[
https://issues.apache.org/jira/browse/HADOOP-2621?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12559530#action_12559530
]
stack commented on HADOOP-2621:
-------------------------------
Looks like hbase.hregion.memcache.flush.size is no longer being read/used.
This probably means that the blocking mechanism -- when memcache > twice the
limit, we stop taking on updates -- no longer works. Was useful for when
regionservers were on occasion overwhelmed; it gave them a chance to catch
their breath.
>From Billy's description up on IRC, compaction was being overrun by the number
>of files produced flushing when cluster was under load. In the past there was
>an attempt at striking an equilibrium between flush and compaction rates/sizes.
I'll take a look at this.
> Memcache flush flushing every 60 secs with out considering the max memcache
> size
> --------------------------------------------------------------------------------
>
> Key: HADOOP-2621
> URL: https://issues.apache.org/jira/browse/HADOOP-2621
> Project: Hadoop
> Issue Type: Bug
> Components: contrib/hbase
> Reporter: Billy Pearson
> Fix For: 0.16.0
>
>
> looks like hbase is flushing all memcache to disk every 60 secs causing a lot
> of work for the compactor to keep up because column gets its own mapfile and
> every region is flushed at one time. This could be a vary large number of
> mapfiles to write if a region server is hosting 100 regions all with milti
> columns.
> Idea memcache flush
> keep all data in memory until memcache get larger then the conf size with
> hbase.hregion.memcache.flush.size.
> When we reach this size we should flush the regions that are the largest
> first stopping once we drop back below the memcache max size maybe 20% below
> the max. This will to flush only as needed as each flush takes time to
> compact when compaction runs on a region. while we are flushing a region we
> should also be blocking new updates from happening on that region so the
> region server does not get over ran when a high update load hits a region
> server. By only blocking on the region we are flushing at that time other
> regions will still be able to do updates this.
> We we still want to use the hbase.regionserver.optionalcacheflushinterval we
> should set to to run once an hour so something like that so we can recover
> memory from the memcache on region that do not have a lot updates in memory.
> But running at the default set now of 60 secs is not so good for the
> compactor if it has many regions to handle also not good for a scanner to
> have to scan many small files vs a few larger ones
> Example a compactor may take 15 mins to compact a region in that time we will
> flush 15 times causeing all other regions to get a new mapfile to compact
> when it becomes it turn to get compacted if you had many regions getting
> compacted the last one on the list of say 10 regions would have 10 regions *
> 15 mins each = 150 mapfiles for each column in the last region written before
> the compactor can get to it.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
