[
https://issues.apache.org/jira/browse/BEAM-10409?focusedWorklogId=466929&page=com.atlassian.jira.plugin.system.issuetabpanels:worklog-tabpanel#worklog-466929
]
ASF GitHub Bot logged work on BEAM-10409:
-----------------------------------------
Author: ASF GitHub Bot
Created on: 05/Aug/20 18:18
Start Date: 05/Aug/20 18:18
Worklog Time Spent: 10m
Work Description: yifanmai commented on a change in pull request #12185:
URL: https://github.com/apache/beam/pull/12185#discussion_r465915603
##########
File path:
sdks/python/apache_beam/runners/portability/fn_api_runner/translations.py
##########
@@ -685,6 +687,264 @@ def fix_side_input_pcoll_coders(stages, pipeline_context):
return stages
+def eliminate_common_siblings(stages, context):
+ # type: (Iterable[Stage], TransformContext) -> Iterable[Stage]
+ """Runs common subexpression elimination for common siblings.
+
+ If stages have common input, an identical transform, and one output each,
+ then all but one stages will be eliminated, and the output of the remaining
+ will be connected to the original output PCollections of the eliminated
+ stages. This elimination runs only once, not recursively, and will only
+ eliminate the first stage after a common input, rather than a chain of
+ stages.
+ """
+
+ SiblingKey = collections.namedtuple(
+ 'SiblingKey', ['spec_urn', 'spec_payload', 'inputs', 'environment_id'])
+
+ def get_sibling_key(transform):
+ """Returns a key that will be identical for common siblings."""
+ transform_output_keys = list(transform.outputs.keys())
+ # Return None as the sibling key for ineligible transforms.
+ if len(transform_output_keys
+ ) != 1 or transform.spec.urn != common_urns.primitives.PAR_DO.urn:
+ return None
+ return SiblingKey(
+ spec_urn=transform.spec.urn,
+ spec_payload=transform.spec.payload,
+ inputs=tuple(transform.inputs.items()),
+ environment_id=transform.environment_id)
+
+ # Group stages by keys.
+ stages_by_sibling_key = collections.defaultdict(list)
+ for stage in stages:
+ transform = only_transform(stage.transforms)
+ stages_by_sibling_key[get_sibling_key(transform)].append(stage)
+
+ # Eliminate stages and build the output PCollection remapping dictionary.
+ pcoll_id_remap = {}
+ for sibling_key, sibling_stages in stages_by_sibling_key.items():
+ if sibling_key is None or len(sibling_stages) == 1:
+ continue
+ output_pcoll_ids = [
+ only_element(stage.transforms[0].outputs.values())
+ for stage in sibling_stages
+ ]
+ to_delete_pcoll_ids = output_pcoll_ids[1:]
+ for to_delete_pcoll_id in to_delete_pcoll_ids:
+ pcoll_id_remap[to_delete_pcoll_id] = output_pcoll_ids[0]
+ del context.components.pcollections[to_delete_pcoll_id]
+ del sibling_stages[1:]
+
+ # Yield stages while remapping output PCollections if needed.
+ for sibling_key, sibling_stages in stages_by_sibling_key.items():
+ for stage in sibling_stages:
+ input_keys_to_remap = []
+ for input_key, input_pcoll_id in stage.transforms[0].inputs.items():
+ if input_pcoll_id in pcoll_id_remap:
+ input_keys_to_remap.append(input_key)
+ for input_key_to_remap in input_keys_to_remap:
+ stage.transforms[0].inputs[input_key_to_remap] = pcoll_id_remap[
+ stage.transforms[0].inputs[input_key_to_remap]]
+ yield stage
+
+
+def pack_combiners(stages, context):
+ # type: (Iterable[Stage], TransformContext) -> Iterator[Stage]
+ """Packs sibling CombinePerKey stages into a single CombinePerKey.
+
+ If CombinePerKey stages have a common input, one input each, and one output
+ each, pack the stages into a single stage that runs all CombinePerKeys and
+ outputs resulting tuples to a new PCollection. A subsequent stage unpacks
+ tuples from this PCollection and sends them to the original output
+ PCollections.
+ """
+
+ class _UnpackFn(core.DoFn):
+ """A DoFn that unpacks a packed to multiple tagged outputs.
+
+ Example:
+ tags = (T1, T2, ...)
+ input = (K, (V1, V2, ...))
+ output = TaggedOutput(T1, (K, V1)), TaggedOutput(T2, (K, V1)), ...
+ """
+
+ def __init__(self, tags):
+ self._tags = tags
+
+ def process(self, element):
+ key, values = element
+ return [
+ core.pvalue.TaggedOutput(tag, (key, value))
+ for tag, value in zip(self._tags, values)
+ ]
+
+ def _get_fallback_coder_id():
+ return context.add_or_get_coder_id(
+ coders.registry.get_coder(object).to_runner_api(None))
+
+ def _get_component_coder_id_from_kv_coder(coder, index):
+ assert index < 2
+ if coder.spec.urn == common_urns.coders.KV.urn and len(
+ coder.component_coder_ids) == 2:
+ return coder.component_coder_ids[index]
+ return _get_fallback_coder_id()
+
+ def _get_key_coder_id_from_kv_coder(coder):
+ return _get_component_coder_id_from_kv_coder(coder, 0)
+
+ def _get_value_coder_id_from_kv_coder(coder):
+ return _get_component_coder_id_from_kv_coder(coder, 1)
+
+ def _try_fuse_stages(a, b):
+ if a.can_fuse(b, context):
+ return a.fuse(b)
+ else:
+ raise ValueError
+
+ def _try_merge_environments(env1, env2):
Review comment:
Yup, replaced with `Stage._merge_environments` instead.
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
Issue Time Tracking
-------------------
Worklog Id: (was: 466929)
Time Spent: 2.5h (was: 2h 20m)
> Add combiner packing to graph optimizer phases
> ----------------------------------------------
>
> Key: BEAM-10409
> URL: https://issues.apache.org/jira/browse/BEAM-10409
> Project: Beam
> Issue Type: Improvement
> Components: runner-core
> Reporter: Yifan Mai
> Assignee: Yifan Mai
> Priority: P2
> Time Spent: 2.5h
> Remaining Estimate: 0h
>
> Some use cases of Beam (e.g. [TensorFlow
> Transform|https://github.com/tensorflow/transform]) create thousands of
> Combine stages with a common parent. The large number of stages can cause
> performance issues on some runners. To alleviate, a graph optimization phase
> could be added to the translations module that packs compatible Combine
> stages into a single stage.
> The graph optimization for CombinePerKey would work as follows: If
> CombinePerKey stages have a common input, one input each, and one output
> each, pack the stages into a single stage that runs all CombinePerKeys and
> outputs resulting tuples to a new PCollection. A subsequent stage unpacks
> tuples from this PCollection and sends them to the original output
> PCollections.
> There is an additional issue with supporting this for CombineGlobally:
> because of the intermediate KeyWithVoid stage between the CombinePerKey
> stages and the input stage, the CombinePerKey stages do not have a common
> input stage, and cannot be packed. To support CombineGlobally, a common
> sibling elimination graph optimization phase can be used to combine the
> KeyWithVoid stages. After this, the CombinePerKey stages would have a common
> input and can be packed.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
