Yes I agree with you.
Individually each HBase daemon can block clients and/or reject writes and/or
defer compactions as write latency increases. This is a second order indication
of cluster stress. By the time write latency falls outside bounds, it may be
too late to shut down gracefully. But anyway, we should do these things.
Metrics can be used for HBase and applications/writers and cluster admins to
globally know directly that the cluster is under stress in order to coordinate
a response. Such a response could happen concurrently on multiple levels.
But getting back to your point, there are certainly opportunities for HBase to
take self preserving actions where currently it does not.
- Andy
________________________________
From: elsif <[email protected]>
To: [email protected]
Sent: Tue, November 10, 2009 1:25:20 AM
Subject: Re: HBase Exceptions on version 0.20.1
The larger issue here is that any hbase cluster will reach this tipping
point at some point in its lifetime as more and more data is added. We
need to have a graceful method to put the cluster into safe mode until
more resources can be added or the load on the cluster has been
reduced. We cannot allow hbase to run itself into the ground causing
data loss or corruption under any circumstances.
*
*
Andrew Purtell wrote:
> You should consider provisioning more nodes to get beyond this ceiling you
> encountered.
>
> DFS write latency spikes from 3 seconds to 6 seconds, to 15! Flushing cannot
> happen fast enough to avoid an OOME. Possibly there was even insufficient CPU
> to GC. The log entries you highlighted indicate the load you are exerting on
> your current cluster needs to be spread out over more resources than
> currently allocated.
>
> This:
>
>> 2009-11-06 09:15:37,144 WARN org.apache.hadoop.hbase.util.Sleeper: We slept
>> 286007ms, ten times longer than scheduled: 10000
>>
>
> indicates a thread that wanted to sleep for 10 seconds was starved for CPU
> for 286 seconds. Obviously Zookeeper timeouts and resulting HBase process
> shutdowns, missed DFS heartbeats possibly resulting in spurious declaration
> of dead datanodes, and other serious problems will result from this.
>
> Did your systems start to swap?
>
> When region servers shut down, the master notices this and splits their HLogs
> into per region reconstruction logs. These are the "oldlogfile.log" files.
> The master log will shed light on why this particular reconstruction log was
> botched. Would have happened at the master. The region server probably did do
> a clean shutdown. I suspect DFS was in extremis due to overloading so the
> split failed. The checksum error indicates incomplete write at the OS level.
> Did a datanode crash?
>
> HBASE-1956 is about making the DFS latency metric exportable via the
> Hadoop metrics layer, perhaps via Ganglia. Write latency above 1 or 2
> seconds is a warning. Anything above 5 seconds is an alarm. It's a
> good indication that an overloading condition is in progress.
>
> The Hadoop stack, being pre 1.0, has some rough edges. Response to
> overloading is one of them. For one thing, HBase could be better about
> applying backpressure to writing clients when the system is under stress. We
> will get there. HBASE-1956 is a start.
>
> - Andy
>
>
