luceneutil
luceneutil copied to clipboard
mikemccand
Demonstrate ordinal map building on refresh
We're considering building ordinal maps eagerly. This can already be done to some extent using the current API. This PR shows how that should be done.
I'm don't know if the change is worth committing as is. The reader states that we compute are unused and there's no useful performance data we would be getting.
Testing: python3 src/python/localrun.py -source wikimedium10k -facets -nrt
Thank you for the review @mikemccand! I abandoned the idea of showing reader states computed on refresh, since there’s not a good place to showcase that in the benchmarks. Instead, I made a change to IndexState to pre-compute reader states that will be used lated by SearchPerfTest.
Hi @mikemccand! I came across this old PR again. I think what we're doing here makes sense. We precompute reader states, which will be stored in IndexState.ssdvFacetStates, and retrieved later in SearchPerfTest. Right now, when running the benchmarks, we're measuring the time we spend building those reader states, because we do it lazily.
Which do you prefer, loading the reader states eagerly (this PR) or lazily (existing functionality)?
Which do you prefer, loading the reader states eagerly (this PR) or lazily (existing functionality)?
+1 to do it eagerly (this PR) -- thank you! The change looks good -- since you are now a Lucene committer, I'll add you as collaborator here and then you can merge it? Yay :) Thank you!
