[
https://issues.apache.org/jira/browse/IGNITE-21047?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17812648#comment-17812648
]
Maksim Zhuravkov commented on IGNITE-21047:
-------------------------------------------
1. It seems that there are only few places where such optimization can be
applied w/o additional changes to other operators.
Changing ExecutionContext::execute to run in the caller thread a few times,
before attempting to schedule cases, both NPEs and incorrect results when
running SQL logic tests (sqllite/select with sub queries) - See attached patch.
2. Changing push to accept a batch of rows: most operators either pass-through
(like project, that calls the next operator w/o any additional scheduling)
or do batching prior to schedule a next run, meaning that some sort of
batching is already in place.
> Sql. Avoid spamming execution tasks when possible
> -------------------------------------------------
>
> Key: IGNITE-21047
> URL: https://issues.apache.org/jira/browse/IGNITE-21047
> Project: Ignite
> Issue Type: Improvement
> Components: sql
> Reporter: Konstantin Orlov
> Assignee: Maksim Zhuravkov
> Priority: Major
> Labels: ignite-3
> Fix For: 3.0.0-beta2
>
> Attachments:
> IGNITE-21047__Sql__Avoid_spamming_execution_tasks_when_possible.patch
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> Need to revise all usages of
> {{org.apache.ignite.internal.sql.engine.exec.ExecutionContext#execute}} and
> check whether spawning a new task is legit in every particular case or it's
> better to do the work right now.
> For example, lets take a look at
> {{org.apache.ignite.internal.sql.engine.exec.rel.ScanNode#request}}:
>
> {code:java}
> @Override
> public void request(int rowsCnt) throws Exception {
> assert rowsCnt > 0 && requested == 0 : "rowsCnt=" + rowsCnt + ",
> requested=" + requested;
> checkState();
> requested = rowsCnt;
> if (!inLoop) {
> context().execute(this::push, this::onError);
> }
> }
> {code}
> in case of the very first request we will spawn a new task, but it would be
> better to drain the first batch of rows as well.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
