[
https://issues.apache.org/jira/browse/HDFS-9802?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15145279#comment-15145279
]
Kihwal Lee commented on HDFS-9802:
----------------------------------
It should be possible to selectively save blocks to trash based on the file's
time stamp. Although {{RollingUpgradeStatus}} does not provide start time, we
can either add a filed or make datanode record the time when it gets a relevant
heartbeat response.
> Selectively save blocks to trash dir during rolling upgrades
> ------------------------------------------------------------
>
> Key: HDFS-9802
> URL: https://issues.apache.org/jira/browse/HDFS-9802
> Project: Hadoop HDFS
> Issue Type: Bug
> Reporter: Kihwal Lee
>
> Currently datanodes save any invalidated blocks to the trash directory during
> a rolling upgrade. Compared to the "previous" directory for full upgrade, the
> trash can grow quickly. This is especially true when blocks are created and
> then quickly deleted.
> Since trash is mainly meant to be defense against faulty new namenode and
> used for rolling back, saving new blocks in trash does not add much value. If
> anything, datanodes run out of space quicker.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
