[jira] [Commented] (ACCUMULO-3710) Scanning with many singleton ranges crashes tserver

["Josh Elser (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/ACCUMULO-3710?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14544153#comment-14544153 ] 

Josh Elser commented on ACCUMULO-3710:
--------------------------------------

bq. It'd be good to have some mitigation for this, in order to avoid the potentially large number of ranges being issued, like in ACCUMULO-3602.

Agreed. Doing some more chunking in the client to prevent spamming a tserver w/ an exorbitant number of ranges would be good to do. My hunch is that optimizing small, disjoint collections of ranges more efficiently is different work.

> Scanning with many singleton ranges crashes tserver
> ---------------------------------------------------
>
>                 Key: ACCUMULO-3710
>                 URL: https://issues.apache.org/jira/browse/ACCUMULO-3710
>             Project: Accumulo
>          Issue Type: Bug
>          Components: client, tserver
>    Affects Versions: 1.6.1
>            Reporter: Dylan Hutchison
>             Fix For: 1.8.0
>
>
> Setup: single-node standalone 1.6.1 Accumulo instance.
> Use case: scan ~1M individual rows, scattered across a ~15GB table.  
> The following steps crash the TabletServer:
> 1. Gather a List of Range objects, each one a singleton range spanning an entire row.
> 2. Create a BatchScanner with one read thread.
> 3. Set the ranges via BatchScanner.setRanges()
> 4. Start iterating through the scanner.
> One solution is to batch the reads into groups of ~10k ranges idea.  
> Comment from Josh Elser:
> {quote}
> Taking a quick glance at the code, it looks like this would be a good place to do some optimization in the BatchScanner's impl (TabletServerBatchReaderImpl). The BatchScanner will bin the ranges to the tablets and the servers hosting those tablets. Normally, this would be spread out, but, in your single server case, all 1M rows would all go to a single TabletServer in one RPC call.
> I'm guessing a good optimization here would be to check the size of a batch of Ranges for a single tabletserver, and when above a certain threshold, split the batch in half and try to reprocess each half (the recursion would naturally keep splitting until we get down to some high-watermark).
> Point being, if your client VM constructed the Ranges without issue, the BatchScanner impl should be smart enough to not knock over a TabletServer.
> {quote}
> Verified to cause an OOME via  tserver_localhost.out:
> {quote}
> #
> # java.lang.OutOfMemoryError: Java heap space
> # -XX:OnOutOfMemoryError="kill -9 %p"
> #   Executing /bin/sh -c "kill -9 12833"...
> {quote}



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)