Fleet: roll out 3.5.0 -> 3.10.1; 96.7% of observed runs (2026-07-28 sweep) predate the parallel-queue fixes

[nanoDBA]

Problem

Version distribution across the 30-day window:

Version	Servers	Runs	Share
3.5.0.2026.04.23	40/40	1773	96.7%
3.8.1.2026.07.25	40/40	61	3.3%

Not true skew -- a single in-flight rollout. Every server ran 3.5.0 on its scheduled job for the whole window and picked up 3.8.1 on 2026-07-25 (1-2 runs each so far).

Consequence

96.7% of observed runs predate:

• 3.5.7 parallel-queue-leadership fix (v31y) -- the cause of the PARALLEL_COMPLETE half of the SILENT_NOOP findings
• the icjq effective_counter fix -- cause of PRIORITY_PASS_EMPTY on 31/40
• 3.5.4 @FirstTimeFullScanCapRows and 3.5.9 @AbortOnIntegrityError -- why UNUSED_FEATURES fires 40/40 for both

Note

The rollout does NOT fix the p0 lazy-mop-up CONTINUE bug, which is still present in current source. Sequence the rollout after that fix ships, or expect the COMPLETED-variant no-ops to persist.

Process gap

The effective_counter defect shipped as a silent point fix. A defect that silently disables the primary code path for ~3 months warrants a release advisory, not a changelog line.

Status (updated 2026-07-30)

v3.8.4 released (this epic's full remediation batch). Rollout target rescoped from 3.8.1 to 3.8.4. All fixes named above as rollout motivators are confirmed shipped as of v3.8.4, including the p0 lazy-mop-up CONTINUE fix this issue flagged as a blocker. Still an ops/deploy task, not a code fix -- needs a fresh diagnostic sweep to confirm actual fleet version distribution before this can be closed.

[nanoDBA]

Rollout artifacts ready: target moved to v3.10.1 (further fixes shipped since the last rescope -- including a mop-up discovery defect that had been silently broken since April and is directly relevant to this fleet's symptoms). A deploy/verify/rollback checklist with the operational changes to brief on (previously-false-green jobs will now legitimately fail; runs will do more real work; preset adaptive sampling; new monitoring surfaces) has been produced. Execution and the post-rollout confirmation sweep remain with ops.

[nanoDBA]

Maintainer reports a 3.8 build was deployed fleet-wide a few days ago (consistent with the sweep observing 3.8.1 arriving on all 40 servers on 2026-07-25). Rescoped as a 3.8.x -> 3.11.0 delta: the queue-leadership and effective_counter fixes are already live in production, which also explains the long-running-job monitoring alerts (runs stopped silently quitting early). Still outstanding on 3.8.1: the COMPLETED-variant silent-no-op fall-through (fixed 3.8.2), the mop-up discovery breakage (fixed 3.10.0), and the 3.9.0-3.11.0 improvements. The rollout checklist's pre-flight query confirms the exact installed build per instance, and the confirmation sweep requires the updated diagnostic proc on the servers first.

[nanoDBA]

Confirmed: the fleet build is 3.8.1 (2026-07-25). That means the P0 silent-queue-abandonment fall-through -- fixed in 3.8.2, three days after that build -- is still live in production, along with the missing failed-job grading for silent no-ops and the mop-up discovery breakage. Exposure is event-driven (job kills are rare on this fleet), but any kill or failover re-opens the hole until the queue drains. Recommendation: prioritize this rollout to v3.11.0.

[nanoDBA]

Release cut: v3.11.1 is now the deploy artifact for this rollout (supersedes v3.11.0; adds the plan-feedback state-filter fix from the SQL 2025 compatibility audit). All code work on the parent epic is complete -- this rollout and the post-deploy confirmation sweep are the only remaining items.