implement FilterFn for ChunkedArray

[a10y]

No description provided.

[a10y]

Heh, turns out the naive implementation is 3x slower than develop, pretty much entirely due to find_chunk_idx slowness

[a10y]

Alright so currently what I'm seeing is for Q15 vortex-compressed-file:

• develop: ~112ms
• This PR, using ChunkedArray::find_chunk_idx: ~700ms, most of that in find_chunk_idx
• The latest commit that uses a binary search on pre-materialized chunk indices: ~106ms

The new flame graph is showing ~6% less time spent in filter, which is what we'd expect roughly just going off the numbers

[robert3005]

Right, this can go lower if all the children also implement filter. Good to know that we need to unpack values

[a10y]

Most of the children do implement FilterFn, the exception is extension array

EDIT: about half do. The missing ones here are FoR and Extension

[robert3005]

FoR has it. Notable omission is Bitpacked which could repack itself and it likely would be faster. Also sparse is missing which likely leads to some extra allocation

[robert3005]

fwiw we should make them in separate prs

[github-actions]

Vortex benchmarks

Benchmark suite	Current: d5ecb80	Previous: 750b9c8	Ratio
tpch_q1/vortex-in-memory-no-pushdown	467629937 ns/iter (± 3750524)	482302101 ns/iter (± 5129904)	0.97
tpch_q1/vortex-in-memory-pushdown	515046187 ns/iter (± 1745901)	527534072 ns/iter (± 3463267)	0.98
tpch_q1/arrow	454002116 ns/iter (± 2291129)	451801735 ns/iter (± 3758415)	1.00
tpch_q1/parquet	665259754 ns/iter (± 3313569)	661718171 ns/iter (± 2674937)	1.01
tpch_q1/vortex-file-compressed	610874701 ns/iter (± 2392209)	621200118 ns/iter (± 6749288)	0.98
tpch_q1/vortex-file-uncompressed	633865011 ns/iter (± 2741863)	592370294 ns/iter (± 8018622)	1.07
tpch_q2/vortex-in-memory-no-pushdown	144574917 ns/iter (± 990792)	160802521 ns/iter (± 4678476)	0.90
tpch_q2/vortex-in-memory-pushdown	145267883 ns/iter (± 876640)	159124621 ns/iter (± 3511202)	0.91
tpch_q2/arrow	122184510 ns/iter (± 572539)	127992668 ns/iter (± 614847)	0.95
tpch_q2/parquet	158407129 ns/iter (± 845449)	167523912 ns/iter (± 976972)	0.95
tpch_q2/vortex-file-compressed	154158990 ns/iter (± 731391)	170749370 ns/iter (± 1095064)	0.90
tpch_q2/vortex-file-uncompressed	164798709 ns/iter (± 1537024)	176352506 ns/iter (± 1861263)	0.93
tpch_q3/vortex-in-memory-no-pushdown	154082645 ns/iter (± 945209)	164894661 ns/iter (± 1688574)	0.93
tpch_q3/vortex-in-memory-pushdown	182599172 ns/iter (± 2140747)	204285819 ns/iter (± 1733504)	0.89
tpch_q3/arrow	148805278 ns/iter (± 1268268)	163147591 ns/iter (± 2200349)	0.91
tpch_q3/parquet	339122776 ns/iter (± 2983826)	364056632 ns/iter (± 3161647)	0.93
tpch_q3/vortex-file-compressed	600109967 ns/iter (± 2475613)	238624885 ns/iter (± 3477685)	2.51
tpch_q3/vortex-file-uncompressed	346678665 ns/iter (± 4154641)	264574306 ns/iter (± 4886537)	1.31
tpch_q4/vortex-in-memory-no-pushdown	116713076 ns/iter (± 680431)	119477803 ns/iter (± 1823670)	0.98
tpch_q4/vortex-in-memory-pushdown	140401608 ns/iter (± 1067465)	148794718 ns/iter (± 4839612)	0.94
tpch_q4/arrow	112121350 ns/iter (± 969314)	126580600 ns/iter (± 3893440)	0.89
tpch_q4/parquet	223556527 ns/iter (± 2134392)	233726536 ns/iter (± 2398398)	0.96
tpch_q4/vortex-file-compressed	608122542 ns/iter (± 1411424)	214291423 ns/iter (± 4481945)	2.84
tpch_q4/vortex-file-uncompressed	301939005 ns/iter (± 3284469)	223841277 ns/iter (± 4788043)	1.35
tpch_q5/vortex-in-memory-no-pushdown	300395241 ns/iter (± 2339398)	331680262 ns/iter (± 3094149)	0.91
tpch_q5/vortex-in-memory-pushdown	314484406 ns/iter (± 2458080)	352998734 ns/iter (± 3928054)	0.89
tpch_q5/arrow	296284099 ns/iter (± 4279535)	322701287 ns/iter (± 4014088)	0.92
tpch_q5/parquet	456248488 ns/iter (± 3726847)	491143044 ns/iter (± 5029280)	0.93
tpch_q5/vortex-file-compressed	346751667 ns/iter (± 4316017)	373255236 ns/iter (± 8149263)	0.93
tpch_q5/vortex-file-uncompressed	354653588 ns/iter (± 6344729)	370011144 ns/iter (± 7597998)	0.96
tpch_q6/vortex-in-memory-no-pushdown	41564617 ns/iter (± 702737)	45597608 ns/iter (± 395126)	0.91
tpch_q6/vortex-in-memory-pushdown	91211372 ns/iter (± 453285)	94817690 ns/iter (± 1631882)	0.96
tpch_q6/arrow	37046230 ns/iter (± 687907)	39935601 ns/iter (± 153591)	0.93
tpch_q6/parquet	154464040 ns/iter (± 1078498)	158515552 ns/iter (± 1076605)	0.97
tpch_q6/vortex-file-compressed	77106628 ns/iter (± 349019)	82131752 ns/iter (± 447633)	0.94
tpch_q6/vortex-file-uncompressed	166363477 ns/iter (± 1183907)	164012224 ns/iter (± 1392620)	1.01
tpch_q7/vortex-in-memory-no-pushdown	570770557 ns/iter (± 5628418)	628570919 ns/iter (± 13993824)	0.91
tpch_q7/vortex-in-memory-pushdown	618735527 ns/iter (± 6725917)	685014994 ns/iter (± 4076828)	0.90
tpch_q7/arrow	555973167 ns/iter (± 6869244)	591575674 ns/iter (± 9570461)	0.94
tpch_q7/parquet	733884143 ns/iter (± 7867692)	766331481 ns/iter (± 7275189)	0.96
tpch_q7/vortex-file-compressed	856993945 ns/iter (± 4929339)	664591921 ns/iter (± 7273788)	1.29
tpch_q7/vortex-file-uncompressed	740677757 ns/iter (± 5450486)	736981022 ns/iter (± 11454205)	1.01
tpch_q8/vortex-in-memory-no-pushdown	210734148 ns/iter (± 1364876)	216322787 ns/iter (± 1450325)	0.97
tpch_q8/vortex-in-memory-pushdown	222583044 ns/iter (± 896335)	225993479 ns/iter (± 1636984)	0.98
tpch_q8/arrow	210921272 ns/iter (± 691187)	213002888 ns/iter (± 798766)	0.99
tpch_q8/parquet	481543044 ns/iter (± 2300449)	480810962 ns/iter (± 5440907)	1.00
tpch_q8/vortex-file-compressed	269483461 ns/iter (± 4783378)	250163486 ns/iter (± 533009)	1.08
tpch_q8/vortex-file-uncompressed	292301158 ns/iter (± 9043288)	289075003 ns/iter (± 2810625)	1.01
tpch_q9/vortex-in-memory-no-pushdown	405813821 ns/iter (± 2341096)	411252124 ns/iter (± 5191902)	0.99
tpch_q9/vortex-in-memory-pushdown	406254784 ns/iter (± 2090194)	402404864 ns/iter (± 10300688)	1.01
tpch_q9/arrow	401193441 ns/iter (± 3486194)	407465021 ns/iter (± 4936156)	0.98
tpch_q9/parquet	702975198 ns/iter (± 3508374)	690287931 ns/iter (± 2228853)	1.02
tpch_q9/vortex-file-compressed	454991563 ns/iter (± 4574292)	448813755 ns/iter (± 3147241)	1.01
tpch_q9/vortex-file-uncompressed	487958941 ns/iter (± 3139067)	493338488 ns/iter (± 3770312)	0.99
tpch_q10/vortex-in-memory-no-pushdown	240534797 ns/iter (± 1106459)	237008689 ns/iter (± 999451)	1.01
tpch_q10/vortex-in-memory-pushdown	263164189 ns/iter (± 1948364)	269284156 ns/iter (± 2766466)	0.98
tpch_q10/arrow	232776848 ns/iter (± 1296696)	232630226 ns/iter (± 1125035)	1.00
tpch_q10/parquet	491261591 ns/iter (± 3634032)	491838979 ns/iter (± 2172643)	1.00
tpch_q10/vortex-file-compressed	599147762 ns/iter (± 2913480)	438008551 ns/iter (± 1560277)	1.37
tpch_q10/vortex-file-uncompressed	397099274 ns/iter (± 3897509)	342085663 ns/iter (± 1049070)	1.16
tpch_q11/vortex-in-memory-no-pushdown	220454180 ns/iter (± 2158905)	228213202 ns/iter (± 1929879)	0.97
tpch_q11/vortex-in-memory-pushdown	218691415 ns/iter (± 2299972)	227285629 ns/iter (± 1260022)	0.96
tpch_q11/arrow	176616687 ns/iter (± 1047656)	181955973 ns/iter (± 481860)	0.97
tpch_q11/parquet	190341853 ns/iter (± 1991299)	190665473 ns/iter (± 1014961)	1.00
tpch_q11/vortex-file-compressed	232418289 ns/iter (± 1398293)	234354109 ns/iter (± 2669234)	0.99
tpch_q11/vortex-file-uncompressed	238199675 ns/iter (± 2755331)	236021473 ns/iter (± 2076840)	1.01
tpch_q12/vortex-in-memory-no-pushdown	178888983 ns/iter (± 564729)	181725183 ns/iter (± 665937)	0.98
tpch_q12/vortex-in-memory-pushdown	257220480 ns/iter (± 1246588)	245639824 ns/iter (± 1117857)	1.05
tpch_q12/arrow	168685105 ns/iter (± 435972)	167848416 ns/iter (± 951860)	1.00
tpch_q12/parquet	358077917 ns/iter (± 2317907)	355743248 ns/iter (± 652283)	1.01
tpch_q12/vortex-file-compressed	577060066 ns/iter (± 2253366)	583693695 ns/iter (± 3524490)	0.99
tpch_q12/vortex-file-uncompressed	353753839 ns/iter (± 1158424)	345717904 ns/iter (± 982840)	1.02
tpch_q13/vortex-in-memory-no-pushdown	233378060 ns/iter (± 6792292)	212026022 ns/iter (± 2139385)	1.10
tpch_q13/vortex-in-memory-pushdown	225897062 ns/iter (± 2687703)	215199549 ns/iter (± 2801269)	1.05
tpch_q13/arrow	220008054 ns/iter (± 3476872)	216288457 ns/iter (± 7582246)	1.02
tpch_q13/parquet	365108165 ns/iter (± 8761788)	339023424 ns/iter (± 4982544)	1.08
tpch_q13/vortex-file-compressed	262370823 ns/iter (± 5771851)	242532188 ns/iter (± 3105220)	1.08
tpch_q13/vortex-file-uncompressed	253289040 ns/iter (± 5878713)	230206578 ns/iter (± 4731127)	1.10
tpch_q14/vortex-in-memory-no-pushdown	39976992 ns/iter (± 402331)	40066809 ns/iter (± 123691)	1.00
tpch_q14/vortex-in-memory-pushdown	87308864 ns/iter (± 536472)	88371331 ns/iter (± 993349)	0.99
tpch_q14/arrow	41047857 ns/iter (± 171059)	39916154 ns/iter (± 413480)	1.03
tpch_q14/parquet	222397328 ns/iter (± 984220)	219644722 ns/iter (± 1478739)	1.01
tpch_q14/vortex-file-compressed	90954221 ns/iter (± 1205267)	84406219 ns/iter (± 549154)	1.08
tpch_q14/vortex-file-uncompressed	145316171 ns/iter (± 3272870)	132056672 ns/iter (± 2526723)	1.10
tpch_q15/vortex-in-memory-no-pushdown	71163237 ns/iter (± 568725)	70240367 ns/iter (± 1211358)	1.01
tpch_q15/vortex-in-memory-pushdown	125755999 ns/iter (± 1421365)	120877627 ns/iter (± 662063)	1.04
tpch_q15/arrow	67457581 ns/iter (± 563588)	65814738 ns/iter (± 745743)	1.02
tpch_q15/parquet	295049250 ns/iter (± 2104210)	292964357 ns/iter (± 4176947)	1.01
tpch_q15/vortex-file-compressed	160178907 ns/iter (± 1035165)	149166262 ns/iter (± 3739034)	1.07
tpch_q15/vortex-file-uncompressed	280301533 ns/iter (± 3911470)	245576987 ns/iter (± 4228385)	1.14
tpch_q16/vortex-in-memory-no-pushdown	120676960 ns/iter (± 607001)	121362328 ns/iter (± 501981)	0.99
tpch_q16/vortex-in-memory-pushdown	125649150 ns/iter (± 652981)	125642905 ns/iter (± 739682)	1.00
tpch_q16/arrow	105605122 ns/iter (± 1175945)	106264398 ns/iter (± 1435624)	0.99
tpch_q16/parquet	122493799 ns/iter (± 590708)	120515396 ns/iter (± 330045)	1.02
tpch_q16/vortex-file-compressed	137353235 ns/iter (± 669637)	135440640 ns/iter (± 2489773)	1.01
tpch_q16/vortex-file-uncompressed	136512392 ns/iter (± 797010)	133236531 ns/iter (± 804638)	1.02
tpch_q17/vortex-in-memory-no-pushdown	680071206 ns/iter (± 5626853)	633750402 ns/iter (± 21999972)	1.07
tpch_q17/vortex-in-memory-pushdown	699765651 ns/iter (± 16552381)	644886261 ns/iter (± 28816494)	1.09
tpch_q17/arrow	614163957 ns/iter (± 15290848)	604794412 ns/iter (± 19325686)	1.02
tpch_q17/parquet	600504141 ns/iter (± 3375464)	591494702 ns/iter (± 5886351)	1.02
tpch_q17/vortex-file-compressed	616490641 ns/iter (± 6109550)	618751453 ns/iter (± 5094786)	1.00
tpch_q17/vortex-file-uncompressed	678439856 ns/iter (± 5153733)	662045625 ns/iter (± 5462241)	1.02
tpch_q18/vortex-in-memory-no-pushdown	1102704467 ns/iter (± 10488831)	1107670912 ns/iter (± 28124407)	1.00
tpch_q18/vortex-in-memory-pushdown	1102693920 ns/iter (± 18019955)	1146350113 ns/iter (± 25935429)	0.96
tpch_q18/arrow	1091548880 ns/iter (± 11287668)	1088592199 ns/iter (± 13463439)	1.00
tpch_q18/parquet	1269701350 ns/iter (± 9244969)	1295921800 ns/iter (± 29563581)	0.98
tpch_q18/vortex-file-compressed	1137231576 ns/iter (± 14509774)	1104461804 ns/iter (± 12804403)	1.03
tpch_q18/vortex-file-uncompressed	1199433179 ns/iter (± 13902541)	1142531705 ns/iter (± 9323161)	1.05
tpch_q19/vortex-in-memory-no-pushdown	172419237 ns/iter (± 558277)	173585771 ns/iter (± 474023)	0.99
tpch_q19/vortex-in-memory-pushdown	255865266 ns/iter (± 777800)	263934184 ns/iter (± 1443569)	0.97
tpch_q19/arrow	158947649 ns/iter (± 563373)	158169603 ns/iter (± 605941)	1.00
tpch_q19/parquet	482869625 ns/iter (± 2182275)	485808293 ns/iter (± 3098680)	0.99
tpch_q19/vortex-file-compressed	739251560 ns/iter (± 3891957)	748242129 ns/iter (± 1861891)	0.99
tpch_q19/vortex-file-uncompressed	361335402 ns/iter (± 1566759)	370691529 ns/iter (± 2848285)	0.97
tpch_q20/vortex-in-memory-no-pushdown	276953400 ns/iter (± 4328394)	271361202 ns/iter (± 2598718)	1.02
tpch_q20/vortex-in-memory-pushdown	296899338 ns/iter (± 2968668)	298142045 ns/iter (± 3138390)	1.00
tpch_q20/arrow	253871397 ns/iter (± 4922124)	268988087 ns/iter (± 5230411)	0.94
tpch_q20/parquet	368063779 ns/iter (± 5920070)	376959670 ns/iter (± 2535342)	0.98
tpch_q20/vortex-file-compressed	337625170 ns/iter (± 3942064)	331887989 ns/iter (± 5057692)	1.02
tpch_q20/vortex-file-uncompressed	419819929 ns/iter (± 2448535)	387430930 ns/iter (± 3744062)	1.08
tpch_q21/vortex-in-memory-no-pushdown	881470571 ns/iter (± 5570260)	860034433 ns/iter (± 6315368)	1.02
tpch_q21/vortex-in-memory-pushdown	912642083 ns/iter (± 12158266)	894832538 ns/iter (± 2774021)	1.02
tpch_q21/arrow	873175820 ns/iter (± 4749117)	860412313 ns/iter (± 4835728)	1.01
tpch_q21/parquet	1033653879 ns/iter (± 10326375)	1031986215 ns/iter (± 15406311)	1.00
tpch_q21/vortex-file-compressed	1930001242 ns/iter (± 9281258)	1055005493 ns/iter (± 8739734)	1.83
tpch_q21/vortex-file-uncompressed	1316949047 ns/iter (± 3918408)	1091486875 ns/iter (± 12444901)	1.21
tpch_q22/vortex-in-memory-no-pushdown	96238393 ns/iter (± 774157)	96963365 ns/iter (± 1134408)	0.99
tpch_q22/vortex-in-memory-pushdown	96618668 ns/iter (± 1183404)	97059184 ns/iter (± 835576)	1.00
tpch_q22/arrow	66826555 ns/iter (± 2434487)	66658773 ns/iter (± 251660)	1.00
tpch_q22/parquet	96074157 ns/iter (± 5283274)	95223939 ns/iter (± 748371)	1.01
tpch_q22/vortex-file-compressed	103282488 ns/iter (± 460786)	101715042 ns/iter (± 409685)	1.02
tpch_q22/vortex-file-uncompressed	110932476 ns/iter (± 603983)	109602568 ns/iter (± 584164)	1.01

This comment was automatically generated by workflow using github-action-benchmark.

[robert3005]

@a10y you have a bug somewhere

Failed to get min_index from SparseArray: index 0 out of bounds from 0 to 0
Backtrace:
disabled backtrace

[a10y]

perhaps we should just return SparseArray directly from BPA::patches. I know we always return Array, but turning it into a SparseArray is pretty much the first thing we always do with it

[robert3005]

eh, I think we should handle non sparse patches. I will work on that

[a10y]

We depend pretty deeply on this being a SparseArray so I figured it should be checked in the constructor, along with the non-empty indices when present thing

[robert3005]

I think this is fine for now, we will want to change it eventually but then we can find all those places

[a10y]

Alright benchmark comment is updated. Most queries stayed roughly same, the following got slower, some by a considerable margin:

• Q1
• Q3
• Q4
• Q7
• Q10
• Q15
• Q16
• Q21
• Q22

Will do some profiling investigation in the AM

[a10y]

Alright, the issue was with filter_slices, it was creating so many slices that the overhead of slicing was just eating everything else up, especially when it came time to re-canonicalize before passing the RecordBatches back to DF.

Now, filter_slices will instead look at all of the slices and determine which chunks must be sliced, passed through, or dropped completely. I'll kick off a full bench run but for Q16 this saves us ~20% VS what we had on develop.

[a10y]

this was a leftover from a time before dtype was able to be passed into this method, we had needed to examine the first chunk to get the DType but that's no longer the case

[a10y]

Ok benchmark run just completed, looks like the 20ms speed up i was seeing locally for Q16 was maybe just test jitter, in the PR run it's roughly same as develop. Most queries are 0-10% faster. A number are considerably slower, this time due to slow take implementation. May try and implement the accelerated search_sorted here as well

[robert3005]

@a10y we should just merge this and make follow up prs. There's a lot of noise in the runs so comparing to latest develop runs is a bit difficult. Anyway I will review this and we can make another pr?

[a10y]

Alright I'm ok with that if you are. We have the issues for tracking already. I just hate to see graphs go up :P

[robert3005]

our search_sorted method is defined on primitive arrays and last I checked it wasn't any different (could be different now since 1.81 changed binary_search)

[a10y]

It's that the chunk_ends are usually compressed, so decompressing them for every call to find_chunk_idx adds up, along with the scalar/dtype overhead.

[robert3005]

I meant that you can do this https://github.com/spiraldb/vortex/blob/develop/vortex-array/src/compute/search_sorted.rs#L276-L279

[robert3005]

i.e. you can do chunk_ends.search_sorted(&idx, SearchSortedSide::Right)

[robert3005]

https://github.com/spiraldb/vortex/blob/develop/vortex-array/src/array/chunked/mod.rs#L98-L101 is the right thing to do

[a10y]

Ah yea, that is probably better than what I am doing. Lemme change

[robert3005]

one thing that the clippy suggestion does is avoid bounds checking but it would only show up on large arrays

[a10y]

I forgot that you can do &mut vector[start...end], which I think is both clear and avoids bounds check

[robert3005]

Since I spent some much time looking at weird arrays I know this has a weird behaviour with empty chunks but we can handle it in a followup. Needs adjustment to find_chunk_idx like I did in a previous pr

[robert3005]

I think I meant something else, you can avoid the scalar/array overhead but still use the search sorted logic, i.e. keep find_chunk_idx in chunked array but you can create a mirror of that function over &[u64]. In the current variant you end up with scalar overhead.

The code I linked was to show that our searching logic is implemented for primitive array slices as well as arrays and you need it to handle empty chunks.

[a10y]

like that?