dsmiley commented on PR #4652:
URL: https://github.com/apache/solr/pull/4652#issuecomment-6608422285
Again, had Claude Opus re-aquiant myself with this one... (I'm quoting some
key aspects of its analysis)
> Part 2, the eDisMax empty-clause fix, is safe and could merge now.
Agree.
----
:robot: Part 1, dropping DisMax's pf boost for single-term queries, is the
risky one. It's correct and well built, but it removes a boost that DisMax
applies today, and users can't opt out
Concerns about disturbing users:
1. pf-only fields lose all effect on single-term queries. Take qf=text
pf=title^5, a common setup. Today q=wireless boosts docs with "wireless" in
title. After this
change, title does nothing for any single-term query. Single-term
queries are a large share of real traffic. The ref guide and upgrade notes say
so, but this is
a real ranking regression, not a cosmetic one.
2. Different weights shift too. With qf=title^2 body and pf=title^10
body^5, the single-term pf boost works today as an extra field weighting.
Removing it changes
rank order even when pf and qf list the same fields.
3. The case for it is mostly consistency. It matches eDisMax and fits
phrase semantics. Performance gains are small and not benchmarked. Nothing is
broken today.
People still on classic DisMax are likely long-running deployments that
tuned weights by hand and least want ranking churn.
4. No opt-out. For a ranking change with no bug behind it, I'd want an
escape hatch, or at least the 11.0-only restriction plus the existing
major-changes-in-solr-11 note.
----
I agree with all that. So in a nuthshell; I'm really hesitent to move this
forward, sorry.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
