[
https://issues.apache.org/jira/browse/BEAM-7131?focusedWorklogId=244979&page=com.atlassian.jira.plugin.system.issuetabpanels:worklog-tabpanel#worklog-244979
]
ASF GitHub Bot logged work on BEAM-7131:
----------------------------------------
Author: ASF GitHub Bot
Created on: 20/May/19 10:24
Start Date: 20/May/19 10:24
Worklog Time Spent: 10m
Work Description: robertwb commented on pull request #8558: [BEAM-7131]
Spark: cache executable stage output to prevent re-computation
URL: https://github.com/apache/beam/pull/8558#discussion_r285523785
##########
File path:
runners/spark/src/main/java/org/apache/beam/runners/spark/translation/SparkBatchPortablePipelineTranslator.java
##########
@@ -224,6 +224,11 @@ private static void translateImpulse(
MetricsAccumulator.getInstance());
JavaRDD<RawUnionValue> staged = inputRdd.mapPartitions(function);
+ // Prevent potentially expensive re-computation of executable stage
+ if (outputs.size() > 1) {
+ staged.cache();
Review comment:
That's really unfortunate.
So the current design is to cache, when the (by default True flag is
enabled) (1) collections that are used more than once and (2) all outputs of
transforms with multiple outputs.
That seems a reasonable first step. I think we can do better (e.g. cache
all but one output from multi-output transforms, uncache once we're done with
all consumers), but could probably call that future work.
This PR also seems to change a bunch of stuff with how coding works. Could
you factor that out into a separate PR?
----------------------------------------------------------------
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: 244979)
Time Spent: 3h 10m (was: 3h)
> Spark portable runner appears to be repeating work (in TFX example)
> -------------------------------------------------------------------
>
> Key: BEAM-7131
> URL: https://issues.apache.org/jira/browse/BEAM-7131
> Project: Beam
> Issue Type: Bug
> Components: runner-spark
> Reporter: Kyle Weaver
> Assignee: Kyle Weaver
> Priority: Major
> Time Spent: 3h 10m
> Remaining Estimate: 0h
>
> I've been trying to run the TFX Chicago taxi example [1] on the Spark
> portable runner. TFDV works fine, but the preprocess step
> (preprocess_flink.sh [2]) fails with the following error:
> RuntimeError: AlreadyExistsError: file already exists [while running
> 'WriteTransformFn/WriteTransformFn']
> Assets are being written multiple times to different temp directories, which
> is okay, but the error occurs when they are copied to the same permanent
> output directory. Specifically, the copy tree operation in transform_fn_io.py
> [3] is run twice with the same output directory. The error doesn't occur when
> that code is modified to allow overwriting existing files, but that's only a
> shallow fix. While the TF transform should probably be made idempotent, this
> is also an issue with the Spark runner, which shouldn't be repeating work
> like this regularly (in the absence of a failure condition).
> [1] [https://github.com/tensorflow/tfx/tree/master/tfx/examples/chicago_taxi]
> [2]
> [https://github.com/tensorflow/tfx/blob/master/tfx/examples/chicago_taxi/preprocess_flink.sh]
> [3]
> [https://github.com/tensorflow/transform/blob/master/tensorflow_transform/beam/tft_beam_io/transform_fn_io.py#L33-L45]
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
