I'm still trying to optimize ikiwiki for a site using album, and checking which pages depend on which pages is still taking too long. Here's another go at fixing that, using Will's suggestion from should optimise pagespecs:
A hash, by itself, is not optimal because the dependency list holds two things: page names and page specs. The hash would work well for the page names, but you'll still need to iterate through the page specs. I was thinking of keeping a list and a hash. You use the list for pagespecs and the hash for individual page names. To make this work you need to adjust the API so it knows which you're adding. -- Will
If you have P pages and refresh after changing C of them, where an average page has E dependencies on exact page names and D other dependencies, this branch should drop the complexity of checking dependencies from O(P * (D+E) * C) to O(C + PE + PD*C). Pages that use inline or map have a large value for E (e.g. one per inlined page) and a small value for D (e.g. one per inline).
Test 1: a wiki with about 3500 pages and 3500 photos, and a change that touches about 350 pages and 350 photos
Test 2: the docwiki (about 700 objects not excluded by docwiki.setup, mostly pages), docwiki.setup modified to turn off verbose, and a change that touches the 98 pages of plugins/*.mdwn
In both tests I rebuilt the wiki with the target ikiwiki version, then touched the appropriate pages and refreshed.
Results of test 1: without this branch it took around 5:45 to rebuild and around 5:45 again to refresh (so rebuilding 10% of the pages, then deciding that most of the remaining 90% didn't need to change, took about as long as rebuilding everything). With this branch it took 5:47 to rebuild and 1:16 to refresh.
Results of test 2: rebuilding took 14.11s without, 13.96s with; refreshing three times took 7.29/7.40/7.37s without, 6.62/6.56/6.63s with.
(This benchmarking was actually done with my album branch, since that's what the huge wiki needs; that branch doesn't alter core code beyond the ready/depends-exact branch point, so the results should be equally valid.)
We discussed this on irc; I had some worries that things may have been switched to
add_depends_exactthat were not pure page names. My current feeling is it's all safe, but who knows. It's easy to miss something. Which makes me think this is not a good interface.
Why not, instead, make
add_dependssmart. If it's passed something that is clearly a raw page name, it can add it to the exact depends hash. Else, add it to the pagespec hash. You can tell if it's a pure page name by matching on
Good thinking. Done in commit 68ce514a 'Auto-detect "simple dependencies"', with a related bugfix in e8b43825 "Force %depends_exact to lower case".
Performance impact: Test 2 above takes 0.2s longer to rebuild (probably from all the calls to lc, which are, however, necessary for correctness) and has indistinguishable performance for a refresh.
Test 1 took about 6 minutes to rebuild and 1:25 to refresh; those are pessimistic figures, since I've added 90 more photos and 90 more pages (both to the wiki as a whole, and the number touched before refreshing) since testing the previous version of this branch. --smcv
Also I think there may be little optimisation value left in 7227c2debfeef94b35f7d81f42900aa01820caa3, since the "regular" dependency lists will be much shorter.
You're probably right, but IMO it's not worth reverting it - a set (hash with dummy values) is still the right data structure. --smcv
Sounds like inline pagenames has an already exstant bug WRT pages moving, which this should not make worse. Would be good to verify.
If you mean the standard "add a better match for a link-like construct" bug that also affects sidebar, then yes, it does have the bug, but I'm pretty sure this branch doesn't make it any worse. I could solve this at the cost of making pagenames less useful for interactive use, by making it not respect LinkingRules, but instead always interpret its paths as relative to the top of the wiki - that's actually all that album needs. --smcv
Re coding, it would be nice if
refresh()could avoid duplicating the debug message, etc in the two cases. --Joey
Fixed in commit f805d566 "Avoid duplicating debug message..." --smcv