hbase-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Bluemetrix Development <bmdevelopm...@gmail.com>
Subject Re: OOME Java heap space
Date Wed, 24 Feb 2010 17:29:33 GMT

So, 3 nodes (2 CPU 8GB RAM) - one shared master/RS and two other RS.
Two tables: table1 about 8M rows and table2 about 10M rows.
A few days ago, I went to insert another 8M or so into table1 using MR
and HBase crashed
fairly early on in the process. Haven't been able to bring it back up
since then due to all 3 RS reporting OOME.

I'm going to go ahead and try to re-insert the data into table1 and
see if I can get the full 16M in
without issue. Or try to investigate the crash again. If there is
anything you can think of that I should
watch out for, please let me know.

I'll report back with some results.

On Wed, Feb 24, 2010 at 11:38 AM, Stack <stack@duboce.net> wrote:
> Then, the content of the file is not the issue.  Something else is going on.
> The location of an OOME is not necessarily a pointer at the culprit.
> It can be, but usually the culprit has done its nasty work and some
> innocent trying to get a bit of memory ends up throwing the OOME.
> So, lets start over.  How many regionservers.  How many regions.  You
> are just trying to start up hbase?
> St.Ack
> On Wed, Feb 24, 2010 at 7:28 AM, Bluemetrix Development
> <bmdevelopment@gmail.com> wrote:
>> Hi,
>> Scanning the KV pairs using HFile as you suggested, the biggest value
>> I came across was about 2400 characters long string, with that
>> particular row having 25000 cells from what I can tell. Is this big
>> enough to cause the problem?
>> There were a dozen or so values over 1000 chars long, but mainly small
>> values under 100 chars as I mentioned earlier.
>> If I have to set a hard limit on the length of cell values, that is
>> not a problem for the moment - I can chop these strings down.
>> Thanks
>> On Tue, Feb 23, 2010 at 3:40 PM, Bluemetrix Development
>> <bmdevelopment@gmail.com> wrote:
>>> Well, the cells themselves should not be too big. Just a few Strings
>>> (url length) or ints at the most per cell.
>>> Its just that there could be 10M (or maybe even 100M) cells per row.
>>> I'm on the latest 0.20.3.
>>> I'll try to find the big record as you suggested earlier and see what
>>> it looks like.
>>> Thanks
>>> On Tue, Feb 23, 2010 at 3:18 PM, Stack <stack@duboce.net> wrote:
>>>> On Tue, Feb 23, 2010 at 10:40 AM, Bluemetrix Development
>>>> <bmdevelopment@gmail.com> wrote:
>>>>> If this is the case tho, how big is too big?
>>>> Each cell and its coordinates is read into memory.  If not enough
>>>> memory, then OOME.
>>>> Or does it depend on my
>>>>> disk/memory resources?
>>>>> I'm currently using dynamic column qualifiers, so I could have been reaching
>>>>> rows with 10s of millions of unique column qualifiers each.
>>>> This should be fine as long as you are on a recent hbase.
>>>> I'd say it a big cell or many big cells concurrently that caused the OOME.
>>>>> Or, with other tables using timestamps as another dimension to the
>>>>> data, and therefore
>>>>> reaching 10s of millions of versions.
>>>>> (I was trying to get HBase back up so I could count these numbers.)
>>>>> What limits should I use for the time being for number of qualifiers
>>>>> and number of timestamps/versions?
>>>> Shouldn't be an issue.
>>>> St.Ack

View raw message