Description
Recently heard about an issue where a large Hadoop installation which had Accumulo running was experiencing long pauses in the Namenode. Inspecting NN audit logs, it was found that the user running Accumulo issues a getContentSummary("/") call just before the NN pauses were experienced.
In DefaultServlet.java, we use this call to compute the total HDFS disk usage and present a ratio of space that Accumulo uses relative to the total available space.
It's still unclear why this was causing issues in this case (as this operation should only be acquiring a read-lock in the namenode), it was recommended to me that Accumulo use the JMX metrics for the NN instead of making this call.
Attachments
Issue Links
- is related to
-
HDFS-8046 Allow better control of getContentSummary
- Closed