Re: Slow row deletion performance in comparison to insertion

[Jeff Whiting]

0.90.4-cdh3u3 is the version I'm running.

~Jeff

On 6/27/2012 5:50 PM, Ted Yu wrote:

I created HBASE-6287 <https://issues.apache.org/jira/browse/HBASE-6287> for porting HBASE-5941 to trunk.

Jeff:
What version of HBase are you using ?

Since HBASE-5941 is an improvement, a vote may be raised for porting it to 
other branches.

On Wed, Jun 27, 2012 at 4:15 PM, Jeff Whiting <je...@qualtrics.com <mailto:je...@qualtrics.com>> wrote:

    Looking at HBASE-6284 it seems that deletes are not batched at the 
regionserver level so that
    is the reason for the performance degradation.  Additionally HBASE-5941 
with the locks is also
    contributing to the performance degradation.

    So until those changes get into an hbase release I just have to live with 
the slower
    performance.  Is there anything I need to do on my end?

    Just as a sanity check, I tried setting a timestamp in the delete object 
but it made no
    difference.  I'll batch my deletes at end as you suggested (as memory 
allows).

    Thanks,
    ~Jeff

    On 6/27/2012 4:11 PM, Ted Yu wrote:

        Amit:
        Can you point us to the JIRA or changelist in 0.89-fb ?

        Thanks


        On Wed, Jun 27, 2012 at 3:05 PM, Amitanand Aiyer <amitanan...@fb.com
        <mailto:amitanan...@fb.com>> wrote:

            There was some difference in the way locks are taken for batched 
deletes
            and puts.  This was fixed for 89.

            I wonder if the same could be the issue here.

            Sent from my iPhone

            On Jun 27, 2012, at 2:04 PM, "Jeff Whiting" <je...@qualtrics.com
            <mailto:je...@qualtrics.com>> wrote:

                I'm struggling to understand why my deletes are taking longer 
than my

            inserts.  My understanding is that a delete is just an insertion of 
a
            tombstone.  And I'm deleting the entire row.

                I do a simple loop (pseudo code) and insert the 100 byte rows:

                for (int i=0; i < 50000; i++)
                {
                   puts.append(new Put(rowkey[i], oneHundredBytes[i]));

                   if (puts.size() % 1000 == 0)
                   {
                       Benchmark.start();
                       table.batch(puts);
                       Benchmark.stop();
                   }
                }


                The above takes about 8282ms total.

                However the delete takes more than twice as long:

                Iterator it = table.getScannerScan(rowkey[0],

            rowkey[50000-1]).iterator();

                while(it.hasNext())
                {
                   r = it.next();
                   deletes.append(new Delete(r.getRow()));
                   if (deletes.size() % 1000 == 0)
                   {
                       Benchmark.start();
                       table.batch(deletes);
                       Benchmark.stop();
                   }
                }

                The above takes 17369ms total.

                I'm only benchmarking the deletion time and not the scan time.

            Additionally if I batch the deletes into one big one at the end 
(rather
            than while I'm scanning) it takes about the same amount of time. I 
am
            deleting the entire row so I wouldn't think it would be doing a 
read before
            the delete (
            
http://mail-archives.apache.org/mod_mbox/hbase-user/201206.mbox/%3CE83D30E8F408F94A96F992785FC29D82063395D6@s2k3mntaexc1.mentacapital.local%3E
            ).

                Any thoughts on why it is slower and how I can speed it up?

                Thanks,
                ~Jeff

                --
                Jeff Whiting
                Qualtrics Senior Software Engineer
                je...@qualtrics.com <mailto:je...@qualtrics.com>

-- Jeff Whiting

    Qualtrics Senior Software Engineer
    je...@qualtrics.com <mailto:je...@qualtrics.com>

--
Jeff Whiting
Qualtrics Senior Software Engineer
je...@qualtrics.com