[
https://issues.apache.org/jira/browse/BEAM-11881?focusedWorklogId=563210&page=com.atlassian.jira.plugin.system.issuetabpanels:worklog-tabpanel#worklog-563210
]
ASF GitHub Bot logged work on BEAM-11881:
-----------------------------------------
Author: ASF GitHub Bot
Created on: 09/Mar/21 17:08
Start Date: 09/Mar/21 17:08
Worklog Time Spent: 10m
Work Description: TheNeuralBit commented on a change in pull request
#14135:
URL: https://github.com/apache/beam/pull/14135#discussion_r590552721
##########
File path: sdks/python/apache_beam/dataframe/partitionings_test.py
##########
@@ -36,8 +36,8 @@ class PartitioningsTest(unittest.TestCase):
}).set_index(['shape', 'color', 'size'])
def test_index_is_subpartition(self):
- ordered_list = [Nothing(), Index([3]), Index([1, 3]), Index(), Singleton()]
- for loose, strict in zip(ordered_list[:1], ordered_list[1:]):
+ ordered_list = [Singleton(), Index([3]), Index([1, 3]), Index(), Nothing()]
+ for loose, strict in zip(ordered_list[:-1], ordered_list[1:]):
Review comment:
Note the change to make this `-1`. This test was only verifying
`Singleton()` vs `Index([3])` then exiting the loop. The checks between
different `Index()` partitionings would have failed, but they weren't being
executed.
##########
File path: sdks/python/apache_beam/dataframe/frames.py
##########
@@ -130,7 +130,7 @@ def fillna(self, value, method, axis, **kwargs):
lambda df,
value: df.fillna(value, method=method, axis=axis, **kwargs),
[self._expr, value_expr],
- preserves_partition_by=partitionings.Singleton(),
+ preserves_partition_by=partitionings.Index(),
Review comment:
In general, what we previously called `preserves=Singleton()` (or
equivalently `preserves=Index()`), i.e. "this expression will preserve any
partitioning", now must be indicated by `preserves=Index()`.
Also what we previously called `preserves=Nothing()`, i.e. "this expression
cannot preserve any partitioning", is now represented with
`preserves=Singleton()`. This is perhaps a little surprising but I think it's
easy to reason about, just think of it as "the only thing this expression
preserves is Singleton() parititioning, which is trivially preserved".
All the `preserves=` updates are applying this mapping
##########
File path: sdks/python/apache_beam/dataframe/transforms.py
##########
@@ -331,8 +339,7 @@ def expr_to_stages(expr):
required_partitioning = expr.requires_partition_by()
for stage in common_stages([expr_to_stages(arg) for arg in expr.args()
if arg not in inputs]):
- if all(output_is_partitioned_by(arg, stage, required_partitioning)
- for arg in expr.args() if not is_scalar(arg)):
+ if output_partitioning_in_stage(expr, stage) is not None:
Review comment:
Instead of duplicating a portion of the logic from
`output_is_partitioned_by`, this new implementation moves all of the logic into
`output_partitioning_in_stage`. Here we just check for a None output indicating
this expression cannot be computed in this stage.
----------------------------------------------------------------
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:
[email protected]
Issue Time Tracking
-------------------
Worklog Id: (was: 563210)
Time Spent: 2h 40m (was: 2.5h)
> DataFrame subpartitioning order is incorrect
> --------------------------------------------
>
> Key: BEAM-11881
> URL: https://issues.apache.org/jira/browse/BEAM-11881
> Project: Beam
> Issue Type: Bug
> Components: sdk-py-core
> Reporter: Brian Hulette
> Priority: P2
> Labels: dataframe-api
> Time Spent: 2h 40m
> Remaining Estimate: 0h
>
> Currently we've defined
> Nothing() < Index([i]) < Index([i,j]) < .. < Index() < Singleton()
> s.t. Singleton is a subpartitoning of Index, is a subpartitioning of
> Index([i,j]), but this is incorrect. The order should be
> Singleton() < Index([i]) < Index([i,j]) < .. < Index() < Nothing()
> s.t. every other partitioning is a subpartitioning of Singleton. This is
> logical, since Singleton will collect the largest amount of data on a single
> node, partitioning by a single index will be alittle more distributed, and
> partitioning by the full Index() will be the most distribtued.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
