Claire McGinty created BEAM-10042:
-------------------------------------
Summary: Add alternate constructor to improve byte encoding
performance in SortValues
Key: BEAM-10042
URL: https://issues.apache.org/jira/browse/BEAM-10042
Project: Beam
Issue Type: Improvement
Components: extensions-java-sorter
Reporter: Claire McGinty
The `SortValues` transform operates on key-groups of `KV<PrimaryKeyT,
Iterable<KV<SecondaryKeyT, ValueT>>>`. From those key groups it iterates
through each element and uses `CoderUtils.encodeToByteArray` on each
SecondaryKeyT-ValueT pair. This operation can be expensive and its parallelism
is limited by the # of key groups.
I'd like to propose adding an alternative to `SortValuesDoFn` that operates on
`KV<PrimaryKeyT, Iterable<KV<byte[], byte[]>>>` and can skip the encoding step
within the key-group. The user's pipeline may be able to encode the data to
bytes in a prior step in a much more parallelized and efficient way (i.e. in a
`MapElements` transform). I've seen performance gains in every Dataflow metric
from from patching this in my team's pipeline.
(I would visualize the alternative vs pre-existing constructors to look similar
to generic vs specific Avro constructors, where the generic constructor has a
static type and specific Avro has a parameterized T.)
What do you think?
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
