smartconnpool: unblock Close and preserve Setting on reopen

[arthurschreiber]

Description

Four small changes to the smartconnpool. The first three address shutdown stalls when the MySQL backend is unhealthy; the fourth fixes a long-standing silent loss of connection settings on reopen. Each is scoped to a specific code path; healthy pools see no behavior change.

1. tryReturnConn closes eagerly when the pool is over capacity

Whenever pool.active > pool.capacity — driven either by Close()/SetCapacity(0) taking capacity to 0 or by any SetCapacity(N) reducing it below the current active count — Recycled connections are closed immediately instead of being handed to whichever waiter is still in the waitlist. Without this, conns get passed waiter-to-waiter during the drain and the setCapacity loop's 10ms-sleep polls never observe a non-empty stack to close. The check is active > capacity rather than capacity == 0 so partial reductions get the same drain semantics as a full Close — SetCapacity(N → M) with 0 < M < N and a non-empty waitlist also gets a clean drain.

2. Pool-wide lifetime context, cancelled on Close

The pool now holds a context.Context (in a small heap-allocated lifetime struct, behind an atomic.Pointer) that is cancelled at the very start of CloseWithContext, before setCapacity(0).

The pointer indirection is deliberate: inlining context.Context (16B) + CancelFunc (8B) directly into ConnPool pushed the struct's size across a Go allocator size-class boundary and exposed a latent 16-byte alignment requirement of the lock-free stack's atomic-128 ops, causing SIGBUS on ARM64. Keeping the inline footprint at 8 bytes preserves the size class.

3. The idle worker, put(nil), and the maxLifetime put branch all use the lifetime ctx for their connect calls

Previously these three sites passed context.Background() to connect(...), so a Close() racing with an in-flight reopen was extended by the full db-connect-timeout-ms per blocked connect, completely ignoring PoolCloseTimeout. They now use pool.connectCtx(), so a Close unblocks any in-flight connect immediately. put stays synchronous on the caller's goroutine — the same shape as before this PR — but no longer holds up Close.

4. connReopen no longer silently drops the connection's Setting

connReopen used to read Setting() after replacing dbconn.Conn with a freshly-connected conn, so it always read nil and the conn silently migrated from settings[bucket] to clean on every reopen — both for the idle worker and for the maxLifetime path.

Fixing connReopen itself to unconditionally capture Setting before the replace broke the get()/getWithSetting error paths, which rely on the bare-reopen behavior to fall back to a settingless conn when ResetSetting fails. So connReopen now takes an explicit setting *Setting parameter: callers refreshing a conn in place (closeIdleResources and the maxLifetime branch of put) capture the previous Setting and pass it through; callers recovering from a ResetSetting failure pass nil for a bare reopen.

What this PR does not do

Under a sick backend during normal operation (not Close), Taint and Recycle still synchronously block on the connect attempt for the full backend connect timeout. That matches the behavior before this PR — fixing it requires moving the connect off the caller's goroutine, which has its own design trade-offs and is a separate conversation. The Close-time stalls are the immediate operational problem this PR addresses.

Related Issue(s)

None. Discovered during a deep review of the smartconnpool stall surface.

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

New and updated coverage:

• TestCloseDoesNotHandOffToWaiters — Close doesn't hand Recycled conns to waiters once capacity is 0.
• TestSetCapacityReductionDrainsWithWaiters — SetCapacity(N → M) with 0 < M < N and hold-forever waiters completes within its deadline; recycled conns are closed when active > capacity rather than handed off and stuck.
• TestIdleWorkerConnectCancelsOnClose — idle worker's connect unblocks when Close fires.
• TestTaintConnectCancelsOnClose — Taint's synchronous reopen unblocks when Close fires.
• TestRecycleMaxLifetimeReopenCancelsOnClose — the maxLifetime reopen in put unblocks when Close fires.
• TestTaintWakesWaiter and TestRecycleMaxLifetimeWakesWaiter — queued waiters are still served by the synchronous replacement.
• TestRecycleMaxLifetimePreservesSetting and TestIdleWorkerReopenPreservesSetting — reopened conns return to the same settings stack.
• TestIdleTimeoutDoesntLeaveLingeringConnection — tightened to account for the two-loop idle refresh path without racing the worker.
• TestStressCloseDuringReconnectStorm — stress-covers Close racing with blocked reconnects and verifies shutdown completes without leaking conns.
• TestStressWaiterStormDuringDrain — stress-covers queued waiters during Close drain and verifies returned conns are closed rather than handed off.

Existing TestCloseDuringWaitForConn × 50 iterations under -race continues to pass.

Backport justification

These are production-reliability fixes for vttablet shutdown against an unhealthy backend. With a dead/unreachable MySQL:

• The idle worker's connReopen, plus the put(nil) and maxLifetime branches of put, all blocked for the full backend connect timeout per in-flight connect, ignoring PoolCloseTimeout. Observed Close latency goes from the intended ~10s up to many minutes.
• Concurrent Recycles during Close get cycled waiter-to-waiter instead of landing on a stack the drain loop can see, adding more polling latency.
• Settings configured on connections silently disappear after every idle-timeout refresh and every conn rotation that goes through the reopen path, forcing repeated re-application.

A vttablet restart against a dead primary is exactly when fast shutdown matters most. The changes are scoped to the Close, drain, and reopen paths; healthy pools see no behavior change beyond the (long-broken) setting preservation now actually working.

Deployment Notes

One user-visible behavior change: connections being Recycled while the pool is over its configured capacity — during Close(), SetCapacity(0), or any SetCapacity(N) that reduces capacity below current active — are now closed rather than handed to queued waiters. Those waiters will return with ErrConnPoolClosed once Close() finishes signalling the close channel, or with their own ctx timeout otherwise.

AI Disclosure

This PR was implemented by Claude Code based on a deep review and step-by-step fix plan I directed.

[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

Two small fixes to smartconnpool.ConnPool to ensure Close() returns promptly when the MySQL backend is unhealthy: connections being recycled during a close are closed eagerly rather than handed to waiters, and the idle-timeout worker's in-flight connect is now cancelled at close start instead of blocking for the full backend connect timeout.

Changes:

• tryReturnConn now closes connections immediately when capacity has been driven to 0 (Close or explicit SetCapacity(0)), preventing waiter-to-waiter handoff loops that starve the setCapacity drain.
• Introduces a heap-allocated workerLifetime{ctx, cancel} (held behind atomic.Pointer to preserve ConnPool's size class and 16-byte alignment for the lock-free stack's atomic-128 ops) that is cancelled at the start of CloseWithContext; the idle worker uses this context for connReopen.
• Two new regression tests: TestCloseDoesNotHandOffToWaiters and TestIdleWorkerConnectCancelsOnClose.

Reviewed changes

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

File	Description
go/pools/smartconnpool/pool.go	Adds workerLifetime struct + atomic pointer, cancels it at Close start, uses pool worker ctx for idle-worker connReopen, and closes recycled conns eagerly when capacity is 0.
go/pools/smartconnpool/pool_test.go	Adds two regression tests covering the no-handoff-after-Close guarantee and the cancellation of the idle-worker reopen on Close.

[codecov]

Codecov Report

❌ Patch coverage is 90.47619% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 56.52%. Comparing base (70c7a72) to head (03e65e6).
⚠️ Report is 277 commits behind head on main.

Files with missing lines	Patch %	Lines
go/pools/smartconnpool/pool.go	90.47%	2 Missing ⚠️

❗ There is a different number of reports uploaded between BASE (70c7a72) and HEAD (03e65e6). Click for more details.

HEAD has 1 upload less than BASE

Flag	BASE (70c7a72)	HEAD (03e65e6)
	1	0

Additional details and impacted files

@@             Coverage Diff             @@
##             main   #20122       +/-   ##
===========================================
- Coverage   69.67%   56.52%   -13.16%     
===========================================
  Files        1614        8     -1606     
  Lines      216793      966   -215827     
===========================================
- Hits       151044      546   -150498     
+ Misses      65749      420    -65329     

Flag	Coverage Δ
partial	56.52% <90.47%> (?)

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.

[Copilot]

Pull request overview

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

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.

[mattlord]

LGTM!

One tiny nit / non-blocking cleanup: go/pools/smartconnpool/pool_test.go:1943-1946 still says the replacement conn is opened off the caller goroutine, but the final implementation is synchronous again. I’d update that comment to avoid confusing future us.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.

Comments suppressed due to low confidence (1)

go/pools/smartconnpool/pool.go:483

• In put(nil), the pool now attempts connNew(pool.connectCtx()) even when the pool is already closed (pool.close.Load()==nil). Since connectors are not required to honor ctx cancellation (e.g. tests’ newConnector ignores ctx), a late Recycle/Taint after Close can block in the user connector and even attempt opening new backend connections after shutdown. Restore the fast-path that decrements active and returns without calling connect when the pool is already closed.

	if conn == nil {
		// Taint, or Recycle on a closed conn: open a replacement so a
		// queued waiter can be served. We use the pool's lifetime context
		// so a Close in flight unblocks this connect, instead of waiting
		// for the backend's connect timeout.
		var err error
		conn, err = pool.connNew(pool.connectCtx())
		if err != nil {
			pool.closedConn()
			return
		}

[timvaillancourt]

It's too bad the context package doesn't have a "nice" pre-cancelled context

I just wanted to add an alternative in case it is somehow more efficent or preferable:

var alreadyCancelled, _ = context.WithDeadline(context.Background(), time.Time{})

[timvaillancourt]

LGTM @arthurschreiber. Added one alternative option if it's interesting to you

[arthurschreiber]

Added one alternative option if it's interesting to you

Thanks. I don't think this makes any difference either way, so I'll keep it as-is! 🙇‍♂️