You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Lite window start: the FinOps High Impact read holds the store read lock for over 12 s, so database-state maintenance skips every server #9233
Starting Lite with its window (3.13.0-nightly.20261121.713, a small test install of SQL Server 2022 servers, a store of about half a gigabyte) loads the FinOps panels at once. About 20 s after the start the log shows the store write lock out of reach for its whole 5 s budget:
InvestigationSweepService: "Investigation sweep failed ... Could not acquire the local database write lock ... within 5 s. Held by LocalDataService.FinOps.Workload.GetExpensiveQueriesAsync (read lock, held 4 s), GetHighImpactQueriesAsync (read lock, held 3.9 s), GetTopResourceConsumersAsync (read lock, held 2 s) and 6 more."
GetDatabaseStateDeviationsAsync, one line per server: "skipped this cycle's database-state maintenance: could not acquire the store write lock within 5s (Held by LocalDataService.FinOps.Workload.GetHighImpactQueriesAsync (read lock, held 12.5 s) ... QueryStoreSliceRepairService.RewriteArchiveFileAsync (read lock, held 9.2 s) ...". By then the High Impact read had held its read lock for 12.5 to 13.8 s.
A later sweep failure, about 14 minutes into the run, names the same read again with nothing else. The FinOps tab had not been opened by hand then, so the read also runs on its own (a timer or the tab's refresh). A headless Lite on the same store does not log these lines.
#9207 (for #9196) sizes a writer's park from recent read holds, and says plainly that a read held longer than 80 percent of the writer's budget cannot be outwaited. A single 12 s High Impact read is that case, so #9207 does not cover it.
Proposed
Bound how long the High Impact read holds the read lock (read in slices, or from a snapshot copy), so no single panel read holds it for longer than a writer's budget.
Do not run the FinOps panel loads at window start before the first collection cycle has written; start them after it, or one at a time.
When database-state maintenance skips, log one line per cycle naming all servers, not one identical line per server.
Tests
A windowed start against a seeded store of that size: no "could not acquire the store write lock" line in the first minute.
The FinOps loads wait for the first cycle's completion (a unit test on the start order).
A read that holds the read lock for 15 s while database-state maintenance asks for the write lock: maintenance gets it within its 5 s budget, or the read yields.
Activity
Open the skip-log window at the first skip of a burst and leave it fi…
Count skips instead of timing them in the skip-log window test (#9233)
Merge pull request #9235 from erikdarlingdata/fix/9233-high-impact-re…
erikdarlingdata commentedon Nov 21, 2026
Claude posting for Erik Darling. Fixed by #9235, merged to dev. It ships in the next release.