[F64116045]

What changes were proposed in this pull request?

SCM flushes the transaction buffer before serving a DB checkpoint.
If a checkpoint download overlaps Ratis transaction apply, flush() can commit the data for index N before the recorded transaction index is updated from N-1 to N. The checkpoint then contains data through N but records N-1.

This PR uses flushIfNeeded(0) to skip the flush while a transaction is being applied.
A slightly stale but consistent checkpoint is safe, as the follower can catch up through Ratis log replay.

Before

Ratis apply path                         gRPC checkpoint path
----------------                         --------------------

beginApplyingTransaction()

addToBuffer(dataForIndexN)
// latestTrxInfo is still N-1
                                         download()
                                           flush()
                                             putWithBatch(transactionInfo, N-1)
                                             commitBatchOperation()
                                           checkpoint:
                                             index N-1 + data through N
                                             inconsistent

updateLatestTrxInfo(N)
endApplyingTransaction()

After

Ratis apply path                         gRPC checkpoint path
----------------                         --------------------

beginApplyingTransaction()

addToBuffer(dataForIndexN)
// latestTrxInfo is still N-1
                                         download()
                                           flushIfNeeded(0)
                                             applyingTransactions > 0
                                             return
                                           checkpoint:
                                             index N-1 + data through N-1
                                             consistent

updateLatestTrxInfo(N)
endApplyingTransaction()

                                         next download()
                                           flushIfNeeded(0)
                                             commitBatchOperation()
                                           checkpoint:
                                             index N + data through N
                                             consistent

What is the link to the Apache JIRA?

https://issues.apache.org/jira/browse/HDDS-16452

How was this patch tested?

The regression test fails with the old flush() because data for index N is committed during transaction apply, and passes with this change.

CI: https://github.com/F64116045/ozone/actions/runs/40821330694

[jojochuang]

lgtm trivial fix

[jojochuang]

https://issues.apache.org/jira/browse/HDDS-16453 is the same issue that happens to StatefulServiceStateManagerImpl

[jojochuang]

CI is green in author's fork https://github.com/F64116045/ozone/actions/runs/40821330694/job/123277460627

[F64116045]

https://issues.apache.org/jira/browse/HDDS-16453 is the same issue that happens to StatefulServiceStateManagerImpl

Thanks for the heads-up! I'm working on HDDS-16453 too.