[
https://issues.apache.org/jira/browse/IGNITE-21116?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17818768#comment-17818768
]
Roman Puchkovskiy commented on IGNITE-21116:
--------------------------------------------
Further analysis showed that the race between write intent cleanup and the
indexing of that write intent is not a problem. The index backfiller, when
encountering a write intent, will invoke the standard Write Intent Resolution
procedure that will either yield a committed version (which the backfiller will
index), or yield an indication that the transaction has been aborted (and the
backfiller will skip the write intent), or hang until the transaction gets
resolved (this state can never be achieved if some partitions are offline, but
the index will just hang in the BUILDING state which seems alright; if the
building node gets rebooted, the backfill process will be recovered, it will
hit the same write intent and hang once again).
> After index backfill has started, ignore attempts to finish a tx started
> before the index had appeared
> ------------------------------------------------------------------------------------------------------
>
> Key: IGNITE-21116
> URL: https://issues.apache.org/jira/browse/IGNITE-21116
> Project: Ignite
> Issue Type: Improvement
> Reporter: Roman Puchkovskiy
> Assignee: Roman Puchkovskiy
> Priority: Major
> Labels: ignite-3
> Fix For: 3.0.0-beta2
>
>
> Let's say a transaction tx1 had started before an index I1 has appeared. Then
> I1 switches to the BACKFILLING state, and only then the transaction tries to
> finish. This happens after IGNITE-21115 ensuring that all transactions
> started before I1 appearance by live coordinators have finished, so I1
> belongs to a dead/unknown coordinator.
> We cannot just commit/abort it as it will cause a race between cleaning up
> the Write Intents of tx1 and the Backfiller already working on I1. So we
> should actually leave the transaction state unchanged and don't return any
> response, so we effectively turn the transaction into an ABANDONED one. Its
> state will be resolved by the Backfiller as soon as it hits first Write
> Intent of this transaction and performs Write Intent resolution.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
