Andrew Purtell wrote:
Should we up the default lease period from 60 to 90 or 120
seconds?
Maybe. Increasing the lease period might just move the problem around. On the other hand this is the prescription already for a heavily loaded cluster.
Unless any objection, I'll change the default to 120 for RC2. In
hadoop, the roughly equivalent interval,
mapred.tasktracker.expiry.interval, is ten minutes. Will make hbase
more sluggish recognizing downed regionservers but more robust against
the occasional hogging process running on same host.
We can revisit after we integrate zookeeper (hbase-546).
St.Ack
St.Ack
1) In master log, lease timeout notices.
2) In regionserver logs, quiesce/restart.
3) In master log, errors related to META going away.
4) In master log, reassignment of META to a regionserver
still up.
5) In master log, new start messages from the regionservers
that have finished restarting.
6) Import continues (eventually).
Before #1 there are no errors or anything out of the
ordinary in either the master or regionserver logs.
I upped the regionserver lease period from 60 to 120
seconds, reinitialized, and ran the import again. No
problems.
- Andy
