Page MenuHomePhabricator

[M] Analyse results from the synonyms patch
Closed, ResolvedPublic

Description

Once the synonyms patch is deployed to production I think we ought to:

  • modify AnalyzeResults.php so that it gathers response times from api calls
  • run AnalyzeResults.php with the patch turned on and off (after having run it once more, to prepopulate cache), and paste the results in a comment - also gather data for response times when the patch is on/off and paste here (e.g. median response time, average response time of slowest 10% of calls, average response time of fastest 10% of calls)
  • discuss results and determine whether to remove the feature flag from T258053 or whether more work is necessary and if so, what work is necessary

Event Timeline

Cparle updated the task description. (Show Details)
CBogen subscribed.

Possible follow on work to be discussed:

  • gather and rate new data relevant to synonyms (e.g. data for non-English languages) and store it in the labeled data repository with the tag "synonyms"
  • run AnalyzeResults.php with the patch turned on and off just on the data tagged with "synonyms", and paste the results in a comment
CBogen renamed this task from Analyse results from the synonyms patch to [M] Analyse results from the synonyms patch.Sep 29 2021, 4:38 PM

... actually, never mind - going to try again from toolforge to reduce noise due to network variance

Here are the results of the analysis run on toolforge

Synonyms

F1 Score0.63958625148152
Precision@100.88936170212766
Precision@250.8572319201995
Precision@500.82246199117214
Precision@1000.79898516783763
Recall0.56042296072508
Average precision0.47271739186655
Min response time82ms
Max response time10205ms
Average response time367ms
Median response time249ms
Average fastest 10%116ms
Average slowest 10%1383ms
Average 2nd slowest 10%487ms

Control (without synonyms)

F1 Score0.64963041182682
Precision@100.87753510140406
Precision@250.84721424468696
Precision@500.81352087114338
Precision@1000.78035520115984
Recall0.58180442595044
Average precision0.4853565651381
Min response time76ms
Max response time2084ms
Average response time211ms
Median response time180ms
Average fastest 10%99ms
Average slowest 10%456ms
Average 2nd slowest 10%373ms

Ok so the upshot of the above is some queries take a very long time - e.g. searching for "wikimedia commons" with the synonyms patch on causes a timeout.