[
https://issues.apache.org/jira/browse/IGNITE-17615?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Alexander Lapin updated IGNITE-17615:
-------------------------------------
Epic Link: IGNITE-21221 (was: IGNITE-21174)
> Close local cursors on primary replica expiration
> -------------------------------------------------
>
> Key: IGNITE-17615
> URL: https://issues.apache.org/jira/browse/IGNITE-17615
> Project: Ignite
> Issue Type: Improvement
> Reporter: Sergey Uttsel
> Assignee: Vladislav Pyatkov
> Priority: Major
> Labels: ignite-3
> Time Spent: 10m
> Remaining Estimate: 0h
>
> h3. Motivation
> According to our tx protocol, it’s impossible to commit a transaction if any
> of the enlisted primary replicas have expired. It also means that there’s no
> sense in preserving tx related volatile state such as locks and cursors. Pay
> attention, that it’s still useful to preserve txnState in the
> txnStateLocalMap because it will ease write intent resolution procedure.
> Locks release on primary replica expiration was already implemented, so this
> ticket is only about closing cursors on primary expiration.
> h3. Definition of Done
> * On primary replica expiration all RW-scoped cursors are closed.
> h3. Implementation Notes
> 1.In
> `org.apache.ignite.internal.table.distributed.replicator.PartitionReplicaListener#onPrimaryExpired`
> we release all tx locks
> {code:java}
> futs.add(allOf(txFuts).whenComplete((unused, throwable) ->
> releaseTxLocks(txId)));
> {code}
> Seems reasonable to reuse same event to close the cursors. Worth mentioning
> that given action should be asynchronous. I believe that we may do the
> cursors close in partition striped pool. See StripedThreadPoolExecutor for
> more details. Another option here is to introduce special dedicated cleanup
> thread and use it instead. That will be a part of TX Resourse Cleanup design.
> 2. So, that was about when to close, let’s clarify what to close. Seems that
> it’s trivial. We have
> `org.apache.ignite.internal.table.distributed.replicator.PartitionReplicaListener#cursors`
> right in partition replica listeners. We even have corresponding helper
> method
> `org.apache.ignite.internal.table.distributed.replicator.PartitionReplicaListener#closeAllTransactionCursors`
>
> All in all, seems that it's required to substitute
>
> {code:java}
> futs.add(allOf(txFuts).whenComplete((unused, throwable) ->
> releaseTxLocks(txId)));{code}
> with
> {code:java}
> futs.add(allOf(txFuts).whenComplete((unused, throwable)
> -> {
> releaseTxLocks(txId);
> try {
> closeAllTransactionCursors(txId);
> } catch (Exception e) {
> LOG.warn("Unable to close cursor on primary
> replica expiration", e);
> }
> }));{code}
> Tests are trickey though.
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
