrvVSchema payload is nil in go/vt/discovery/keyspace_events.go:641-655. The new deleted-state reaper sits after if vs == nil { return true }. If onSrvKeyspace has marked the keyspace deleted and the SrvVSchema watcher later invokes this callback with a nil value/error, the callback keeps returning true and the orphan listener survives. Since deleted should always mean this listener is done, check kss.deleted before the nil fast path and add a nil-payload regression.