[
https://issues.apache.org/jira/browse/BEAM-6766?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Claire McGinty updated BEAM-6766:
---------------------------------
Description:
Hi! Spotify has been internally prototyping and testing an implementation of
the sort merge join using Beam primitives and we're interested in contributing
it open-source – probably to Beam's extensions package in its own `smb` module
or as part of the joins module?
We've tested this with Avro files using Avro's GenericDatumWriter/Reader
directly (although this could theoretically be expanded to other serialization
formats). We'd add two transforms*, an SMB write and an SMB join.
SMB write would take in one PCollection and a # of buckets and:
1) Apply a partitioning function to the input to assign each record to one
bucket. (the user code would have to statically specify a # of buckets... hard
to see a way to do this dynamically.)
2) Group by that bucket ID and within each bucket perform an in-memory sort on
join key. If the grouped records are too large to fit in memory, fall back to
an external sort (although if this happens, user should probably increase
bucket size so every group fits in memory).
3) Directly write the contents of bucket to a sequentially named file.
4) Write a metadata file to the same output path with info about hash
algorithm/# buckets.
SMB join would take in the input paths for 2 or more Sources, all of which are
written in a bucketed and partitioned way, and :
1) Verify that the metadata files have compatible bucket # and hash algorithm.
2) Expand the input paths to enumerate the `ResourceIds` of every file in the
paths. Group all inputs with the same bucket ID.
3) Within each group, open a file reader on all `ResourceIds`. Sequentially
read files one record at a time, outputting tuples of all record pairs with
matching join key.
* These could be implemented either directly as `PTransforms` with the writer
being a `DoFn` but I semantically do like the idea of extending
`FileBasedSource`/`Sink` with abstract classes like
`SortedBucketSink`/`SortedBucketSource`... if we represent the elements in a
sink as KV pairs of <Bucket #, <Sorted Iterable of Records>>>, so that the # of
elements in the PCollection == # of buckets == # of output files, we could just
implement something like `SortedBucketSink` extending `FileBasedSink` with a
dynamic file naming function. I'd like to be able to take advantage of the
existing write/read implementation logic in the `io` package as much as
possible although I guess some of those are package private.
--
>From our internal testing, we've seen some substantial performance
>improvements using the right bucket size--not only by avoiding a shuffle
>during the join step, but also in storage costs, since we're getting better
>compression in Avro by storing sorted records.
Please let us know what you think/any concerns we can address! Our
implementation isn't quite production-ready yet, but we'd like to start a
discussion about it early.
was:
Hi! Spotify has been internally prototyping and testing an implementation of
the sort merge join using Beam primitives and we're interested in contributing
it open-source – probably to Beam's extensions/join-library package?
We've tested this with Avro files using Avro's GenericDatumWriter/Reader
directly (although this could theoretically be expanded to other serialization
types). We'd add two transforms, an SMB write and an SMB join.
SMB write would take in one PCollection and a # of buckets and:
1) Apply a partitioning function to the input to assign each record to one
bucket.
2) Group by that bucket ID and within each bucket perform an in-memory sort on
join key. If the grouped records are too large to fit in memory, fall back to
an external sort (although if this happens, user should probably increase
bucket size so every group fits in memory).
3) Directly write the contents of bucket to a sequentially named file.
4) Write a metadata file to the same output path with info about hash
algorithm/# buckets.
SMB join would take in the input paths for 2 or more Sources, all of which are
written in a bucketed and partitioned way, and :
1) Verify that the metadata files have compatible bucket # and hash algorithm.
2) Expand the input paths to enumerate the `ResourceIds` of every file in the
paths. Group all inputs with the same bucket ID.
3) Within each group, open a file reader on all `ResourceIds`. Sequentially
read files one record at a time, outputting tuples of all record pairs with
matching join key.
--
We've seen some substantial performance improvements using the right bucket
size--not only by avoiding a shuffle during the join step, but also in storage
costs, since we're getting better compression in Avro by storing sorted records.
Please let us know what you think/any concerns we can address! Our
implementation isn't quite production-ready yet, but we'd like to start a
discussion about it early.
> Sort Merge Bucket Join support in Beam
> --------------------------------------
>
> Key: BEAM-6766
> URL: https://issues.apache.org/jira/browse/BEAM-6766
> Project: Beam
> Issue Type: Improvement
> Components: io-ideas, sdk-java-join-library
> Reporter: Claire McGinty
> Priority: Major
>
> Hi! Spotify has been internally prototyping and testing an implementation of
> the sort merge join using Beam primitives and we're interested in
> contributing it open-source – probably to Beam's extensions package in its
> own `smb` module or as part of the joins module?
> We've tested this with Avro files using Avro's GenericDatumWriter/Reader
> directly (although this could theoretically be expanded to other
> serialization formats). We'd add two transforms*, an SMB write and an SMB
> join.
> SMB write would take in one PCollection and a # of buckets and:
> 1) Apply a partitioning function to the input to assign each record to one
> bucket. (the user code would have to statically specify a # of buckets...
> hard to see a way to do this dynamically.)
> 2) Group by that bucket ID and within each bucket perform an in-memory sort
> on join key. If the grouped records are too large to fit in memory, fall back
> to an external sort (although if this happens, user should probably increase
> bucket size so every group fits in memory).
> 3) Directly write the contents of bucket to a sequentially named file.
> 4) Write a metadata file to the same output path with info about hash
> algorithm/# buckets.
> SMB join would take in the input paths for 2 or more Sources, all of which
> are written in a bucketed and partitioned way, and :
> 1) Verify that the metadata files have compatible bucket # and hash algorithm.
> 2) Expand the input paths to enumerate the `ResourceIds` of every file in the
> paths. Group all inputs with the same bucket ID.
> 3) Within each group, open a file reader on all `ResourceIds`. Sequentially
> read files one record at a time, outputting tuples of all record pairs with
> matching join key.
>
> * These could be implemented either directly as `PTransforms` with the writer
> being a `DoFn` but I semantically do like the idea of extending
> `FileBasedSource`/`Sink` with abstract classes like
> `SortedBucketSink`/`SortedBucketSource`... if we represent the elements in a
> sink as KV pairs of <Bucket #, <Sorted Iterable of Records>>>, so that the #
> of elements in the PCollection == # of buckets == # of output files, we could
> just implement something like `SortedBucketSink` extending `FileBasedSink`
> with a dynamic file naming function. I'd like to be able to take advantage of
> the existing write/read implementation logic in the `io` package as much as
> possible although I guess some of those are package private.
> --
> From our internal testing, we've seen some substantial performance
> improvements using the right bucket size--not only by avoiding a shuffle
> during the join step, but also in storage costs, since we're getting better
> compression in Avro by storing sorted records.
> Please let us know what you think/any concerns we can address! Our
> implementation isn't quite production-ready yet, but we'd like to start a
> discussion about it early.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
