ensemble change may cause pendingAddOps of ledgerHandle to be unable to be resent, and the pulsar partition may become unavailable. #4459
Description
BUG REPORT
Describe the bug
When restarting bookie or in a high-traffic back pressure scenario, ensemble change will occur. If the value of newEnsemble is exactly the same as origEnsemble, replaced = EnsembleUtils.diffEnsemble(origEnsemble, newEnsemble) returns an empty HashSet, calling the unsetSuccessAndSendWriteRequest method will not be able to resend the request, resulting in all ledger requests being blocked.
To Reproduce
Steps to reproduce the behavior:
- restart bookie
- multiple ensemble changes occurred (I don't know why the bookie lists of the two ensemble changes are exactly the same):
2024-06-26 16:22:52.0453 [BookKeeperClientWorker-OrderedExecutor-28-0] INFO org.apache.bookkeeper.client.LedgerHandle - New Ensemble: [10.199.102.18:3181, 10.200.48.84:3181] for ledger: 320092
2024-06-26 16:22:53.0542 [BookKeeperClientWorker-OrderedExecutor-28-0] INFO org.apache.bookkeeper.client.LedgerHandle - New Ensemble: [10.199.102.18:3181, 10.200.48.84:3181] for ledger: 320092
Expected behavior
Partition messages can be sent normally
Screenshots
Additional context
Add any other context about the problem here.