reddycharan opened a new pull request #2166: Enhance deferLedgerLockReleaseOfFailedLedger in ReplicationWorker URL: https://github.com/apache/bookkeeper/pull/2166 Descriptions of the changes in this PR: **(Note: Working on testcases and should add more comments to the changes)** **Issue:** In the past, ReplicationWorker (RW) retrial logic is enhanced to backoff replication after threshold number of replication failures of a ledger. This is to help in a pathological situation where data (ledger/entry) is unavailable. But this is sub-optimal solution, since there is possibility that each RW can try recovering a ledger threshold number of times, before a RW defers ledgerLockRelease. Also each time a RW tries to recover it would read entry/fragment sequentially and writes to new bookies until it finds a missing entry (completely unavailable) before failing on replication of that ledger. This is done for each retrial and it bloats the storage and overreplication need to detect and delete it, which runs once a day by default. So because of this cluster can run out of storage space and may become RO cluster. Also this puts quite a bit of load on cluster in vain. **So the new proposal is to** - On each RW. remember the state in addition to the counter. State must include the entries which RW failed to read. - Counter and state must kept around in each RW node. And exponential backup should be used for deferLedgerLockReleaseOfFailedLedger - During next attempt by RW, try to read the failed entries which is noted in the state. Read must be successful before proceeding replicating. - With this model we avoid duplicate copies on each attempt. At the most each RW will create only one copy
---------------------------------------------------------------- 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. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
