ndimiduk commented on pull request #1933:
URL: https://github.com/apache/hbase/pull/1933#issuecomment-649091065
> > I think 5 mins is too short.
>
> I was thinking about it for huge clusters and yes it is quite possible for
all plan completion to take longer time but I don't have any concrete data as
of now. If 5 min is short, should we not update default value of config:
`hbase.normalizer.period` to higher interval?
I've also been thinking that 5 minutes is too short of a period, I believe
originally it was 30 minutes. It's configurable, and an operator can make it as
they like. Anyway, I think any changes to default settings are not
backwards-compatible and need a separate ticket.
> As far as plan computation and submission is concerned, we have
ReentrantLock to prevent concurrent execution but it's about how much extra
time we want to wait before we assume last round of all split/merge tasks
should have been completed (failed/succeeded).
This goes against my previous suggestion of logging the number of
actions/procedures submitted and finishing. To do this, the normalizer would
have to block on completion of all pids in order to record the timestamp, or
else record all pid's submitted in the previous run, and check all their
statuses (do we have a procedure completion time stored anywhere) in order to
make a decision... No, I think it's better to not try to get into that,
especially if we think we'll bump the normalizer run interval to something much
larger, like every 24 hours.
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
