On 24 May 2013 00:52, Stephen Watt <sw...@redhat.com> wrote: > Hi Folks > > Hadoop's pluggable filesystem architecture supports the ability to enable > an alternate filesystem for use with Hadoop by writing a plugin for it. We > now have several alternate filesystems that have Hadoop FileSystem plugins > and because this isn't a very well understood topic, I've been working on a > page on the project wiki to bring this all together - > http://wiki.apache.org/hadoop/HCFS. At the same time, the Ambari project > has been opening up Ambari to support any configured Hadoop FileSystem (as > opposed to just HDFS) over at > https://issues.apache.org/jira/browse/AMBARI-1817 > > My team (over at Red Hat) have been working on writing a Hadoop FileSystem > plugin for the glusterfs filesystem and have been finding that some of the > expected semantics of the operations within the Abstract FileSystem class > are a little ambiguous. With that said, we've joined Steve Loughran in > attempting to clarify these for both the Hadoop 1.0 and the Hadoop 2.0 > FileSystem class over at https://issues.apache.org/jira/browse/HADOOP-9371 > > It seems to me that once we had these semantics defined, it would be good > for consistency of implementation if we could make sure they are well > understood and properly implemented by the community of folks writing > Hadoop FileSystem plugins. To that end, we might work to ensure that those > semantics are tested within an exhaustive test framework that focuses on > the abstract Hadoop FileSystem layer. Each FileSystem provider could run > the tests to ensure their plugin implementation and behavior is consistent > with the expectation. Perhaps a broader extension of > https://issues.apache.org/jira/browse/HADOOP-9258. > > I have a plan for starting those tests, pulling up the Swift ones when they are checked in. Big tests that do scale, and that verify the assumptions that MR, HBase &c are where we are weakest. The defacto definition of FS sematics are the apps, and its them that currently find the problems (e.g MAPREDUCE-5264)
> If folks are interested in these goals, I could host a > workshop/discussion/hackday in Mountain View to get local people together > (perhaps a Google Hangout for the remote folks) to keep the ball rolling on > the semantics discussion and test creation. As a side note, I think this > could also turn out be quite an effective means of introducing FileSystem > vendors to the ASF and getting them contributing to these aspects of the > project. > > Can we start with some G+ hangouts to get to know each other and have some broader participation (myself, the others working on Swift, people who have done S3 (Tom, some of the amazon folk), etc...), Then when a workshop is held, it's got some clearer objectives "how do we test this". I would want the FS semantics to be locked down in some online discussions/JIRA rather than come back after a night's sleep to discover it had be defined with tests. -steve
