Our firefox browser tests running on local instances kept complaining of losing connection to the firefox web driver after 45 seconds. We never really managed to figure out the issue and went with throttling the browsertests to ensure that only one run per instance ( https://gerrit.wikimedia.org/r/#/c/155337/ ).
Timo proposed a patch to add chromium and a xvfb server on the contint slaves which propose to maintain an xvfb service on display port 99. That is what triggered the race condition warning to me.
In mediawiki_selenium we use ruby gems headless defaults:
require "headless" headless = Headless.new headless.start
Then destroy it:
at_exit do headless.destroy end
Looking at headless code ( https://github.com/leonid-shevtsov/headless/blob/master/lib/headless.rb ) the default are:
- display port 99
- auto picking disabled (since display port is set)
- reuses display
When the first test start, that boot an xvfb on display port 99. A second test starting in parallel will end up reusing the same display. Imho that would cause a bunch of race conditions between the two browsers being run.
mediawiki_selenium should probably be invoked with:
Headless.new(autopick: true, reuse: false)
We could have that tweakable via env variable such as HEADLESS_DISPLAY, HEADLESS_AUTOPICK and HEADLESS_REUSE.
Version: wmf-deployment
Severity: normal