staging-v22.2.18: release-22.2: rangefeed: fix premature checkpoint due to intent resolution race #118633
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
![blathers-crl[bot]](https://avatars.githubusercontent.com/in/59700?s=60&v=4){kind=small}
It was possible for rangefeeds to emit a premature checkpoint, before all writes below its timestamp had been emitted. This in turn would cause changefeeds to not emit these write events at all. This could happen because `PushTxn` may return a false `ABORTED` status for a transaction that has in fact been committed, if the transaction record has been GCed (after resolving all intents). The timestamp cache does not retain sufficient information to disambiguate a committed transaction from an aborted one in this case, so it pessimistically assumes an abort (see `Replica.CanCreateTxnRecord` and `batcheval.SynthesizeTxnFromMeta`). However, the rangefeed txn pusher trusted this `ABORTED` status, ignoring the pending txn intents and allowing the resolved timestamp to advance past them before emitting the committed intents. This can lead to the following scenario: - A rangefeed is running on a lagging follower. - A txn writes an intent, which is replicated to the follower. - The closed timestamp advances past the intent. - The txn commits and resolves the intent at the original write timestamp, then GCs its txn record. This is not yet applied on the follower. - The rangefeed pushes the txn to advance its resolved timestamp. - The txn is GCed, so the push returns ABORTED (it can't know whether the txn was committed or aborted after its record is GCed). - The rangefeed disregards the "aborted" txn and advances the resolved timestamp, emitting a checkpoint. - The follower applies the resolved intent and emits an event below the checkpoint, violating the checkpoint guarantee. - The changefeed sees an event below its frontier and drops it, never emitting these events at all. This patch fixes the bug by submitting a barrier command to the leaseholder which waits for all past and ongoing writes (including intent resolution) to complete and apply, and then waits for the local replica to apply the barrier as well. This ensures any committed intent resolution will be applied and emitted before the transaction is removed from resolved timestamp tracking. Epic: none Release note (bug fix): fixed a bug where a changefeed could omit events in rare cases, logging the error "cdc ux violation: detected timestamp ... that is less or equal to the local frontier". This can happen if a rangefeed runs on a follower replica that lags significantly behind the leaseholder, a transaction commits and removes its transaction record before its intent resolution is applied on the follower, the follower's closed timestamp has advanced past the transaction commit timestamp, and the rangefeed attempts to push the transaction to a new timestamp (at least 10 seconds after the transaction began). This may cause the rangefeed to prematurely emit a checkpoint before emitting writes at lower timestamps, which in turn may cause the changefeed to drop these events entirely, never emitting them.
Epic: none Release note: None
blathers-crl
bot
force-pushed
the
blathers/backport-staging-v22.2.18-118477
branch
from
cb22102 to
9c9ac22
Compare
blathers-crl
bot
added
blathers-backport
|
This change is |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Backport 2/2 commits from #118477 on behalf of @erikgrinaker.
/cc @cockroachdb/release
Only the 2 last commits should be reviewed here, the earlier commits in this PR are from the following prerequisite backports:
TestProcessorContextCancellation#118470PushTxnResponse.AmbiguousAbort#118472BarrierRequest.WithLeaseAppliedIndex#118474This backport includes new RPC protocol additions that are necessary to fix the bug. See discussion in 23.2 backport: #118413.
This backport had no significant conflicts.
Backport 2/2 commits from #117612. Also pull in parts of #118469 and #118265.
Release justification: fixes a bug which could cause changefeeds to omit events in some scenarios.
/cc @cockroachdb/release
It was possible for rangefeeds to emit a premature checkpoint, before all writes below its timestamp had been emitted. This in turn would cause changefeeds to not emit these write events at all.
This could happen because
PushTxnmay return a falseABORTEDstatus for a transaction that has in fact been committed, if the transaction record has been GCed (after resolving all intents). The timestamp cache does not retain sufficient information to disambiguate a committed transaction from an aborted one in this case, so it pessimisticallyassumes an abort (see
Replica.CanCreateTxnRecordandbatcheval.SynthesizeTxnFromMeta).However, the rangefeed txn pusher trusted this
ABORTEDstatus, ignoring the pending txn intents and allowing the resolved timestamp to advance past them before emitting the committed intents. This can lead to the following scenario:This patch fixes the bug by submitting a barrier command to the leaseholder which waits for all past and ongoing writes (including intent resolution) to complete and apply, and then waits for the local replica to apply the barrier as well. This ensures any committed intent resolution will be applied and emitted before the transaction is removed from resolved timestamp tracking.
Resolves #104309.
Epic: none
Release note (bug fix): fixed a bug where a changefeed could omit events in rare cases, logging the error "cdc ux violation: detected timestamp ... that is less or equal to the local frontier". This can happen if a rangefeed runs on a follower replica that lags significantly behind the leaseholder, a transaction commits and removes its transaction record before its intent resolution is applied on the follower, the follower's closed timestamp has advanced past the transaction commit timestamp, and the rangefeed attempts to push the transaction to a new timestamp (at least 10 seconds after the transaction began). This may cause the rangefeed to prematurely emit a checkpoint before emitting writes at lower timestamps, which in turn may cause the changefeed to drop these events entirely, never emitting them.
Release justification: