Last modified: 2014-05-28 18:46:32 UTC
Try this url: http://localhost:8000/huwiki/Vegy%C3%BCletek_%C3%B6sszegk%C3%A9plet-t%C3%A1bl%C3%A1zata?oldid=14167011 Puts Parsoid in a coma. Discovered via production logs after Parsoid cluster load spiked y'day where most parsoid processes were stuck.
It does look like a tokenizer issue: [subbu@earth tests] ./fetch-wt.js --prefix huwiki 14167011 > inf.loop.wt [subbu@earth tests] node parse --trace peg --prefix huwiki < inf.loop.wt ... some tokens emitted ... ... stuck ...
Change 135611 had a related patch set uploaded by GWicke: Bug 65812: Speed up processing of huge sync token chunks https://gerrit.wikimedia.org/r/135611
(In reply to Gerrit Notification Bot from comment #2) > Change 135611 had a related patch set uploaded by GWicke: > Bug 65812: Speed up processing of huge sync token chunks > > https://gerrit.wikimedia.org/r/135611 Sorry for the spam, this was actually intendend for bug 65812.
(In reply to Gabriel Wicke from comment #3) > Sorry for the spam, this was actually intendend for bug 65812. Never mind..
{{:Sablon:összegtáblázat}} is the transclusion in question in huwiki which generates a 408K token chunk in the tokenizer prior to the fix and which seemed to essentially slow down in async-ttm after about close to 128K tokens were processed and we traced this to a slowdown in concatenation once the accum size crossed a threshold.
Change 135611 merged by jenkins-bot: Bug 65812: Speed up processing of huge sync token chunks https://gerrit.wikimedia.org/r/135611
Fixed by https://gerrit.wikimedia.org/r/135611, and further improved by https://gerrit.wikimedia.org/r/135723 to the point where this huge test case now parses in about 66 seconds.
