[
https://issues.apache.org/jira/browse/SOLR-13132?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17153994#comment-17153994
]
Michael Gibney commented on SOLR-13132:
---------------------------------------
I just attached [^SOLR-13132-benchmarks.tgz], some naive/sanity-check
benchmarks. The results below are mostly for without any filterCache, but I
included hooks in the included scripts to easily change the filterCache size
for evaluation. For purpose of this evaluation, fgSet == base search result
domain. All results discussed here are for single-valued string fields, but
multivalued string fields are also included in the benchmark attachment
(results for multi-valued didn't differ substantially from those for
single-valued).
There's a row for each docset domain recall percentage (percentage of \*:*
domain returned by main query/fg), and a column for each field cardinality;
cell values indicate latency (QTime) in ms against a single core with 3 million
docs, no deletes; each value is the average of 10 repeated invocations of the
the relevant request (standard deviation isn't captured here, but was quite
low, fwiw).
{code:java}
MASTER:
cdnlty: 10 100 1k 10k 100k 1m
.1% 23 32 75 124 118 109
1% 28 43 98 412 1057 582
10% 78 90 135 430 3291 5530
20% 139 151 192 484 3353 7610
30% 198 207 250 538 3462 9537
40% 253 265 307 595 3474 11206
50% 321 341 374 655 3497 13155
{code}
{code:java}
SOLR-13132 sweep_collection=false, filterCacheSize=0
cdnlty: 10 100 1k 10k 100k 1m
.1% 24 37 74 119 108 74
1% 29 44 97 403 1021 563
10% 77 92 136 417 3161 5356
20% 144 156 197 486 3233 7257
30% 199 209 254 534 3322 9224
40% 254 276 314 599 3393 10937
50% 323 352 368 643 3403 12718
{code}
{code:java}
SOLR-13132 sweep_collection=true
cdnlty: 10 100 1k 10k 100k 1m
.1% 99 99 106 111 114 145
1% 102 106 112 110 122 157
10% 168 173 177 175 193 241
20% 241 245 249 249 266 341
30% 307 312 318 313 337 409
40% 382 386 390 390 414 494
50% 449 455 459 460 487 569
{code}
{code:java}
SOLR-13132 sweep_collection=false, filterCacheSize=12000 (very large!)
.1% 21 17 33 45 37 57
1% 7 17 41 270 885 525
10% 35 44 57 375 3239 5866
20% 71 78 89 204 3324 7800
30% 97 108 120 220 3335 9684
40% 130 141 152 248 3258 11582
50% 159 171 183 270 3299 13754
{code}
The last of these is with a _very_ large filterCache configured. It pretty
clearly benefits cardinality through 10k, but has a slight negative impact on
100k and above (filterCache overhead with no benefit). This is as expected;
it's worth noting that in real-world situations, filterCache is likely to be
smaller and also used by other features, so this test probably underestimates
the negative system-wide impact of an undersized filterCache, and also
underestimates the sufficient filterCache size threshold wrt field cardinality.
Because there were low-level changes to the code to collect counts, I also did
a sanity check comparing simple facet term count performance (no skg) of master
wrt SOLR-13132. I won't post the results in line here, but there appeared to be
no change whatsoever. The results are included in the attached tar file (under
the "results" directory).
Side note: the negative performance impact of sweep collection for
low-cardinality docsets is mainly due to the fact that the full count
accumulation domain (for sweep collection) becomes the union of the result
DocSet, fg DocSet, and bg DocSet. Where bg DocSet is often large (e.g., \*:*),
we're essentially just seeing the effect of collecting term facet counts over a
high cardinality DocSet domain. As a contrived example, setting an artificially
restricted bgSet (e.g., {{\{!prefix f=id v=999999}}}) is considerably faster:
{code:java}
SOLR-13132 sweep_collection=true, tiny bgSet
cdnlty: 10 100 1k 10k 100k 1m
.1% 2 1 1 2 2 8
1% 4 6 6 6 9 17
10% 50 48 46 47 55 78
20% 91 92 93 93 104 146
30% 135 137 138 141 152 198
40% 180 182 185 186 198 253
50% 223 226 228 229 245 316
{code}
> Improve JSON "terms" facet performance when sorted by relatedness
> ------------------------------------------------------------------
>
> Key: SOLR-13132
> URL: https://issues.apache.org/jira/browse/SOLR-13132
> Project: Solr
> Issue Type: Improvement
> Components: Facet Module
> Affects Versions: 7.4, master (9.0)
> Reporter: Michael Gibney
> Priority: Major
> Attachments: SOLR-13132-benchmarks.tgz,
> SOLR-13132-with-cache-01.patch, SOLR-13132-with-cache.patch,
> SOLR-13132.patch, SOLR-13132_testSweep.patch
>
> Time Spent: 1.5h
> Remaining Estimate: 0h
>
> When sorting buckets by {{relatedness}}, JSON "terms" facet must calculate
> {{relatedness}} for every term.
> The current implementation uses a standard uninverted approach (either
> {{docValues}} or {{UnInvertedField}}) to get facet counts over the domain
> base docSet, and then uses that initial pass as a pre-filter for a
> second-pass, inverted approach of fetching docSets for each relevant term
> (i.e., {{count > minCount}}?) and calculating intersection size of those sets
> with the domain base docSet.
> Over high-cardinality fields, the overhead of per-term docSet creation and
> set intersection operations increases request latency to the point where
> relatedness sort may not be usable in practice (for my use case, even after
> applying the patch for SOLR-13108, for a field with ~220k unique terms per
> core, QTime for high-cardinality domain docSets were, e.g.: cardinality
> 1816684=9000ms, cardinality 5032902=18000ms).
> The attached patch brings the above example QTimes down to a manageable
> ~300ms and ~250ms respectively. The approach calculates uninverted facet
> counts over domain base, foreground, and background docSets in parallel in a
> single pass. This allows us to take advantage of the efficiencies built into
> the standard uninverted {{FacetFieldProcessorByArray[DV|UIF]}}), and avoids
> the per-term docSet creation and set intersection overhead.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org
For additional commands, e-mail: issues-h...@lucene.apache.org
