zacharymorn commented on PR #972:
URL: https://github.com/apache/lucene/pull/972#issuecomment-1168197720
> > I feel the effect would be similar?
>
> Indeed, sorry I had misread your code!
>
No worry, thanks still for the suggestion!
>
> No, it shouldn't matter. Bulk scorers sometimes help yield better
performance because it's easier for them to amortize computation across docs,
but if they don't yield better performance, there's no point in using a bulk
scorer instead of a regular scorer.
Ok I see, makes sense.
> I agree that it looks like a great speedup, we should get this in! The
benchmark only tests performance of top-level disjunctions of term queries that
have two clauses. I'd be curious to get performance numbers for queries like
the below ones to see if we need to fine-tune a bit more when this new scorer
gets used. Note that I don't think we need to get the performance better for
all these queries to merge the change, we could start by only using this new
scorer for the (common) case of a top-level disjunction of 2 term queries, and
later see if this scorer can handle more disjunctions.
>
> ```
> OrAndHigMedAndHighMed: (+including +looking) (+date +finished) #
disjunction of conjunctions, which don't have as good score upper bounds as
term queries
> OrHighPhraseHighPhrase: "united states" "new york" # disjunction of phrase
queries, which don't have as good score upper bounds as term queries and are
slow to advance
> AndHighOrMedMed: +be +(mostly interview) # disjunction within conjunction
that leads iteration
> AndMedOrHighHigh: +interview +(at united) # disjunction within conjunction
that doesn't lead iteration
> ```
Sounds good! I have run these queries through benchmark and the results look
somewhat consistent:
```
TaskQPS baseline StdDevQPS
my_modified_version StdDev Pct diff p-value
OrHighPhraseHighPhrase 28.89 (8.7%) 24.19
(4.7%) -16.3% ( -27% - -3%) 0.000
AndHighOrMedMed 101.24 (6.6%) 101.09
(3.0%) -0.1% ( -9% - 10%) 0.927
AndMedOrHighHigh 81.44 (6.3%) 81.62
(3.7%) 0.2% ( -9% - 10%) 0.895
OrAndHigMedAndHighMed 128.26 (7.0%) 136.94
(3.7%) 6.8% ( -3% - 18%) 0.000
PKLookup 221.47 (11.7%) 236.93
(9.1%) 7.0% ( -12% - 31%) 0.035
```
```
TaskQPS baseline StdDevQPS
my_modified_version StdDev Pct diff p-value
OrHighPhraseHighPhrase 27.73 (9.1%) 23.73
(4.6%) -14.4% ( -25% - 0%) 0.000
AndHighOrMedMed 97.09 (13.1%) 99.30
(4.3%) 2.3% ( -13% - 22%) 0.462
AndMedOrHighHigh 75.87 (15.2%) 80.04
(5.7%) 5.5% ( -13% - 31%) 0.128
PKLookup 219.70 (15.7%) 238.75
(12.4%) 8.7% ( -16% - 43%) 0.053
OrAndHigMedAndHighMed 121.83 (13.7%) 134.79
(4.4%) 10.6% ( -6% - 33%) 0.001
```
```
TaskQPS baseline StdDevQPS
my_modified_version StdDev Pct diff p-value
OrHighPhraseHighPhrase 27.42 (16.2%) 23.99
(4.0%) -12.5% ( -28% - 9%) 0.001
AndHighOrMedMed 96.61 (15.8%) 100.09
(3.6%) 3.6% ( -13% - 27%) 0.321
AndMedOrHighHigh 75.72 (16.8%) 79.53
(4.9%) 5.0% ( -14% - 32%) 0.200
OrAndHigMedAndHighMed 122.33 (16.9%) 136.60
(4.5%) 11.7% ( -8% - 39%) 0.003
PKLookup 207.94 (21.6%) 233.10
(16.5%) 12.1% ( -21% - 63%) 0.046
```
Looks like we may need to restrict the scorer to only term queries, or
improve it for phrase queries?
--
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: issues-unsubscr...@lucene.apache.org
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org
For additional commands, e-mail: issues-h...@lucene.apache.org
