crunch-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Josh Wills (JIRA)" <>
Subject [jira] [Commented] (CRUNCH-296) Support new distributed execution engines (e.g., Spark)
Date Tue, 10 Dec 2013 17:50:11 GMT


Josh Wills commented on CRUNCH-296:

[~gabriel.reid] that error makes sense to me, actually-- the CombineFn generic definition
has more constraints on it than the DoFn<S, Pair<K, V>> has at that point (i.e.,
there's a constraint that S must be Pair<K, Iterable<V>>). I'm fine with that

Re: caching, I agree with you. I hesitated at first, but then I realized that a lot of the
use of materialize() is really to signal a split point in a pipeline, and that returning the
PCollection instead of an Iterable will be more literate.

> Support new distributed execution engines (e.g., Spark)
> -------------------------------------------------------
>                 Key: CRUNCH-296
>                 URL:
>             Project: Crunch
>          Issue Type: Improvement
>          Components: Core
>            Reporter: Josh Wills
>            Assignee: Josh Wills
>         Attachments: CRUNCH-296.patch, CRUNCH-296b.patch, CRUNCH-296c.patch, CRUNCH-296d.patch,
> I've been working on this off-and-on for awhile, but it's currently in a state where
I feel like it's worth sharing: I came up with an implementation of the Crunch APIs that runs
on top of Apache Spark instead of MapReduce.
> My goal for this is pretty simple; I want to be able to change any instances of "new
MRPipeline(...)" to "new SparkPipeline(...)", not change anything else at all, and have my
pipelines run on Spark instead of as a series of MR jobs. Turns out that we can pretty much
do exactly that. Not everything works yet, but lots of things do-- joins and cogroups work,
the PageRank and TfIdf integration tests work. Some things that do not work that I'm aware
of: in-memory joins and some of the more complex file output handling rules, but I believe
that these things are fixable. Some thing that might work or might not: HBase inputs and outputs
on top of Spark.
> This is just an idea I had, and I would understand if other people don't want to work
on this or don't think it's the right direction for the project. My minimal request would
be to include the refactoring of the core APIs necessary to support plugging in new execution
frameworks so I can keep working on this stuff.

This message was sent by Atlassian JIRA

View raw message