Wouldn't it be possible to take the recordreader class as an config variable and then have a concrete implementation that instantiates the configured record reader? (like streaminputformat)
What I meant about about the splits wasn't so much about the number of maps - but how the allocation of files to each map job is done. Currently the logic (in MultiFileInputFormat) doesn't take the location into account. All we need to do is sort the files by location in the getSplits() method and then do the binning. That way - files in the same split will be co-located. (ok - there are multiple locations for each file - but I would think choosing _a_ location and binning based on that would be better then doing so randomly). -----Original Message----- From: Enis Soztutar [mailto:[EMAIL PROTECTED] Sent: Thursday, April 24, 2008 12:27 AM To: core-user@hadoop.apache.org Subject: Re: Best practices for handling many small files A shameless attempt to defend MultiFileInputFormat : A concrete implementation of MultiFileInputFormat is not needed, since every InputFormat relying on MultiFileInputFormat is expected to have its custom RecordReader implementation, thus they need to override getRecordReader(). An implementation which returns (sort of) LineRecordReader is under src/examples/.../MultiFileWordCount. However we may include it if any generic (for example returning SequenceFileRecordReader) implementation pops up. An InputFormat returns <numSplits> many Splits from getSplits(JobConf job, int numSplits), which is the number of maps, not the number of machines in the cluster. Last of all, MultiFileSplit class implements getLocations() method, which returns the files' locations. Thus it's the JT's job to assign tasks to leverage local processing. Coming to the original question, I think #2 is better, if the construction of the sequence file is not a bottleneck. You may, for example, create several sequence files in parallel and use all of them as input w/o merging. Joydeep Sen Sarma wrote: > million map processes are horrible. aside from overhead - don't do it if u share the cluster with other jobs (all other jobs will get killed whenever the million map job is finished - see https://issues.apache.org/jira/browse/HADOOP-2393) > > well - even for #2 - it begs the question of how the packing itself will be parallelized .. > > There's a MultiFileInputFormat that can be extended - that allows processing of multiple files in a single map job. it needs improvement. For one - it's an abstract class - and a concrete implementation for (at least) text files would help. also - the splitting logic is not very smart (from what i last saw). ideally - it should take the million files and form it into N groups (say N is size of your cluster) where each group has files local to the Nth machine and then process them on that machine. currently it doesn't do this (the groups are arbitrary). But it's still the way to go .. > > > -----Original Message----- > From: [EMAIL PROTECTED] on behalf of Stuart Sierra > Sent: Wed 4/23/2008 8:55 AM > To: core-user@hadoop.apache.org > Subject: Best practices for handling many small files > > Hello all, Hadoop newbie here, asking: what's the preferred way to > handle large (~1 million) collections of small files (10 to 100KB) in > which each file is a single "record"? > > 1. Ignore it, let Hadoop create a million Map processes; > 2. Pack all the files into a single SequenceFile; or > 3. Something else? > > I started writing code to do #2, transforming a big tar.bz2 into a > BLOCK-compressed SequenceFile, with the file names as keys. Will that > work? > > Thanks, > -Stuart, altlaw.org > > >