spark-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Josh Rosen <rosenvi...@gmail.com>
Subject Re: [PySpark]: reading arbitrary Hadoop InputFormats
Date Thu, 07 Nov 2013 03:03:28 GMT
I opened a pull request to add custom serializer support to PySpark:
https://github.com/apache/incubator-spark/pull/146

My pull request adds the plumbing for transferring data from Java to Python
using formats other than Pickle.  For example, look at how textFile() uses
MUTF8Deserializer to read strings from Java.  Hopefully this provides all
of the functionality needed to support MsgPack.

- Josh


On Thu, Oct 31, 2013 at 11:11 AM, Josh Rosen <rosenville@gmail.com> wrote:

> Hi Nick,
>
> This is a nice start.  I'd prefer to keep the Java sequenceFileAsText()
> and newHadoopFileAsText() methods inside PythonRDD instead of adding them
> to JavaSparkContext, since I think these methods are unlikely to be used
> directly by Java users (you can add these methods to the PythonRDD
> companion object, which is how readRDDFromPickleFile is implemented:
> https://github.com/apache/incubator-spark/blob/branch-0.8/core/src/main/scala/org/apache/spark/api/python/PythonRDD.scala#L255
> )
>
> For MsgPack, the UnpicklingError is because the Python worker expects to
> receive its input in a pickled format.  In my prototype of custom
> serializers, I modified the PySpark worker to receive its
> serialization/deserialization function as input (
> https://github.com/JoshRosen/spark/blob/59b6b43916dc84fc8b83f22eb9ce13a27bc51ec0/python/pyspark/worker.py#L41)
> and added logic to pass the appropriate serializers based on each stage's
> input and output formats (
> https://github.com/JoshRosen/spark/blob/59b6b43916dc84fc8b83f22eb9ce13a27bc51ec0/python/pyspark/rdd.py#L42
> ).
>
> At some point, I'd like to port my custom serializers code to PySpark; if
> anyone's interested in helping, I'd be glad to write up some additional
> notes on how this should work.
>
> - Josh
>
> On Wed, Oct 30, 2013 at 2:25 PM, Nick Pentreath <nick.pentreath@gmail.com>wrote:
>
>> Thanks Josh, Patrick for the feedback.
>>
>> Based on Josh's pointers I have something working for JavaPairRDD ->
>> PySpark RDD[(String, String)]. This just calls the toString method on each
>> key and value as before, but without the need for a delimiter. For
>> SequenceFile, it uses SequenceFileAsTextInputFormat which itself calls
>> toString to convert to Text for keys and values. We then call toString
>> (again) ourselves to get Strings to feed to writeAsPickle.
>>
>> Details here: https://gist.github.com/MLnick/7230588
>>
>> This also illustrates where the "wrapper function" api would fit in. All
>> that is required is to define a T => String for key and value.
>>
>> I started playing around with MsgPack and can sort of get things to work
>> in
>> Scala, but am struggling with getting the raw bytes to be written properly
>> in PythonRDD (I think it is treating them as pickled byte arrays when they
>> are not, but when I removed the 'stripPickle' calls and amended the length
>> (-6) I got "UnpicklingError: invalid load key, ' '. ").
>>
>> Another issue is that MsgPack does well at writing "structures" - like
>> Java
>> classes with public fields that are fairly simple - but for example the
>> Writables have private fields so you end up with nothing being written.
>> This looks like it would require custom "Templates" (serialization
>> functions effectively) for many classes, which means a lot of custom code
>> for a user to write to use it. Fortunately for most of the common
>> Writables
>> a toString does the job. Will keep looking into it though.
>>
>> Anyway, Josh if you have ideas or examples on the "Wrapper API from
>> Python"
>> that you mentioned, I'd be interested to hear them.
>>
>> If you think this is worth working up as a Pull Request covering
>> SequenceFiles and custom InputFormats with default toString conversions
>> and
>> the ability to specify Wrapper functions, I can clean things up more, add
>> some functionality and tests, and also test to see if common things like
>> the "normal" Writables and reading from things like HBase and Cassandra
>> can
>> be made to work nicely (any other common use cases that you think make
>> sense?).
>>
>> Thoughts, comments etc welcome.
>>
>> Nick
>>
>>
>>
>> On Fri, Oct 25, 2013 at 11:03 PM, Patrick Wendell <pwendell@gmail.com
>> >wrote:
>>
>> > As a starting point, a version where people just write their own
>> "wrapper"
>> > functions to convert various HadoopFiles into String <K, V> files could
>> go
>> > a long way. We could even have a few built-in versions, such as dealing
>> > with Sequence files that are <String, String>. Basically, the user
>> needs to
>> > write a translator in Java/Scala that produces textual records from
>> > whatever format that want. Then, they make sure this is included in the
>> > classpath when running PySpark.
>> >
>> > As Josh is saying, I'm pretty sure this is already possible, but we may
>> > want to document it for users. In many organizations they might have 1-2
>> > people who can write the Java/Scala to do this but then many more people
>> > who are comfortable using python once it's setup.
>> >
>> > - Patrick
>> >
>> > On Fri, Oct 25, 2013 at 11:00 AM, Josh Rosen <rosenville@gmail.com>
>> wrote:
>> >
>> > > Hi Nick,
>> > >
>> > > I've seen several requests for SequenceFile support in PySpark, so
>> > there's
>> > > definitely demand for this feature.
>> > >
>> > > I like the idea of passing MsgPack'ed data (or some other structured
>> > > format) from Java to the Python workers.  My early prototype of custom
>> > > serializers (described at
>> > >
>> > >
>> >
>> https://cwiki.apache.org/confluence/display/SPARK/PySpark+Internals#PySparkInternals-customserializers
>> > > )
>> > > might be useful for implementing this.  Proper custom serializer
>> support
>> > > would handle the bookkeeping for tracking each stage's input and
>> output
>> > > formats and supplying the appropriate deserialization functions to the
>> > > Python worker, so the Python worker would be able to directly read the
>> > > MsgPack'd data that's sent to it.
>> > >
>> > > Regarding a wrapper API, it's actually possible to initially transform
>> > data
>> > > using Scala/Java and perform the remainder of the processing in
>> PySpark.
>> > >  This involves adding the appropriate compiled to the Java classpath
>> and
>> > a
>> > > bit of work in Py4J to create the Java/Scala RDD and wrap it for use
>> by
>> > > PySpark.  I can hack together a rough example of this if anyone's
>> > > interested, but it would need some work to be developed into a
>> > > user-friendly API.
>> > >
>> > > If you wanted to extend your proof-of-concept to handle the cases
>> where
>> > > keys and values have parseable toString() values, I think you could
>> > remove
>> > > the need for a delimiter by creating a PythonRDD from the
>> newHadoopFile
>> > > JavaPairRDD and adding a new method to writeAsPickle (
>> > >
>> > >
>> >
>> https://github.com/apache/incubator-spark/blob/master/core/src/main/scala/org/apache/spark/api/python/PythonRDD.scala#L224
>> > > )
>> > > to dump its contents as a pickled pair of strings.  (Aside: most of
>> > > writeAsPickle() would probably need be eliminated or refactored when
>> > adding
>> > > general custom serializer support).
>> > >
>> > > - Josh
>> > >
>> > > On Thu, Oct 24, 2013 at 11:18 PM, Nick Pentreath
>> > > <nick.pentreath@gmail.com>wrote:
>> > >
>> > > > Hi Spark Devs
>> > > >
>> > > > I was wondering what appetite there may be to add the ability for
>> > PySpark
>> > > > users to create RDDs from (somewhat) arbitrary Hadoop InputFormats.
>> > > >
>> > > > In my data pipeline for example, I'm currently just using Scala
>> (partly
>> > > > because I love it but also because I am heavily reliant on quite
>> custom
>> > > > Hadoop InputFormats for reading data). However, many users may
>> prefer
>> > to
>> > > > use PySpark as much as possible (if not for everything). Reasons
>> might
>> > > > include the need to use some Python library. While I don't do it
>> yet, I
>> > > can
>> > > > certainly see an attractive use case for using say scikit-learn /
>> numpy
>> > > to
>> > > > do data analysis & machine learning in Python. Added to this my
>> > cofounder
>> > > > knows Python well but not Scala so it can be very beneficial to do
a
>> > lot
>> > > of
>> > > > stuff in Python.
>> > > >
>> > > > For text-based data this is fine, but reading data in from more
>> complex
>> > > > Hadoop formats is an issue.
>> > > >
>> > > > The current approach would of course be to write an ETL-style
>> > Java/Scala
>> > > > job and then process in Python. Nothing wrong with this, but I was
>> > > thinking
>> > > > about ways to allow Python to access arbitrary Hadoop InputFormats.
>> > > >
>> > > > Here is a quick proof of concept:
>> > https://gist.github.com/MLnick/7150058
>> > > >
>> > > > This works for simple stuff like SequenceFile with simple Writable
>> > > > key/values.
>> > > >
>> > > > To work with more complex files, perhaps an approach is to
>> manipulate
>> > > > Hadoop JobConf via Python and pass that in. The one downside is of
>> > course
>> > > > that the InputFormat (well actually the Key/Value classes) must
>> have a
>> > > > toString that makes sense so very custom stuff might not work.
>> > > >
>> > > > I wonder if it would be possible to take the objects that are
>> yielded
>> > via
>> > > > the InputFormat and convert them into some representation like
>> > ProtoBuf,
>> > > > MsgPack, Avro, JSON, that can be read relatively more easily from
>> > Python?
>> > > >
>> > > > Another approach could be to allow a simple "wrapper API" such that
>> one
>> > > can
>> > > > write a wrapper function T => String and pass that into an
>> > > > InputFormatWrapper that takes an arbitrary InputFormat and yields
>> > Strings
>> > > > for the keys and values. Then all that is required is to compile
>> that
>> > > > function and add it to the SPARK_CLASSPATH and away you go!
>> > > >
>> > > > Thoughts?
>> > > >
>> > > > Nick
>> > > >
>> > >
>> >
>>
>
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message