Hi Ted,
hmm, the stripe compaction seems to solve the problem, provided there
will be some statistics available (ideally over RPC on HMaster) for each
stripe (subregion range). In my original idea, the ranges would have to
be determined based on some heuristic (probably a guess could be made
based on file size of all HFiles in the region directory on HDFS).
Jan
On 04/30/2014 12:26 PM, Ted Yu wrote:
Interesting idea.
How would the ranges of rows be determined per region ?
This reminds me of stripe compaction feature which is in 0.98
See HBASE-7667
Cheers
On Apr 30, 2014, at 3:07 AM, Jan Lukavský <[email protected]> wrote:
Hi all,
I have a general idea I'd like to consult. A short description of a problem we
are facing: during mapreduce jobs run over HBase cluster, we very often see
great disproportions in run time of different map tasks (some tasks tend to
finish in minutes or even seconds, while others might take even hours). This
causes the job to run inefficiently and the whole cluster to be underutilized -
reducers have to wait until all the map tasks finish - at least before starting
the sort phase. The number of long running map tasks is usually low, so the
whole cluster basically waits until several machines finish their work. We
tried to get over this by sampling the regions and creating some statistics
(one statistic per mapreduce job), which we then used to tune the input format
splits to make the distribution of running time more even. This seems to work
(although at the time being might cause some issues with data locality, which
we think we can solve).
Now, the questions is, would it be possible to calculate some statistics during
major compactions and store them in the region directory on HDFS? What I mean
by these statistics, I think it could be possible to store for some reasonable
ranges of rows (so that for each region there would be like hundreds of these
ranges):
* total number of rows between specified rows
* total number of KeyValues
* amount of data stored on disk
These statistics could be calculated per column family and subsequently used in
InputFormat to tune the splits to match even distribution as close as possible.
Is anyone else interested in this? Does anyone have any other solution to the
problem I have described? I know we could say manually split regions that take
long time to process, but first, these regions are job-specific (so different
jobs have different regions that take long time to process), and second,
ideally I'm looking for an automated solution.
Thanks for reply,
Jan
--
Jan Lukavský
Vedoucí týmu vývoje
Seznam.cz, a.s.
Radlická 3494/10
15000, Praha 5
[email protected]
http://www.seznam.cz
