[
https://issues.apache.org/jira/browse/IGNITE-1553?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16295263#comment-16295263
]
Alexey Kuznetsov commented on IGNITE-1553:
------------------------------------------
[~agoncharuk] Currently we have a couple of CacheStore implementations :
GridCacheWriteBehindStore, and GridCacheLoaderWriterStore which don't support
sessionEnd() method(No-op implementation). This means that we cannot commit to
database as you described in ticket description above, instead data is
persisted right after calling write\delete etc.
Additionally, we cannot guarantee that custom CacheStore prepares data for
commit when write\delete etc. is called (and we cannot guarantee that data is
flushed when sessionEnd is called)
So, I would propose to optimize prepare step only if CacheStore is instance of
CacheAbstractJdbcStore or CacheJdbcBlobStore(only these stores guarantee
autocommit==false)
> Optimize transaction prepare step when store is enabled
> -------------------------------------------------------
>
> Key: IGNITE-1553
> URL: https://issues.apache.org/jira/browse/IGNITE-1553
> Project: Ignite
> Issue Type: Improvement
> Components: cache
> Affects Versions: ignite-1.4
> Reporter: Alexey Goncharuk
> Assignee: Alexey Kuznetsov
> Labels: important
>
> Currently entries are enlisted in a database transaction after grid
> transaction is in PREPARED state. We can do this in parallel in the following
> fashion (pseudo-code):
> {code}
> fut = tx.prepareAsync();
> db.write(tx.writes());
> fut.get();
> try {
> db.commit();
>
> tx.commit();
> }
> catch (Exception e) {
> tx.rollback();
> }
> {code}
> If this approach is applied, we should be able to reduce latency for
> transactions when write-through is enabled.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
