[
https://issues.apache.org/jira/browse/HBASE-22301?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16825637#comment-16825637
]
Andrew Purtell commented on HBASE-22301:
----------------------------------------
Thanks [~busbey]
[~dmanning] convinced me a count based threshold might be better. Let me attach
a new version of this patch. Please let me know what you think. We could commit
either option.
The drawback to rolling on detection of a single latency outlier is that the
threshold might have to be set really high to exclude "slow syncs" which are
really GC activity. We went back and forth on what threshold might be
appropriate to avoid false positives yet still be effective. I still want to do
something simple, so instead of creating a new latency threshold we reuse the
existing slow sync warn threshold, and then count how many times we exceed the
slow sync warn threshold within a longer and configurable interval. If we
counted enough warnings over that latter interval, then we ask for a roll. We
can only ask once per interval so retain a limit on pacing to prevent runaway
roll requests during an outage or incident. This can be a better strategy
because while it is possible a single outlier is GC activity, the probability
of many data points being a false positive is less the more there are. Defaults
in new patch is an interval of one minute; and a count based threshold of ten
slow sync warnings within that interval.
> Consider rolling the WAL if the HDFS write pipeline is slow
> -----------------------------------------------------------
>
> Key: HBASE-22301
> URL: https://issues.apache.org/jira/browse/HBASE-22301
> Project: HBase
> Issue Type: Improvement
> Components: wal
> Reporter: Andrew Purtell
> Assignee: Andrew Purtell
> Priority: Minor
> Fix For: 3.0.0, 1.5.0, 2.3.0
>
> Attachments: HBASE-22301-branch-1.patch, HBASE-22301-branch-1.patch,
> HBASE-22301-branch-1.patch
>
>
> Consider the case when a subset of the HDFS fleet is unhealthy but suffering
> a gray failure not an outright outage. HDFS operations, notably syncs, are
> abnormally slow on pipelines which include this subset of hosts. If the
> regionserver's WAL is backed by an impacted pipeline, all WAL handlers can be
> consumed waiting for acks from the datanodes in the pipeline (recall that
> some of them are sick). Imagine a write heavy application distributing load
> uniformly over the cluster at a fairly high rate. With the WAL subsystem
> slowed by HDFS level issues, all handlers can be blocked waiting to append to
> the WAL. Once all handlers are blocked, the application will experience
> backpressure. All (HBase) clients eventually have too many outstanding writes
> and block.
> Because the application is distributing writes near uniformly in the
> keyspace, the probability any given service endpoint will dispatch a request
> to an impacted regionserver, even a single regionserver, approaches 1.0. So
> the probability that all service endpoints will be affected approaches 1.0.
> In order to break the logjam, we need to remove the slow datanodes. Although
> there is HDFS level monitoring, mechanisms, and procedures for this, we
> should also attempt to take mitigating action at the HBase layer as soon as
> we find ourselves in trouble. It would be enough to remove the affected
> datanodes from the writer pipelines. A super simple strategy that can be
> effective is described below:
> This is with branch-1 code. I think branch-2's async WAL can mitigate but
> still can be susceptible. branch-2 sync WAL is susceptible.
> We already roll the WAL writer if the pipeline suffers the failure of a
> datanode and the replication factor on the pipeline is too low. We should
> also consider how much time it took for the write pipeline to complete a sync
> the last time we measured it, or the max over the interval from now to the
> last time we checked. If the sync time exceeds a configured threshold, roll
> the log writer then too. Fortunately we don't need to know which datanode is
> making the WAL write pipeline slow, only that syncs on the pipeline are too
> slow and exceeding a threshold. This is enough information to know when to
> roll it. Once we roll it, we will get three new randomly selected datanodes.
> On most clusters the probability the new pipeline includes the slow datanode
> will be low. (And if for some reason it does end up with a problematic
> datanode again, we roll again.)
> This is not a silver bullet but this can be a reasonably effective mitigation.
> Provide a metric for tracking when log roll is requested (and for what
> reason).
> Emit a log line at log roll time that includes datanode pipeline details for
> further debugging and analysis, similar to the existing slow FSHLog sync log
> line.
> If we roll too many times within a short interval of time this probably means
> there is a widespread problem with the fleet and so our mitigation is not
> helping and may be exacerbating those problems or operator difficulties.
> Ensure log roll requests triggered by this new feature happen infrequently
> enough to not cause difficulties under either normal or abnormal conditions.
> A very simple strategy that could work well under both normal and abnormal
> conditions is to define a fairly lengthy interval, default 5 minutes, and
> then insure we do not roll more than once during this interval for this
> reason.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
