Optimize exact composite primary key seeks

[timsehn]

Summary

• identify the highest multiplier from PR Avoid duplicate seek for single-row inserts #1036 as composite PK file-backed autocommit writes: oltp_update_index_ac and oltp_delete_insert_ac tied at 4.73x
• skip the leaf scan for exact WITHOUT ROWID table primary-key seeks when the encoded sort key already matched the tree
• keep the pending-delete check before returning the tree hit

Local benchmark notes

Focused 100K-row composite autocommit harness, DoltLite-only 21 runs on local Mac:

• baseline oltp_update_index_ac median 41,706 us, trimmed median 39,928 us
• branch oltp_update_index_ac median 43,064 us, trimmed median 41,423 us
• baseline oltp_delete_insert_ac median 58,642 us, trimmed median 55,993 us
• branch oltp_delete_insert_ac median 37,533 us, trimmed median 36,729 us

Local runs had large scheduler outliers; delete/insert showed the clear win among the tied worst metrics.

Tests

• make -j8 doltlite doltlite-lib
• BENCH_ROWS=200 BENCH_RUNS=1 BENCH_MAX_MULTIPLIER=1000 BENCH_AVG_MAX_MULTIPLIER=1000 ./test/sysbench_compare_compositepk.sh
• make -j8 invariant_test corruption_test c-tests && ./invariant_test && ./corruption_test && test/run_c_tests.sh .

[github-actions]

Sysbench-Style Benchmark: Doltlite vs SQLite

In-Memory

Reads

Test	SQLite (us)	Doltlite (us)	Multiplier
oltp_point_select	20,691	30,510	1.47
oltp_range_select	8,841	12,785	1.45
oltp_sum_range	8,463	12,388	1.46
oltp_order_range	2,325	2,930	1.26
oltp_distinct_range	3,105	3,776	1.22
oltp_index_scan	3,386	5,560	1.64
select_random_points	8,883	15,969	1.80
select_random_ranges	2,614	4,468	1.71
covering_index_scan	3,536	3,725	1.05
groupby_scan	26,987	31,363	1.16
index_join	4,936	8,270	1.68
index_join_scan	2,626	5,101	1.94
types_table_scan	932,801	1,249,759	1.34
table_scan	1,062,115	1,407,625	1.33
oltp_read_only	87,385	116,514	1.33
Average			1.46

Writes

Test	SQLite (us)	Doltlite (us)	Multiplier
oltp_bulk_insert	144,738	199,783	1.38
oltp_insert	12,851	22,469	1.75
oltp_update_index	43,525	90,280	2.07
oltp_update_non_index	28,069	56,891	2.03
oltp_delete_insert	39,138	67,231	1.72
oltp_write_only	18,599	41,316	2.22
types_delete_insert	20,454	38,008	1.86
oltp_read_write	55,627	104,493	1.88
Average			1.86

File-Backed

Reads

Test	SQLite (us)	Doltlite (us)	Multiplier
oltp_point_select	54,999	69,484	1.26
oltp_range_select	13,375	25,720	1.92
oltp_sum_range	12,948	25,512	1.97
oltp_order_range	2,826	4,775	1.69
oltp_distinct_range	3,752	5,700	1.52
oltp_index_scan	7,416	12,430	1.68
select_random_points	17,348	44,354	2.56
select_random_ranges	6,217	10,431	1.68
covering_index_scan	6,570	5,582	0.85
groupby_scan	28,311	41,868	1.48
index_join	7,235	13,299	1.84
index_join_scan	3,385	9,456	2.79
types_table_scan	1,041,107	2,090,184	2.01
table_scan	1,193,227	2,592,415	2.17
oltp_read_only	142,499	206,428	1.45
Average			1.79

Writes

Test	SQLite (us)	Doltlite (us)	Multiplier
oltp_bulk_insert	152,287	213,285	1.40
oltp_insert	20,573	29,196	1.42
oltp_update_index	88,179	166,259	1.89
oltp_update_non_index	57,832	106,998	1.85
oltp_delete_insert	69,042	116,115	1.68
oltp_write_only	44,341	76,127	1.72
types_delete_insert	36,062	63,303	1.76
oltp_read_write	90,044	192,590	2.14
Average			1.73

File-Backed (autocommit)

Each statement runs as its own transaction — exposes per-commit
fixed costs that the wrapped-in-BEGIN/COMMIT tests amortize away.
SQLite uses WAL mode with synchronous=FULL in this section so
the comparison uses SQLite's durable WAL autocommit path.

Reads

Reads have no commit cost; these are the same SQL files as the
File-Backed Reads section, included here for symmetry and to
catch any per-statement overhead doltlite pays on the read path.

Test	SQLite (us)	Doltlite (us)	Multiplier
oltp_point_select	34,414	69,428	2.02
oltp_range_select	11,150	25,827	2.32
oltp_sum_range	10,674	25,613	2.40
oltp_order_range	2,751	4,884	1.78
oltp_distinct_range	3,452	5,659	1.64
oltp_index_scan	5,276	12,069	2.29
select_random_points	15,104	44,558	2.95
select_random_ranges	4,111	10,322	2.51
covering_index_scan	4,634	5,640	1.22
groupby_scan	28,068	41,570	1.48
index_join	6,078	13,427	2.21
index_join_scan	3,097	9,392	3.03
types_table_scan	1,043,069	2,096,970	2.01
table_scan	1,199,759	2,595,291	2.16
oltp_read_only	113,138	204,082	1.80
Average			2.12

Writes

Test	SQLite (us)	Doltlite (us)	Multiplier
oltp_bulk_insert_ac	15,986	55,429	3.47
oltp_insert_ac	16,412	71,354	4.35
oltp_update_index_ac	17,502	78,897	4.51
oltp_update_non_index_ac	16,305	62,398	3.83
oltp_delete_insert_ac	17,800	73,776	4.14
oltp_write_only_ac	17,205	70,131	4.08
types_delete_insert_ac	15,750	67,418	4.28
oltp_read_write_ac	20,411	82,754	4.05
Average			4.09

100000 rows, median of 5 invocations per test, workload-only timing via host monotonic clock when available.

Performance Ceiling Check (6x individual, 5x average)

All tests within ceilings.

[github-actions]

Sysbench-Style Benchmark (composite PK): Doltlite vs SQLite

Companion to the classic Sysbench-Style Benchmark. Every workload here
runs against tables with a 2-column INTEGER PRIMARY KEY(a, b) WITHOUT ROWID.
Individual ratios gated at 6×; section averages gated at 5×.

In-Memory

Reads

Test	SQLite (us)	Doltlite (us)	Multiplier
oltp_point_select	32,800	44,579	1.36
oltp_range_select	20,402	25,335	1.24
oltp_sum_range	18,314	24,341	1.33
oltp_order_range	3,753	4,382	1.17
oltp_distinct_range	4,864	5,499	1.13
oltp_index_scan	4,663	7,149	1.53
select_random_points	27,304	36,811	1.35
select_random_ranges	7,545	9,163	1.21
covering_index_scan	4,288	4,774	1.11
groupby_scan	39,844	45,535	1.14
index_join	7,919	11,970	1.51
index_join_scan	4,092	6,838	1.67
types_table_scan	1,119,319	1,523,487	1.36
table_scan	1,289,081	1,737,048	1.35
oltp_read_only	152,002	187,413	1.23
Average			1.31

Writes

Test	SQLite (us)	Doltlite (us)	Multiplier
oltp_bulk_insert	244,153	311,334	1.28
oltp_insert	19,152	33,278	1.74
oltp_update_index	66,910	124,144	1.86
oltp_update_non_index	51,499	85,220	1.65
oltp_delete_insert	49,834	98,616	1.98
oltp_write_only	27,159	57,921	2.13
types_delete_insert	32,577	58,030	1.78
oltp_read_write	100,051	165,534	1.65
Average			1.76

File-Backed

Reads

Test	SQLite (us)	Doltlite (us)	Multiplier
oltp_point_select	125,768	127,376	1.01
oltp_range_select	30,918	54,172	1.75
oltp_sum_range	29,573	52,459	1.77
oltp_order_range	4,843	8,164	1.69
oltp_distinct_range	5,829	9,200	1.58
oltp_index_scan	14,352	21,285	1.48
select_random_points	44,611	95,124	2.13
select_random_ranges	16,763	21,252	1.27
covering_index_scan	12,518	10,971	0.88
groupby_scan	43,031	66,243	1.54
index_join	13,414	27,712	2.07
index_join_scan	5,396	17,812	3.30
types_table_scan	1,329,734	3,417,485	2.57
table_scan	1,559,274	4,179,249	2.68
oltp_read_only	292,869	374,663	1.28
Average			1.80

Writes

Test	SQLite (us)	Doltlite (us)	Multiplier
oltp_bulk_insert	252,864	337,498	1.33
oltp_insert	25,562	49,654	1.94
oltp_update_index	157,397	278,135	1.77
oltp_update_non_index	104,136	181,466	1.74
oltp_delete_insert	104,077	196,140	1.88
oltp_write_only	63,380	128,434	2.03
types_delete_insert	60,712	105,974	1.75
oltp_read_write	163,142	340,851	2.09
Average			1.82

File-Backed (autocommit)

Each statement runs as its own transaction — exposes per-commit
fixed costs that the wrapped-in-BEGIN/COMMIT tests amortize away.
SQLite uses WAL mode with synchronous=FULL in this section so
the comparison uses SQLite's durable WAL autocommit path.

Reads

Reads have no commit cost; these are the same SQL files as the
File-Backed Reads section, included here for symmetry and to
catch any per-statement overhead doltlite pays on the read path.

Test	SQLite (us)	Doltlite (us)	Multiplier
oltp_point_select	69,625	125,202	1.80
oltp_range_select	25,445	53,576	2.11
oltp_sum_range	24,286	51,927	2.14
oltp_order_range	4,280	8,107	1.89
oltp_distinct_range	5,320	9,136	1.72
oltp_index_scan	8,750	20,886	2.39
select_random_points	38,544	94,456	2.45
select_random_ranges	11,090	21,030	1.90
covering_index_scan	7,127	10,970	1.54
groupby_scan	42,292	66,155	1.56
index_join	10,740	27,177	2.53
index_join_scan	4,839	17,895	3.70
types_table_scan	1,307,790	3,366,683	2.57
table_scan	1,543,046	4,184,490	2.71
oltp_read_only	214,183	374,734	1.75
Average			2.18

Writes

Test	SQLite (us)	Doltlite (us)	Multiplier
oltp_bulk_insert_ac	17,569	57,630	3.28
oltp_insert_ac	18,345	79,002	4.31
oltp_update_index_ac	21,589	90,300	4.18
oltp_update_non_index_ac	16,248	71,232	4.38
oltp_delete_insert_ac	19,499	85,501	4.38
oltp_write_only_ac	20,561	87,106	4.24
types_delete_insert_ac	17,136	79,026	4.61
oltp_read_write_ac	27,954	101,282	3.62
Average			4.13

100000 rows, median of 5 invocations per test, workload-only timing via host monotonic clock when available.

Performance Ceiling Check (6x individual, 5x average)

All tests within ceilings.

[github-actions]

Sysbench-Style Benchmark (BLOB PK): Doltlite vs SQLite

Companion to the classic Sysbench-Style Benchmark. Every workload here
runs against tables with a 16-byte big-endian BLOB PRIMARY KEY.
Individual ratios gated at 6×; section averages gated at 5×.

In-Memory

Reads

Test	SQLite (us)	Doltlite (us)	Multiplier
oltp_point_select	30,193	41,612	1.38
oltp_range_select	13,536	18,827	1.39
oltp_sum_range	11,932	18,346	1.54
oltp_order_range	3,128	3,798	1.21
oltp_distinct_range	4,214	4,791	1.14
oltp_index_scan	4,484	7,178	1.60
select_random_points	17,409	25,857	1.49
select_random_ranges	4,190	6,341	1.51
covering_index_scan	4,509	5,157	1.14
groupby_scan	34,817	41,240	1.18
index_join	6,786	11,993	1.77
index_join_scan	4,217	7,611	1.80
types_table_scan	1,122,917	1,612,808	1.44
table_scan	1,303,373	1,821,802	1.40
oltp_read_only	120,633	158,380	1.31
Average			1.42

Writes

Test	SQLite (us)	Doltlite (us)	Multiplier
oltp_bulk_insert	243,514	315,596	1.30
oltp_insert	20,573	38,542	1.87
oltp_update_index	73,255	142,627	1.95
oltp_update_non_index	49,978	90,494	1.81
oltp_delete_insert	49,974	107,058	2.14
oltp_write_only	29,087	65,084	2.24
types_delete_insert	33,170	57,909	1.75
oltp_read_write	81,663	153,167	1.88
Average			1.87

File-Backed

Reads

Test	SQLite (us)	Doltlite (us)	Multiplier
oltp_point_select	123,164	129,168	1.05
oltp_range_select	24,285	53,407	2.20
oltp_sum_range	23,411	52,593	2.25
oltp_order_range	4,094	8,378	2.05
oltp_distinct_range	5,152	9,318	1.81
oltp_index_scan	14,789	22,713	1.54
select_random_points	35,345	89,125	2.52
select_random_ranges	13,314	19,555	1.47
covering_index_scan	13,395	14,516	1.08
groupby_scan	38,259	67,714	1.77
index_join	14,296	36,387	2.55
index_join_scan	7,110	26,068	3.67
types_table_scan	1,340,482	3,974,863	2.97
table_scan	1,594,566	4,832,627	3.03
oltp_read_only	261,787	379,941	1.45
Average			2.09

Writes

Test	SQLite (us)	Doltlite (us)	Multiplier
oltp_bulk_insert	254,245	344,739	1.36
oltp_insert	35,572	60,185	1.69
oltp_update_index	180,609	329,522	1.82
oltp_update_non_index	110,119	199,171	1.81
oltp_delete_insert	119,089	227,668	1.91
oltp_write_only	73,408	143,135	1.95
types_delete_insert	63,954	119,637	1.87
oltp_read_write	156,260	360,550	2.31
Average			1.84

File-Backed (autocommit)

Each statement runs as its own transaction — exposes per-commit
fixed costs that the wrapped-in-BEGIN/COMMIT tests amortize away.
SQLite uses WAL mode with synchronous=FULL in this section so
the comparison uses SQLite's durable WAL autocommit path.

Reads

Reads have no commit cost; these are the same SQL files as the
File-Backed Reads section, included here for symmetry and to
catch any per-statement overhead doltlite pays on the read path.

Test	SQLite (us)	Doltlite (us)	Multiplier
oltp_point_select	68,445	130,153	1.90
oltp_range_select	18,856	53,179	2.82
oltp_sum_range	18,128	52,910	2.92
oltp_order_range	3,678	8,376	2.28
oltp_distinct_range	4,693	9,294	1.98
oltp_index_scan	9,334	22,645	2.43
select_random_points	30,448	89,339	2.93
select_random_ranges	7,922	19,370	2.45
covering_index_scan	8,135	14,124	1.74
groupby_scan	37,395	67,573	1.81
index_join	12,036	35,766	2.97
index_join_scan	6,887	26,264	3.81
types_table_scan	1,364,763	4,049,834	2.97
table_scan	1,622,169	4,871,949	3.00
oltp_read_only	184,402	379,506	2.06
Average			2.54

Writes

Test	SQLite (us)	Doltlite (us)	Multiplier
oltp_bulk_insert_ac	15,855	56,952	3.59
oltp_insert_ac	18,152	89,988	4.96
oltp_update_index_ac	20,986	102,344	4.88
oltp_update_non_index_ac	16,315	78,349	4.80
oltp_delete_insert_ac	17,578	88,459	5.03
oltp_write_only_ac	18,299	88,840	4.85
types_delete_insert_ac	15,424	76,175	4.94
oltp_read_write_ac	24,934	100,463	4.03
Average			4.64

100000 rows, median of 5 invocations per test, workload-only timing via host monotonic clock when available.

Performance Ceiling Check (6x individual, 5x average)

All tests within ceilings.

[github-actions]

Sysbench-Style Benchmark (TEXT PK): Doltlite vs SQLite

Companion to the classic Sysbench-Style Benchmark. Every workload here
runs against tables with a 32-char hex TEXT PRIMARY KEY (UUID-shaped).
Individual ratios gated at 6×; section averages gated at 5×.

In-Memory

Reads

Test	SQLite (us)	Doltlite (us)	Multiplier
oltp_point_select	31,382	47,550	1.52
oltp_range_select	13,800	21,543	1.56
oltp_sum_range	12,688	20,696	1.63
oltp_order_range	3,030	4,027	1.33
oltp_distinct_range	4,122	5,183	1.26
oltp_index_scan	4,964	8,398	1.69
select_random_points	19,185	29,717	1.55
select_random_ranges	4,076	6,637	1.63
covering_index_scan	4,645	5,509	1.19
groupby_scan	33,696	39,798	1.18
index_join	7,086	12,652	1.79
index_join_scan	4,492	8,363	1.86
types_table_scan	1,217,924	1,654,377	1.36
table_scan	1,411,759	1,811,323	1.28
oltp_read_only	133,214	174,755	1.31
Average			1.48

Writes

Test	SQLite (us)	Doltlite (us)	Multiplier
oltp_bulk_insert	234,305	346,644	1.48
oltp_insert	23,845	40,300	1.69
oltp_update_index	79,318	162,602	2.05
oltp_update_non_index	53,611	101,607	1.90
oltp_delete_insert	53,484	121,951	2.28
oltp_write_only	32,638	69,462	2.13
types_delete_insert	34,729	64,204	1.85
oltp_read_write	98,915	174,743	1.77
Average			1.89

File-Backed

Reads

Test	SQLite (us)	Doltlite (us)	Multiplier
oltp_point_select	108,493	127,901	1.18
oltp_range_select	22,846	54,182	2.37
oltp_sum_range	22,203	53,462	2.41
oltp_order_range	4,031	8,304	2.06
oltp_distinct_range	5,145	9,371	1.82
oltp_index_scan	13,445	22,344	1.66
select_random_points	35,147	88,220	2.51
select_random_ranges	11,841	18,192	1.54
covering_index_scan	12,767	13,570	1.06
groupby_scan	35,381	65,546	1.85
index_join	14,563	35,218	2.42
index_join_scan	7,845	24,667	3.14
types_table_scan	1,299,965	3,885,877	2.99
table_scan	1,558,486	4,593,299	2.95
oltp_read_only	242,810	379,726	1.56
Average			2.10

Writes

Test	SQLite (us)	Doltlite (us)	Multiplier
oltp_bulk_insert	246,524	369,184	1.50
oltp_insert	47,904	73,443	1.53
oltp_update_index	192,876	321,613	1.67
oltp_update_non_index	134,218	194,823	1.45
oltp_delete_insert	130,265	222,252	1.71
oltp_write_only	88,672	136,761	1.54
types_delete_insert	67,559	120,137	1.78
oltp_read_write	181,244	352,089	1.94
Average			1.64

File-Backed (autocommit)

Each statement runs as its own transaction — exposes per-commit
fixed costs that the wrapped-in-BEGIN/COMMIT tests amortize away.
SQLite uses WAL mode with synchronous=FULL in this section so
the comparison uses SQLite's durable WAL autocommit path.

Reads

Reads have no commit cost; these are the same SQL files as the
File-Backed Reads section, included here for symmetry and to
catch any per-statement overhead doltlite pays on the read path.

Test	SQLite (us)	Doltlite (us)	Multiplier
oltp_point_select	61,570	124,697	2.03
oltp_range_select	17,943	53,025	2.96
oltp_sum_range	17,418	53,895	3.09
oltp_order_range	3,636	8,486	2.33
oltp_distinct_range	4,693	9,359	1.99
oltp_index_scan	8,963	22,406	2.50
select_random_points	30,131	90,049	2.99
select_random_ranges	7,360	18,108	2.46
covering_index_scan	8,669	13,568	1.57
groupby_scan	35,738	65,684	1.84
index_join	12,392	36,124	2.92
index_join_scan	7,807	25,867	3.31
types_table_scan	1,297,347	3,908,025	3.01
table_scan	1,530,578	4,611,645	3.01
oltp_read_only	178,922	379,181	2.12
Average			2.54

Writes

Test	SQLite (us)	Doltlite (us)	Multiplier
oltp_bulk_insert_ac	23,059	80,161	3.48
oltp_insert_ac	28,173	98,436	3.49
oltp_update_index_ac	33,385	125,213	3.75
oltp_update_non_index_ac	24,548	99,485	4.05
oltp_delete_insert_ac	27,791	110,389	3.97
oltp_write_only_ac	26,363	108,235	4.11
types_delete_insert_ac	22,266	94,038	4.22
oltp_read_write_ac	32,869	119,059	3.62
Average			3.84

100000 rows, median of 5 invocations per test, workload-only timing via host monotonic clock when available.

Performance Ceiling Check (6x individual, 5x average)

All tests within ceilings.