I think it's simplest to publish a single Hadoop tarball and users start the services they want. This is the model we have always followed up to now.
Cheers, Tom On Wed, Oct 12, 2011 at 9:07 AM, Alejandro Abdelnur <t...@cloudera.com> wrote: > Currently common, hdfs and mapred create partial tars which are not usable > unless they are stitched together into a single tar. > > With HADOOP-7642 the stitching happens as part of the build. > > The build currently produces the following tars: > > 1* common TAR > 2* hdfs (partial) TAR > 3* mapreduce (partial) TAR > 4* hadoop (full, the stitched one) TAR > > #1 on its own does not run anything, #2 and #3 on their own don't run. #4 > runs hdfs & mapreduce. > > Questions: > > Q1. Does it make sense to publish #1, #2 & #3? Or #4 is sufficient and you > start the services you want (i.e. Hbase would just use HDFS)? > > Q2. And what about a source TAR, does it make sense to have source TAR per > component or a single TAR for the whole? > > > For simplicity (for the build system and for users) I'd prefer a single > binary TAR and a single source TAR. > > Thanks. > > Alejandro >
