It's weird, I thought we already did something like that and it seems that the old TableInputFormatBase does it but not the new one. From it's javadoc:
* Splits are created in number equal to the smallest between numSplits and * the number of {@link HRegion}s in the table. If the number of splits is * smaller than the number of {@link HRegion}s then splits are spanned across * multiple {@link HRegion}s and are grouped the most evenly possible. In the * case splits are uneven the bigger splits are placed first in the * {@link InputSplit} array. J-D On Sat, Apr 9, 2011 at 9:48 AM, Stack <st...@duboce.net> wrote: > Yes, you could make a different Splitter. Would be nice in the > splitter if you could keep the locality where we have the Map task > running on the TaskTracker that is adjacent to the hosting > RegionServer. That shouldn't be hard. Study the current splitter and > see how it juggles locations. > > Can you put us in contact w/ the person running the cluster (offline > if you prefer)? 150k sounds like regions need to be bigger. > > Thanks, > St.Ack > > On Sat, Apr 9, 2011 at 9:33 AM, Avery Ching <ach...@yahoo-inc.com> wrote: >> The number of regions is pretty insane, but not under my control >> unfortunately. The workaround I suggested is to write another InputFormat >> and InputSplit such that each InputSplit is responsible for a configurable >> number of regions. For example, if i have 100k regions and I configure each >> InputSplit to handle 1k regions, then I'll only have 100 map tasks. Just >> was wondering if anyone else faced these issues. >> >> Thanks for your quick response on a Saturday morning =), >> >> Avery >> >> On Apr 9, 2011, at 9:26 AM, Jean-Daniel Cryans wrote: >> >>> You cannot have more mappers than you have regions, but you can have >>> less. Try going that way. >>> >>> Also 149,624 regions is insane, is that really the case? I don't think >>> i've ever seen such a large deploy and it's probably bound to hit some >>> issues... >>> >>> J-D >>> >>> On Sat, Apr 9, 2011 at 9:15 AM, Avery Ching <ach...@yahoo-inc.com> wrote: >>>> Hi, >>>> >>>> First off, I'd like to say thanks to the developers for HBase, it's been >>>> fun to work with. >>>> >>>> I've been using TableInputFormat to run a Map-Reduce job and ran into an >>>> issue. >>>> >>>> Exception in thread "main" org.apache.hadoop.ipc.RemoteException: >>>> java.io.IOException: java.io.IOException: The number of tasks for this job >>>> 149624 exceeds the configured limit 100000 >>>> >>>> The table i'm accessing has 149624 regions, however my Hadoop instance >>>> won't allow me to start a job with that many map tasks. After briefly >>>> looking at the TableInputFormatBase code, it appears that since TableSplit >>>> only knows about a single region, my job will be forced into having >>>> mappers == # of regions. Since the Hadoop instance I'm using is shared, >>>> I'm concerned that even if configured limit was raised, having Jobs with >>>> so many mappers would eventually cause havoc to the job tracker. >>>> >>>> Given that I have no control over the number of regions in the table >>>> (maintained by someone else), is the only solution to implement another >>>> input format (i.e. MultiRegionTableFormat) that allows InputSplits to have >>>> more than one region? I don't mind doing it, but didn't want to write it >>>> if another solution already exists. >>>> >>>> Apologies if this issue has been raised before, but a quick search didn't >>>> turn anything up for me. >>>> >>>> Thanks, >>>> >>>> Avery >>>> >> >> >