[
https://issues.apache.org/jira/browse/BEAM-8151?focusedWorklogId=342878&page=com.atlassian.jira.plugin.system.issuetabpanels:worklog-tabpanel#worklog-342878
]
ASF GitHub Bot logged work on BEAM-8151:
----------------------------------------
Author: ASF GitHub Bot
Created on: 13/Nov/19 20:56
Start Date: 13/Nov/19 20:56
Worklog Time Spent: 10m
Work Description: lukecwik commented on issue #10091: (WIP) [BEAM-8151]
Swap to create SdkWorkers on demand when processing portable jobs
URL: https://github.com/apache/beam/pull/10091#issuecomment-553598687
I ran the performance benchmark using the fn_api_runner_microbenchmark:
before (at commit 80e8fb59b6d785eb462e6a05b91d029fd8ff50c6):
```
run_single_pipeline, 1 element(s) at start, 100 growth per run: run 1 of 10,
per element time cost: 3.0014 sec
run_single_pipeline, 1 element(s) at start, 100 growth per run: run 2 of 10,
per element time cost: 0.0394958 sec
run_single_pipeline, 1 element(s) at start, 100 growth per run: run 3 of 10,
per element time cost: 0.0220158 sec
run_single_pipeline, 1 element(s) at start, 100 growth per run: run 4 of 10,
per element time cost: 0.0160263 sec
run_single_pipeline, 1 element(s) at start, 100 growth per run: run 5 of 10,
per element time cost: 0.0130143 sec
run_single_pipeline, 1 element(s) at start, 100 growth per run: run 6 of 10,
per element time cost: 0.0111442 sec
run_single_pipeline, 1 element(s) at start, 100 growth per run: run 7 of 10,
per element time cost: 0.00963751 sec
run_single_pipeline, 1 element(s) at start, 100 growth per run: run 8 of 10,
per element time cost: 0.00856183 sec
run_single_pipeline, 1 element(s) at start, 100 growth per run: run 9 of 10,
per element time cost: 0.00796899 sec
run_single_pipeline, 1 element(s) at start, 100 growth per run: run 10 of
10, per element time cost: 0.00760294 sec
Fixed cost 3.49717886575063
Per-element 0.003790947336139101
R^2 0.963500601942131
```
after:
```
run_single_pipeline, 1 element(s) at start, 100 growth per run: run 1 of 10,
per element time cost: 2.8741 sec
run_single_pipeline, 1 element(s) at start, 100 growth per run: run 2 of 10,
per element time cost: 0.0401369 sec
run_single_pipeline, 1 element(s) at start, 100 growth per run: run 3 of 10,
per element time cost: 0.0219585 sec
run_single_pipeline, 1 element(s) at start, 100 growth per run: run 4 of 10,
per element time cost: 0.0161115 sec
run_single_pipeline, 1 element(s) at start, 100 growth per run: run 5 of 10,
per element time cost: 0.0124523 sec
run_single_pipeline, 1 element(s) at start, 100 growth per run: run 6 of 10,
per element time cost: 0.0108133 sec
run_single_pipeline, 1 element(s) at start, 100 growth per run: run 7 of 10,
per element time cost: 0.00963711 sec
run_single_pipeline, 1 element(s) at start, 100 growth per run: run 8 of 10,
per element time cost: 0.00862622 sec
run_single_pipeline, 1 element(s) at start, 100 growth per run: run 9 of 10,
per element time cost: 0.00795245 sec
run_single_pipeline, 1 element(s) at start, 100 growth per run: run 10 of
10, per element time cost: 0.00736677 sec
Fixed cost 3.4638128309249883
Per-element 0.003727303158153187
R^2 0.9529920831060255
```
----------------------------------------------------------------
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: 342878)
Time Spent: 10h 10m (was: 10h)
> Allow the Python SDK to use many many threads
> ---------------------------------------------
>
> Key: BEAM-8151
> URL: https://issues.apache.org/jira/browse/BEAM-8151
> Project: Beam
> Issue Type: Improvement
> Components: sdk-py-core, sdk-py-harness
> Reporter: Luke Cwik
> Assignee: Luke Cwik
> Priority: Major
> Time Spent: 10h 10m
> Remaining Estimate: 0h
>
> We need to use a thread pool which shrinks the number of active threads when
> they are not being used.
>
> This is to prevent any stuckness issues related to a runner scheduling more
> work items then there are "work" threads inside the SDK harness.
>
> By default the control plane should have all "requests" being processed in
> parallel and the runner is responsible for not overloading the SDK with too
> much work.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
