You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@ignite.apache.org by "Sergey Kosarev (JIRA)" <ji...@apache.org> on 2018/10/08 11:04:00 UTC
[jira] [Commented] (IGNITE-8449) Avoid empty acquiring/releasing
checkpointReadLock for stale updates
[ https://issues.apache.org/jira/browse/IGNITE-8449?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16641657#comment-16641657 ]
Sergey Kosarev commented on IGNITE-8449:
----------------------------------------
I realized my changes may cause deadlocks as lock are taken in another way. So It looks like my idea about optimisation can't be realized.
> Avoid empty acquiring/releasing checkpointReadLock for stale updates
> --------------------------------------------------------------------
>
> Key: IGNITE-8449
> URL: https://issues.apache.org/jira/browse/IGNITE-8449
> Project: Ignite
> Issue Type: Bug
> Reporter: Sergey Kosarev
> Assignee: Sergey Kosarev
> Priority: Major
>
> we have in
> org.apache.ignite.internal.processors.cache.distributed.dht.GridDhtPartitionTopologyImpl#update(org.apache.ignite.internal.processors.cache.distributed.dht.preloader.GridDhtPartitionExchangeId, org.apache.ignite.internal.processors.cache.distributed.dht.preloader.GridDhtPartitionMap, boolean)
> currently we have something like this
> {code:java}
> ctx.database().checkpointReadLock();
> try {
> .......
> if (isStaleUpdate(cur, parts)) {
> ....
> return false;
> }
> .....
> }
> finally {
> ctx.database().checkpointReadUnlock();
> }
> {code}
> we'd better do not accquire those lock for isStaleUpdate == true branch. It can significantly decrease contention as this method can be hot (thousands invocations per minute) see also IGNITE-8226
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)