hbase-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Geoff Hendrey" <ghend...@decarta.com>
Subject RE: ScannerTimeoutException
Date Thu, 08 Sep 2011 18:56:40 GMT
hfile.block.cache.size is default (.2). Regionservers have 14GB heap.

I did not disable major compactions.


-----Original Message-----
From: Vidhyashankar Venkataraman [mailto:vidhyash@yahoo-inc.com] 
Sent: Thursday, September 08, 2011 11:08 AM
To: user@hbase.apache.org
Cc: Rohit Nigam; Tony Wang; Parmod Mehta
Subject: Re: ScannerTimeoutException

Jira for 3: HBASE-2077

Dince you are adding a column as opposed to a family, you are doing a sparse scan so it is
likely that issue (3) is happening.

* Have you disabled major compactions as well?
 * How big is your block cache?


On 9/8/11 10:57 AM, "Geoff Hendrey" <ghendrey@decarta.com> wrote:


1)   I have splitting disabled

2)   possible, but each reduce task uses its own JVM. Therefore, GC could explain an individual
failure, but not the exponentially increasing failure rate. Further, I can concurrently start
a second M/R job doing exactly the same thing, and it follows the same pattern: No failures
at first…progressing to virtually all scanners timing out. Both jobs are scanning the same

3)   we are using 90.1…does the bug effect us? Jira issue number?



From: Vidhyashankar Venkataraman [mailto:vidhyash@yahoo-inc.com]
Sent: Thursday, September 08, 2011 10:06 AM
To: user@hbase.apache.org; Geoff Hendrey
Subject: Re: ScannerTimeoutException

There are a few reasons why this occurs

1.      Regionserver is very busy. Splits and compactions together suck precious CPU/IO cycles.
2.      GC pauses. See if you can see any swaps (through iostat/vmstat) or any unusual pauses
in GC logs.
3.      There is an existing scanner bug (which Hbase version are you using?) wherein the
scanner does not send keepalive messages until it sees a row that matches your scan. So the
scanner is busy in the regionserver trying to find a row that matches the scan filter while
the client waits for a response.

There could be more reasons. This is what I could come up with off the cuff.

On 9/8/11 8:34 AM, "Jan Lukavsk‡" <jan.lukavsky@firma.seznam.cz> wrote:

Hi Geoff,

I believe you don't need the scan.addColumn() when you add the whole
family. Although this should not affect the timeouts. If the timeouts
are getting more frequent, do you see compations in you RegionServer's
log? Do the timeouts occur while scanning for the same row(s)?


On 8.9.2011 17:23, Geoff Hendrey wrote:
> Hi Jan -
> The relevant code looks like this. I added addFamily. Still getting the ScannerTimeoutException.
Arghhh. This is really becoming a mission critical problem for my team...Furthermore, the
problem becomes more and more frequent, to the point that the job almost never finishes. At
the beginning of the job, the timout never happens. By the end, it's happening 9 out of 10
>                  Scan scan = new Scan(Bytes.toBytes(key));
>                  scan.setCaching(1);
>                  scan.setMaxVersions(1);
>                  scan.addFamily(Bytes.toBytes("V1"));
>                  scan.addColumn(Bytes.toBytes("V1"), Bytes.toBytes("cluster_map"));
>                  scan.addColumn(Bytes.toBytes("V1"), Bytes.toBytes("version_control_number"));
> -geoff
> -----Original Message-----
> From: Geoff Hendrey
> Sent: Tuesday, September 06, 2011 9:51 AM
> To: Jan Lukavský; user@hbase.apache.org
> Cc: hbase-user@hadoop.apache.org
> Subject: RE: ScannerTimeoutException
> I'll try your suggestions!
> -----Original Message-----
> From: Jan Lukavský [mailto:jan.lukavsky@firma.seznam.cz]
> Sent: Tuesday, September 06, 2011 9:48 AM
> To: user@hbase.apache.org
> Cc: Geoff Hendrey; hbase-user@hadoop.apache.org
> Subject: Re: ScannerTimeoutException
> Hi Geoff,
> we are having these issues when the scanner uses scan.addColumn() and
> the column is sparse, in the sense there are many rows with some other
> column in the same column family. I suppose your problems will vanish if
> you use scan.addFamily() call instead. The same behavior may appear if
> you are reading from region after massive delete (then the timeouts
> settle down after major compation), or when using server side Filters.
> Changing scan.addColumn() to scan.addFamily() brings some overhead,
> which I think could be removed by RegionServer renewing the lease of
> scanner while reading data, not only after first entry to
> HRegionServer.next().
> Would this be worth opening a JIRA?
> Jan
> On 6.9.2011 04:11, Geoff Hendrey wrote:
>> Hi -
>> I found some odd behavior with ResultScanner.next(). Usually the times
>> for next() are couple hundred ms. But occasionally the call to next
>> spikes VERY long. In fact, I have the timeout set to 60 seconds (60000),
>> but once in a while the call to next() itself is interrupted by the
>> ScannerTimeoutException after more than 60 seconds. It seems odd that
>> the call to next itself can be interrupted because "61107ms passed since
>> the last invocation, timeout is currently set to 60000"
>> The only thing I can think of is that a GC kicks in after the call to
>> next begins, but before the call returns, and the server is still
>> ticking the timeout. But this seems to happen periodically with
>> regularity. The odds of the GC kicking in at that exact instant, so
>> often, seem kind of low.
>> -geoff

View raw message