codelipenghui opened a new pull request, #24855:
URL: https://github.com/apache/pulsar/pull/24855
## Summary
Fixes a critical race condition where ledger trimming could delete ledgers
while cursors still pointed to them, causing:
- Negative backlog values from `getNumberOfEntriesInBacklog(isPrecise=false)`
- Cursors pointing to non-existent ledgers after topic reload
- `messagesConsumedCounter > entriesAddedCounter` inconsistency
## Root Cause
The issue occurs when:
1. Messages are acknowledged → cursor position advances in memory immediately
2. Cursor state persists to BookKeeper asynchronously (can be slow)
3. Ledger trimming runs during the persistence delay, using the in-memory
cursor position
4. Ledgers get deleted before cursor state is durably saved
5. On topic reload, cursor reverts to old persistent position pointing to
deleted ledgers
## The Fix
Changed `maybeUpdateCursorBeforeTrimmingConsumedLedger()` in
`ManagedLedgerImpl.java:2704-2705` to use the persistent cursor position
instead of the in-memory position:
```java
// Before:
Position lastAckedPosition = cursor.getMarkDeletedPosition();
// After:
Position lastAckedPosition = cursor.getPersistentMarkDeletedPosition() !=
null
? cursor.getPersistentMarkDeletedPosition() :
cursor.getMarkDeletedPosition();
```
This ensures ledgers are only deleted after cursor advancement has been
durably persisted to BookKeeper.
## Test Coverage
Added `testCursorPointsToDeletedLedgerAfterTrim()` in
`ManagedLedgerTest.java` which:
- Simulates BookKeeper persistence delay (30 seconds)
- Acknowledges messages asynchronously during the delay
- Triggers ledger trimming
- Verifies ledgers are NOT trimmed when persistent position hasn't advanced
yet
## Verification
Without the fix, the test fails because:
- First ledger gets trimmed even though persistent cursor still points to it
- Creates the exact race condition seen in production
With the fix, the test passes because:
- Trimming respects the persistent cursor position
- Ledgers are only deleted after cursor state is durably saved
🤖 Generated with [Claude Code](https://claude.com/claude-code)
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
