Search suggestions can make searches very slow in 3.11

[Jeff Davis]

EG 3.11

Searches can be extremely slow on systems with many bib records, due to the time required to process a large number of potential search suggestions.

On our 3.11.1 test server with 4 million bib records, a keyword search for "rainbew" takes >23 seconds to complete. Almost all of that search time is due to the search.symspell_lookup call in search.symspell_suggest, which has to process 62,000 potential search suggestions in order to return the 3 suggestions we actually want.

This test environment has 7.2 million search.symspell_dictionary entries which were freshly generated using the sideload process. It uses the default values for internal flags (symspell.prefix_length = 6, symspell.max_edit_distance = 3), and the keyword metabib class has max_suggestions = 3.

[Galen Charlton (gmc)]

Confirmed in a test database of 1.9 million bibs, where the search.symspell_lookup() time is about 10 times that of a 3.9 system with the same dataset.

[Galen Charlton (gmc)]

This looks like in part a regression on bug 1931162; as a workaround, dropping symspell.max_edit_distance down to 2 should reduce the number of short prefixes with large number of suggestions that get checked.

[Mike Rylander (mrylander)]

Galen is correct, the core "smarts" of the optimizations embodied in the 1282 upgrade script -- in-function caching, and better operation ordering -- were lost during later development, either by starting that development on top of the version of the code immediately prior to 1282 or during merge conflict cleanup later on.

A branch is available at https://git.evergreen-ils.org/?p=working/Evergreen.git;a=shortlog;h=refs/heads/user/miker/lp-2038472-DYM-optimization-regression with (another) drop-in replacement for search.symspell_lookup that reintroduces those optimizations that were lost.

[Galen Charlton (gmc)]

Tested and a signoff branch pushed to working/user/gmcharlt/lp2038472_signoff

Jeff: I'm curious how this does on your test system.

[Jeff Davis (jdavis-sitka)]

I have tested this code and consent to signing off on it with my name, Jeff Davis, and my email address, <email address hidden>.

On our test system, this fix substantially improved search times and appeared to resolve the issue identified in the bug report. We decided to leave Did You Mean turned off in our environment for now, so our testing of this fix hasn't been quite as definitive as I would like, but I think it should be included in a release.

[Galen Charlton (gmc)]

Pushed down to rel_3_11. Thanks, Mike and Jeff!