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
A retry after a conflict at commit keeps its own @createdTime over a concurrent creator's #3975
A PUT or PATCH that creates a record stamps @createdTime in _writeUpdate's validate (resources/Table.ts, stampRecordTimes), from the commit base it saw then. If its commit then conflicts at commit (ConflictKind.atCommit, or any retry whose log batch is already written) because a concurrent writer created the same record first, the retry reloads that base but never re-runs validate. The retry stores the record with its own @createdTime, which overwrites the creator's. Nothing reports it, and the value stays wrong.
Why it isn't a one-line restamp
The attempt's transaction-log batch is already durable and carries the validate-time times. A retry that keeps that batch must store what the batch says, or the local record and what replication ships disagree. A fix has to re-log the batch when the reloaded base changes the stamped values. That interacts with the retry cap and dedup accounting from #3928.
The refused-before-log case (relogOnRetry), where the batch is written fresh, is fixed separately by the PR that links this issue.
Repro sketch
unitTests/resources/abortMarkersAdoption.test.js's conflictAtCommit helper, on a table with assignCreatedTime: the competitor creates the unseeded id with an older timestamp; mine reads (nothing), then PUTs. Expect created to be the competitor's; it is mine's first-attempt time.
Folded into #3890. It is the same mechanism: a conflict retry never re-runs validate and stampRecordTimes, so the retry keeps its own @createdTime over the concurrent creator's. One constraint from this issue must carry over to that fix: when the attempt's transaction-log batch is already durable (ConflictKind.atCommit), a restamp must re-log the batch, or the stored record and what replication ships will disagree. That re-log interacts with the retry cap and dedup accounting from #3928.
🤖 Claude Opus 5.5, closed for @kriszyp after the 2026-11-09 triage pass.
Activity
kriszyp commentedon Nov 10, 2026
Folded into #3890. It is the same mechanism: a conflict retry never re-runs validate and
stampRecordTimes, so the retry keeps its own@createdTimeover the concurrent creator's. One constraint from this issue must carry over to that fix: when the attempt's transaction-log batch is already durable (ConflictKind.atCommit), a restamp must re-log the batch, or the stored record and what replication ships will disagree. That re-log interacts with the retry cap and dedup accounting from #3928.🤖 Claude Opus 5.5, closed for @kriszyp after the 2026-11-09 triage pass.