[
https://issues.apache.org/jira/browse/PHOENIX-1278?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14144417#comment-14144417
]
Lars Hofhansl commented on PHOENIX-1278:
----------------------------------------
It generally looks like we need some careful tuning as how much work and in
what chunk size we push to the server.
This is a somewhat unexpected (although in hindsight obvious) outcome.
What we want is to keep client and server equally busy. If we have many clients
we want to push more work there. When we only a few client processes we want to
push more work into the server.
Rather that disabling this here, why not take the fact that a table is salted
into account when calculating the desired parallelism, i.e. by combining chunk
into a single task.
This is not necessarily a degradation for salted tables only, right?
> Performance degradation for salted tables with guideposts
> ---------------------------------------------------------
>
> Key: PHOENIX-1278
> URL: https://issues.apache.org/jira/browse/PHOENIX-1278
> Project: Phoenix
> Issue Type: Bug
> Reporter: James Taylor
> Assignee: Anoop Sam John
>
> When a table is salted, we're seeing a degradation in performance using our
> new guidepost-based parallelization. With salted tables, we do a merge sort
> with the results from all the parallel scans. I suspect the cause here is
> that we're doing a merge sort now between more chunks than before (since we
> chunk everything up more now than we used to). We should group the scans
> we're doing for the same bucket together and do a concat with those results
> and then do a merge sort only with the concatenated batches.
> Pls revert PHOENIX-1279 when we implement this.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)