nkeywal created HBASE-7246:
------------------------------
Summary: Assignment#nodeChildrenChanged calls
listChildrenAndWatchForNewChildren, overloading master & zookeper needlessly
Key: HBASE-7246
URL: https://issues.apache.org/jira/browse/HBASE-7246
Project: HBase
Issue Type: Improvement
Components: master, Region Assignment
Affects Versions: 0.96.0
Reporter: nkeywal
Assignment#nodeChildrenChanged listens to change event on hbase/unassigned,
then calls listChildrenAndWatchForNewChildren. As its name says, this function
gets the whole list of unnasigned regions from ZK.
The internal comments says it's only for split creation. I tried to remove it,
it seems to work.
The point is:
- ZK does not allow us to know upfront if it's a split or not
- 50% of the events will be deletions, we don't care about them but we can't
know in advance
- for large assignment, we will have a lot of children (hence events). Common
practise in ZK is to limit the number of subnodes (see the threads
http://www.mail-archive.com/[email protected]/msg01954.html)
Monitoring the master, it was spending 10% of it's time on this during
assignment. It seems it was not on the critical path (yet...), so removing it
did not change the overall performances on a test scenario. In real life
however, this typically adds workload to ZK, and ZK workload is precious on
large failures... Same as master actually.
If it's used only for split, we should have a znode specific for split and not
a global znode.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
