Home | About | Sematext search-lucene.com search-hadoop.com
 Search Hadoop and all its subprojects:

Switch to Threaded View
HBase, mail # dev - HLogSplit error with hadoop-2.0.3-alpha and hbase trunk


Copy link to this message
-
Re: HLogSplit error with hadoop-2.0.3-alpha and hbase trunk
Jimmy Xiang 2013-05-08, 19:46
I got the same error today and filed HBASE-8511.  The log splitting is
stuck there for a while no matter how many retries.

Thanks,
Jimmy
On Wed, May 8, 2013 at 12:17 PM, Anoop John <[EMAIL PROTECTED]> wrote:

> I see we have FSDataInputStream in ProtobufLogReader.. We will test with
> the patch tomorow and report back.
>
> -Anoop-
>
> On Thu, May 9, 2013 at 12:30 AM, Anoop John <[EMAIL PROTECTED]> wrote:
>
> >  >However this doesn't explain why sequencefile has no such problems
> >
> > What I can see is we use FSDataInputStream in case of
> > SequenecFileLogReader (Not FSInputStream) and we use FSDataInputStream
> > #readInt() (WALEdit#read...).  This inturn reads the int from stream by
> > calling in.read() 4 times making sure we read full int (If it is not EOF
> > issue)
> >
> > May be we can try out this way and test tomorow Ram.
> >
> > BTW IOUtils.readFully() is not returning the actual # bytes read it
> looks.
> >
> > -Anoop-
> >
> >  On Thu, May 9, 2013 at 12:07 AM, Sergey Shelukhin <
> [EMAIL PROTECTED]
> > > wrote:
> >
> >> 589824 is divisible by 4096, default buffer size. Que groundless
> >> suspicions. However this doesn't explain why sequencefile has no such
> >> problems. Do you have any more numbers with this?
> >> I am in meetings now, let me try readFully after I come back from lunch
> >>
> >> On Wed, May 8, 2013 at 11:23 AM, ramkrishna vasudevan <
> >>  [EMAIL PROTECTED]> wrote:
> >>
> >> > >The length is from read call. This looks pretty suspicious, if the
> >> stream
> >> > is not EOF why would it return less bytes?
> >> >
> >> > Infact this happened for an entry in the middle of the HLog.  Write
> some
> >> > data using LoadTestTool.  Try to use the HLog tool to read the HLog
> file
> >> > you will end up in this issue. (most of the times).
> >> >
> >> > Regards
> >> > Ram
> >> >
> >> >
> >> >
> >> > On Wed, May 8, 2013 at 11:48 PM, Sergey Shelukhin <
> >> [EMAIL PROTECTED]
> >> > >wrote:
> >> >
> >> > > if (length != intBytes.length) throw new IOException("Failed read of
> >> int
> >> > > length " + length);
> >> > > The length is from read call. This looks pretty suspicious, if the
> >> stream
> >> > > is not EOF why would it return less bytes? I will try to repro
> today.
> >> > >
> >> > > On Wed, May 8, 2013 at 5:46 AM, ramkrishna vasudevan <
> >> > > [EMAIL PROTECTED]> wrote:
> >> > >
> >> > > > On further debugging found that this issue happens with
> >> ProtoBufWriter
> >> > > and
> >> > > > not with sequenceFileWriter.(atleast we could not reproduce it
> with
> >> > > > different runs)
> >> > > >
> >> > > > We can see that the HLog has more data in it but while reading one
> >> of
> >> > the
> >> > > > lines in the HLog this error happens.  So pretty much sure that it
> >> is
> >> > not
> >> > > > EOF.
> >> > > > Verified DFS logs but could not find any exceptions out there too.
> >> > > >
> >> > > > We will try to figure out more on this tomorrow.
> >> > > >
> >> > > > Regards
> >> > > > Ram
> >> > > >
> >> > > >
> >> > > > On Wed, May 8, 2013 at 11:34 AM, ramkrishna vasudevan <
> >> > > > [EMAIL PROTECTED]> wrote:
> >> > > >
> >> > > > > Ok so i tried this out with hadoop 2.0.4 and also with Sergey's
> >> > patch.
> >> > > > >  The issue is reproducible in all version of hadoop but not
> >> always.
> >> > > > > I am able to get the errors like this
> >> > > > >
> >> > > > > 2013-05-07 17:11:08,827 WARN  [SplitLogWorker-ram.sh.intel.com<
> http://splitlogworker-ram.sh.intel.com/>
> >> > > > ,60020,1367961009182]
> >> > > > > org.apache.hadoop.hbase.regionserver.SplitLogWorker: log
> >> splitting of
> >> > > > .logs/
> >> > > > > ram.sh.intel.com,60020,1367960957620-splitting/ram.sh.intel.com
> >> > > > %2C60020%2C1367960957620.1367960993389
> >> > > > > failed, returning error
> >> > > > > java.io.IOException: Error  while reading 1 WAL KVs; started
> >> reading
> >> > at
> >> > > > > 589822 and read up to 589824
> >> > > > > at
> >> > > > >