claudevdm commented on code in PR #38015:
URL: https://github.com/apache/beam/pull/38015#discussion_r3017858877
##########
sdks/python/apache_beam/io/gcp/bigquery_change_history.py:
##########
@@ -1170,16 +1315,46 @@ def expand(self, pbegin: beam.pvalue.PBegin) ->
beam.PCollection:
row_filter=self._row_filter))
| 'CommitQueryResults' >> beam.Reshuffle())
+ emit_raw = self._decompress_shards is not None
+
+ read_sdf = beam.ParDo(
+ _ReadStorageStreamsSDF(
+ batch_arrow_read=self._batch_arrow_read,
+ change_timestamp_column=self._change_timestamp_column,
+ max_split_rounds=self._max_split_rounds,
+ emit_raw_batches=emit_raw))
+ if emit_raw:
+ read_sdf = read_sdf.with_output_types(Tuple[bytes, bytes])
+ else:
+ read_sdf = read_sdf.with_output_types(Dict[str, Any])
+
read_outputs = (
query_results
- | 'ReadStorageStreams' >> beam.ParDo(
- _ReadStorageStreamsSDF(
- batch_arrow_read=self._batch_arrow_read,
- change_timestamp_column=self._change_timestamp_column)).
- with_outputs(_CLEANUP_TAG, main='rows'))
+ | 'ReadStorageStreams' >> read_sdf.with_outputs(
+ _CLEANUP_TAG, main='rows'))
_ = (
read_outputs[_CLEANUP_TAG]
| 'CleanupTempTables' >> beam.ParDo(_CleanupTempTablesFn()))
- return read_outputs['rows']
+ if emit_raw:
+ # Fan out raw Arrow batches across decompress_shards workers
+ # via GBK, then decompress and convert to timestamped row dicts.
+ # Uses a discarding trigger so GBK fires per-element without
+ # waiting for the GlobalWindow to close.
+ num_shards = self._decompress_shards
+ rows = (
+ read_outputs['rows']
+ | 'ShardBatches' >>
+ beam.WithKeys(lambda _, n=num_shards: random.randint(0, n - 1))
+ | 'WindowForGBK' >> beam.WindowInto(
+ GlobalWindows(),
+ trigger=beam_trigger.Repeatedly(beam_trigger.AfterCount(1)),
+ accumulation_mode=(beam_trigger.AccumulationMode.DISCARDING))
+ | 'GroupByShardKey' >> beam.GroupByKey()
Review Comment:
Good idea. I am not sure what the best approach is tbh.
1. Reshuffle instead of grouping into batches. One arrow batch per key.
2. Limited amount of batching (or maybe configurable). Probably do not want
too many arrow batches per key single decompression is expensive.
I am inclined to just replace with reshuffle instead of limiting the
parallelism? The only benefit of batching/limiting keys is combiner lifting,
but arrow batches are already thousands of rows so that might be unnecessary.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
