Assorted project updates (tests, build, etc)
Hey All, 1. The original test infrastructure hosted by the AMPLab has been fully restored and also expanded with many more executor slots for tests. Thanks to Matt Massie at the Amplab for helping with this. 2. We now have a nightly build matrix across different Hadoop versions. It appears that the Maven build is failing tests with some of the newer Hadoop versions. If people from the community are interested, diagnosing and fixing test issues would be welcome patches (they are all dependency related). https://issues.apache.org/jira/browse/SPARK-2232 3. Prashant Sharma has spent a lot of time to make it possible for our sbt build to read dependencies from Maven. This will save us a huge amount of headache keeping the builds consistent. I just wanted to give a heads up to users about this - we should retain compatibility with features of the sbt build, but if you are e.g. hooking into deep internals of our build it may affect you. I'm hoping this can be updated and merged in the next week: https://github.com/apache/spark/pull/77 4. We've moved most of the documentation over to recommending users build with Maven when creating official packages. This is just to provide a single "reference build" of Spark since it's the one we test and package for releases, we make sure all recursive dependencies are correct, etc. I'd recommend that all downstream packagers use this build. For day-to-day development I imagine sbt will remain more popular (repl, incremental builds, etc). Prashant's work allows us to get the "best of both worlds" which is great. - Patrick
Re: Assorted project updates (tests, build, etc)
Just a couple of FYI notes: With Zinc and the scala-maven-plugin, repl and incremental builds are also available to those doing day-to-day development using Maven. As long as you don't have to delve into the extra boilerplate and verbosity of Maven's POMs relative to an SBT build file, there is little day-to-day functional difference between the two -- if anything, I find that Maven supports faster development cycles. On Sun, Jun 22, 2014 at 12:24 AM, Patrick Wendell wrote: > Hey All, > > 1. The original test infrastructure hosted by the AMPLab has been > fully restored and also expanded with many more executor slots for > tests. Thanks to Matt Massie at the Amplab for helping with this. > > 2. We now have a nightly build matrix across different Hadoop > versions. It appears that the Maven build is failing tests with some > of the newer Hadoop versions. If people from the community are > interested, diagnosing and fixing test issues would be welcome patches > (they are all dependency related). > > https://issues.apache.org/jira/browse/SPARK-2232 > > 3. Prashant Sharma has spent a lot of time to make it possible for our > sbt build to read dependencies from Maven. This will save us a huge > amount of headache keeping the builds consistent. I just wanted to > give a heads up to users about this - we should retain compatibility > with features of the sbt build, but if you are e.g. hooking into deep > internals of our build it may affect you. I'm hoping this can be > updated and merged in the next week: > > https://github.com/apache/spark/pull/77 > > 4. We've moved most of the documentation over to recommending users > build with Maven when creating official packages. This is just to > provide a single "reference build" of Spark since it's the one we test > and package for releases, we make sure all recursive dependencies are > correct, etc. I'd recommend that all downstream packagers use this > build. > > For day-to-day development I imagine sbt will remain more popular > (repl, incremental builds, etc). Prashant's work allows us to get the > "best of both worlds" which is great. > > - Patrick >
GraphX's VertexRDD can not be materialized by calling count()
Hi there, Seems one can not materialize VertexRDD by simply calling count method, which is overridden by VertexRDD. But if you call RDD's count, it could materialize it. Is this a feature that designed to get the count without materialize VertexRDD? If so, do you guys think it is necessary to add a materialize method to VertexRDD? I did that and ready to send a pull request, but I just want to make sure this is reasonable. By the way, does count() is the cheapest way to materialize a RDD? Or it just cost the same resources like other actions? Best, -- View this message in context: http://apache-spark-developers-list.1001551.n3.nabble.com/GraphX-s-VertexRDD-can-not-be-materialized-by-calling-count-tp7065.html Sent from the Apache Spark Developers List mailing list archive at Nabble.com.
Checkpointed RDD still causing StackOverflow
Hi, I'm doing iterative computing now, and due to lineage chain, we need to checkpoint the RDD in order to cut off lineage and prevent StackOverflow error. The following code still having StackOverflowError, I checked `isCheckpointed` and the result is true. Also, I write a function to count the lineage, but the lineage is not big. Any idea about that? Please give me some hit so I can dig into the source code and try to fix it. Best, -- View this message in context: http://apache-spark-developers-list.1001551.n3.nabble.com/Checkpointed-RDD-still-causing-StackOverflow-tp7066.html Sent from the Apache Spark Developers List mailing list archive at Nabble.com.
Re: Checkpointed RDD still causing StackOverflow
After checkpoint(), please call count(). This is similar to cache(), the RDD is only marked as to be checked with checkpoint(). -Xiangrui On Sun, Jun 22, 2014 at 3:14 PM, dash wrote: > Hi, > > I'm doing iterative computing now, and due to lineage chain, we need to > checkpoint the RDD in order to cut off lineage and prevent StackOverflow > error. > > The following code still having StackOverflowError, I checked > `isCheckpointed` and the result is true. Also, I write a function to count > the lineage, but the lineage is not big. Any idea about that? Please give me > some hit so I can dig into the source code and try to fix it. > > > > > Best, > > > > -- > View this message in context: > http://apache-spark-developers-list.1001551.n3.nabble.com/Checkpointed-RDD-still-causing-StackOverflow-tp7066.html > Sent from the Apache Spark Developers List mailing list archive at Nabble.com.
Re: Checkpointed RDD still causing StackOverflow
Hi Xiangrui, According to my knowledge, calling count is for materialize the RDD, does collect do the same thing since it also an action? I can not call count because for a Graph object, count does not materialize the RDD. I already send an issue on that. My question is, why there still have stack overflow even if `isCheckpointed` is true? -- View this message in context: http://apache-spark-developers-list.1001551.n3.nabble.com/Checkpointed-RDD-still-causing-StackOverflow-tp7066p7068.html Sent from the Apache Spark Developers List mailing list archive at Nabble.com.