[ https://issues.apache.org/jira/browse/YARN-1717?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13935568#comment-13935568 ]
Zhijie Shen commented on YARN-1717: ----------------------------------- Billie, thanks for your metrics. I've don some simple calculation myself. In long term, if a cluster has x entities written per second, no matter how long the ttl is, the number of entities to delete per second should be x on average. Therefore, let's say throughput of put requests is 100 entities/sec, the number of entities to delete per second will be 100 as well. Given we do the deletion every 5 minutes, we have 30,000 entities to delete per round. According to your measurement, it will take less than 8 sec to complete the deletion. The deletion will delay put request, but every 5 mins, it just happens for 8 secs, i.e., 2.67%. It sounds good to me. +1 for the patch. Will commit it. > Enable offline deletion of entries in leveldb timeline store > ------------------------------------------------------------ > > Key: YARN-1717 > URL: https://issues.apache.org/jira/browse/YARN-1717 > Project: Hadoop YARN > Issue Type: Sub-task > Reporter: Billie Rinaldi > Assignee: Billie Rinaldi > Attachments: YARN-1717.1.patch, YARN-1717.10.patch, > YARN-1717.11.patch, YARN-1717.2.patch, YARN-1717.3.patch, YARN-1717.4.patch, > YARN-1717.5.patch, YARN-1717.6-extra.patch, YARN-1717.6.patch, > YARN-1717.7.patch, YARN-1717.8.patch, YARN-1717.9.patch > > > The leveldb timeline store implementation needs the following: > * better documentation of its internal structures > * internal changes to enable deleting entities > ** never overwrite existing primary filter entries > ** add hidden reverse pointers to related entities -- This message was sent by Atlassian JIRA (v6.2#6252)