[
https://issues.apache.org/jira/browse/TRAFODION-1996?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15322455#comment-15322455
]
Sean Broeder commented on TRAFODION-1996:
-----------------------------------------
My suggestion to fix this is to process each region sequentially rather than
inundate the region server during each cleanup.
In addition I will be submitting a change to skip the WAL for the deletes as a
performance improvement. Normally skipping the WAL will have no effect on the
contests, but in the rare case where a delete is lost there in no problem
because the delete will be repeated in a subsequent scan during the next
cleanup. Data will still be consistent.
> TLOG uses excessive memory when aging old entries
> -------------------------------------------------
>
> Key: TRAFODION-1996
> URL: https://issues.apache.org/jira/browse/TRAFODION-1996
> Project: Apache Trafodion
> Issue Type: Bug
> Components: dtm
> Affects Versions: 2.0-incubating
> Reporter: Sean Broeder
> Assignee: Sean Broeder
> Fix For: 2.1-incubating
>
> Original Estimate: 168h
> Remaining Estimate: 168h
>
> When deleting old entries in the TLOG tables a complete table scan is
> performed and all regions for all TLOG tables are scanned in parallel. This
> was done to try to improve performance by attempting to delete all old
> entries as quickly as possible. However, with extremely large scans on busy
> systems the scans can become quite large and in a worst case scenario can
> starve out other requests in a regions server by dominating the handlers and
> potentially causing a RS failure due to a long GC pause.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
