Backends and performance

Choose a backend using measurements of your dictionary, input and output volume.

Performance in charts

Measured snapshots, not a live benchmark or a claim about the latest npm release. Lower time and memory are better. Expand the tables for values and variation.

Scanning before and after optimization

All 18 scenarios share the same baseline. The dashed line is the original time (1×); a bar at 0.4× takes 40% of that time. Switch operations to see where the gains apply.

Operation

After / before · lower is faster · baseline = 1×

Exact measurements and variation

Scenario	Before · ms	After · ms	After / before	Time change
Tiny ASCII	0.001269675 (0.000014722; 0.000023004)	0.000136628 (0.00000104; 0.00000094)	0.107609	-89.24%
CRLF	2.572690125 (0.057604125; 0.026924375)	1.281498688 (0.02346875; 0.014897125)	0.498116	-50.19%
ASCII + combining	2.80034375 (0.043317625; 0.037565125)	2.659710937 (0.04278125; 0.04969275)	0.94978	-5.02%
Unicode at start	2.463591125 (0.030835938; 0.006440125)	2.329559937 (0.01759375; 0.027726625)	0.945595	-5.44%
ASCII prefix	2.87225525 (0.0160625; 0.068546938)	0.856292328 (0.007295562; 0.035654953)	0.298125	-70.19%
Early Unicode	4.819791625 (0.036427125; 0.114406125)	1.504946594 (0.014993531; 0.094178344)	0.312243	-68.78%
Ordinary text	2.021273437 (0.0326745; 0.0097005)	0.609461578 (0.006155594; 0.002634109)	0.301524	-69.85%
No matches	2.472632813 (0.011533937; 0.012984375)	0.718931641 (0.013856766; 0.023575516)	0.290756	-70.92%
Mixed Unicode	0.865691406 (0.013355469; 0.001882812)	0.862647125 (0.009561219; 0.012480469)	0.996483	-0.35%
Shared prefixes	1.442992187 (0.014175781; 0.060458313)	0.425919594 (0.002939461; 0.012159187)	0.295164	-70.48%
Dense suffixes	2.35503125 (0.020028625; 0.039721313)	2.357835937 (0.02818225; 0.098164062)	1.001191	+0.12%
High-output, no hit	4.63714575 (0.017401; 0.130124875)	4.655984375 (0.0189115; 0.163510375)	1.004063	+0.41%
High-output, sparse	4.782875 (0.043520875; 0.03275)	4.759760375 (0.030213625; 0.098999875)	0.995167	-0.48%
Duplicate patterns	2.288026062 (0.009601563; 0.022078063)	1.725445312 (0.031473938; 0.0095)	0.75412	-24.59%
Early hit	7.3808335 (0.04828125; 0.234719)	2.25433075 (0.05160675; 0.032132812)	0.30543	-69.46%
Late hit	7.4371145 (0.02235425; 0.1510415)	2.260109375 (0.037114625; 0.019760375)	0.303896	-69.61%
10k patterns	1.967362 (0.0115755; 0.017362)	0.601537109 (0.002729172; 0.005005203)	0.305758	-69.42%
Long text	20.820208 (0.292167; 0.155749)	20.587333 (0.39925; 0.385)	0.988815	-1.12%

Timing cells show median ms (within-round MAD; between-round MAD). MAD measures variation, not a confidence interval. — means that statistic was not recorded.

benchmarks-optimization.json · 2026-09-23 · 5 independent rounds · 0d10f03

Environment and full provenance

node: v24.18.0
platform: darwin
arch: arm64
cpu: Apple M4 Max
icu: 78.3
unicode: 17.0
Measured at: 2026-09-23T07:35:25.510Z
Source revision: 0d10f03bade9609faf6bbe086b0dd4666452c4e0
Working tree changes: true
Baseline: b01c9f2
Implementation SHA-256: 3b9119908bfda141b7ebba8f291ebba40f285cb04bbde0caadb87781cade0e36
Baseline SHA-256: c0e85759b5226525ba51f7dc1f95dc427340118dca47fc93c72d613e6e46cdbf
Corpus SHA-256: 4afa3eee0768ace1840771227bd94753d446ab82a09bbaf11d2acf1834462c5e
Report SHA-256: 00e2cd3c6630174fa4be71ff03169a4317e0650f9b8d5856cb6fc81c07b8239f

Later targeted rechecks and remaining tradeoffs

The scan chart keeps its original five-round snapshot. Later measurements below use their own implementation fingerprints and are not spliced into it. Repeated warm construction of the ordinary small dictionary changed by 7.3% (positive means slower). High-output dictionaries use native full enumeration even on no-hit text, so ASCII gains are not universal.

Exact measurements and variation

Scenario	Operation	Before · ms	After · ms	Time change	Rounds / source
Early Unicode	load	0.007536184 (0.000202433; 0.000046631)	0.007769521 (0.000126221; 0.000037882)	+3.1%	7 · scanRecheck
Early Unicode	build	0.006296193 (0.000046499; 0.000082372)	0.0060964 (0.000044271; 0.00004863)	-3.17%	7 · scanRecheck
Early Unicode	match	0.000103466 (0.000000839; 0.000003649)	0.000100652 (0.000000945; 0.000003679)	-2.72%	7 · scanRecheck
Mixed Unicode	load	0.161373859 (0.002421387; 0.001249348)	0.160725586 (0.001869141; 0.002076008)	-0.4%	7 · scanRecheck
Mixed Unicode	build	0.162988445 (0.001462242; 0.001569176)	0.154940594 (0.001020508; 0.000681969)	-4.94%	7 · scanRecheck
Mixed Unicode	match	0.003720205 (0.000103099; 0.000160105)	0.00419518 (0.000515127; 0.000273702)	+12.77%	7 · scanRecheck
Shared prefixes	load	1.019778656 (0.038187469; 0.019308594)	1.076451844 (0.037391937; 0.027854188)	+5.56%	7 · scanRecheck
Shared prefixes	build	1.314447938 (0.019407563; 0.023204406)	0.849103516 (0.008429687; 0.030935547)	-35.4%	7 · scanRecheck
Shared prefixes	match	0.000300666 (0.000001889; 0.000009721)	0.000310079 (0.000001184; 0.000011546)	+3.13%	7 · scanRecheck
Duplicate patterns	load	0.056904215 (0.001910727; 0.000343344)	0.057346436 (0.002123615; 0.001427814)	+0.78%	7 · scanRecheck
Duplicate patterns	build	0.036539266 (0.000353801; 0.000282308)	0.045132162 (0.000180543; 0.001015463)	+23.52%	7 · scanRecheck
Duplicate patterns	match	0.00009635 (0.000000731; 0.000001887)	0.000096013 (0.000001251; 0.000003813)	-0.35%	7 · scanRecheck
Duplicate patterns	load	0.0510448 (0.000971191; 0.000046184)	0.051154216 (0.000773642; 0.000258504)	+0.21%	7 · builderRecheck
Duplicate patterns	build	0.035663778 (0.000324342; 0.000484985)	0.017146688 (0.000130717; 0.000104512)	-51.92%	7 · builderRecheck
Duplicate patterns	match	0.000097522 (0.000000668; 0.000002916)	0.000099024 (0.000000907; 0.000001781)	+1.54%	7 · builderRecheck
Ordinary text	load	0.074622314 (0.002341635; 0.001869955)	0.078218018 (0.003306396; 0.00222526)	+4.82%	7 · builderRecheck
Ordinary text	build	0.075951334 (0.001940672; 0.002001219)	0.081459799 (0.001128416; 0.000841391)	+7.25%	7 · builderRecheck
Ordinary text	match	0.000145013 (0.000002099; 0.000007323)	0.000146319 (0.0000032; 0.000005473)	+0.9%	7 · builderRecheck
Shared prefixes	load	1.012390625 (0.029386719; 0.030472656)	1.033446625 (0.023977906; 0.045339844)	+2.08%	7 · builderRecheck
Shared prefixes	build	1.26889975 (0.042160187; 0.031625)	0.841124344 (0.031281234; 0.051948562)	-33.71%	7 · builderRecheck
Shared prefixes	match	0.000313238 (0.000002728; 0.000009483)	0.000354697 (0.000012394; 0.000044048)	+13.24%	7 · builderRecheck
Mixed Unicode	load	0.161232094 (0.002578449; 0.001061688)	0.158745281 (0.003340008; 0.001036133)	-1.54%	7 · builderRecheck
Mixed Unicode	build	0.164593262 (0.00159798; 0.001666504)	0.15644059 (0.002091469; 0.00152018)	-4.95%	7 · builderRecheck
Mixed Unicode	match	0.004181345 (0.000354461; 0.000251313)	0.003797536 (0.000316915; 0.000310079)	-9.18%	7 · builderRecheck
Shared prefixes	match	0.000295103 (0.000003017; 0.000013225)	0.000304079 (0.000003312; 0.000004858)	+3.04%	7 · presenceRecheck

benchmarks-optimization-recheck.json · 2026-09-23 · 7 independent rounds · 0d10f03

Environment and full provenance

node: v24.18.0
platform: darwin
arch: arm64
cpu: Apple M4 Max
icu: 78.3
unicode: 17.0
Measured at: 2026-09-23T07:50:21.300Z
Source revision: 0d10f03bade9609faf6bbe086b0dd4666452c4e0
Working tree changes: true
Baseline: b01c9f2
Implementation SHA-256: 3b9119908bfda141b7ebba8f291ebba40f285cb04bbde0caadb87781cade0e36
Baseline SHA-256: b6d995a840e4962a368f3f098a3e2c013a05f294ec8860efcd549ce9a5ff16a3
Corpus SHA-256: 4afa3eee0768ace1840771227bd94753d446ab82a09bbaf11d2acf1834462c5e
Report SHA-256: ba12ed87faa1dc14127bf7be75530b5d1477969c874c8f99e461320b2df08680

benchmarks-final-builder-recheck.json · 2026-09-23 · 7 independent rounds · dd97451

Environment and full provenance

node: v24.18.0
platform: darwin
arch: arm64
cpu: Apple M4 Max
icu: 78.3
unicode: 17.0
Measured at: 2026-09-23T08:16:31.795Z
Source revision: dd97451efdce9586bbe7d3f442a410ed77d07853
Working tree changes: true
Baseline: b01c9f2
Implementation SHA-256: c7af20de2e8a6a0064b3755ef656d653b0e6d469429ac614a7f0c03d5bc32a6c
Baseline SHA-256: 59e4199637e07de27cf118aee97e5cc736356eaf5813d2017811c39aa4e937c4
Corpus SHA-256: 4afa3eee0768ace1840771227bd94753d446ab82a09bbaf11d2acf1834462c5e
Report SHA-256: 4c02cf008470fc6b27a6d48f339ee593b22e476f3f268ef9406241caf7bbdcfe

benchmarks-presence-final-recheck.json · 2026-09-23 · 7 independent rounds · dd97451

Environment and full provenance

node: v24.18.0
platform: darwin
arch: arm64
cpu: Apple M4 Max
icu: 78.3
unicode: 17.0
Measured at: 2026-09-23T08:20:14.774Z
Source revision: dd97451efdce9586bbe7d3f442a410ed77d07853
Working tree changes: true
Baseline: b01c9f2
Implementation SHA-256: c7af20de2e8a6a0064b3755ef656d653b0e6d469429ac614a7f0c03d5bc32a6c
Baseline SHA-256: 3fc0b1a2891ab6fbc993d76904205615631e871dc262d41d3dac990d13b34725
Runner SHA-256: f45aeb0a9c3eda6c4b9fd27056167d3888c22f8a8c66f616ab829ea901243c55
Corpus SHA-256: 4afa3eee0768ace1840771227bd94753d446ab82a09bbaf11d2acf1834462c5e
Report SHA-256: 842c493eae1c6965b7456f24f639aea7fa86272975a13dcdd461c496c3a62a38

v1, v2 and optimized v3

With independent UTF-16 ranges as the shared output, optimized v3 is faster than v2 in 14 of 14 scenarios. v1 still wins on some ASCII inputs. Results that fail the correctness check have no bars or speed ratios.

Independent ranges · lower is faster · v2 = 1×. v1 is not comparable for emoji, ZWJ, combining marks and CRLF.

Independent range timings · median ms (MAD; between-round MAD)

Scenario	v1 · ms	v2 · ms	Optimized v3 · ms	Rounds / source
Ordinary text	0.671835937 (0.007359375; 0.019706375)	3.49615625 (0.027041625; 0.029947875)	0.598543625 (0.005460937; 0.003548172)	5 · original
No matches	0.421351562 (0.004225906; 0.008166016)	4.42000525 (0.039885375; 0.0384895)	0.598173172 (0.005952469; 0.024811844)	7 · recheck
Shared prefixes	0.475947906 (0.006143219; 0.009919281)	2.524388 (0.020815125; 0.046349)	0.393973313 (0.003625977; 0.007260742)	5 · original
Dense suffixes	3.114309938 (0.021072938; 0.015197875)	3.259916688 (0.0251745; 0.051653688)	2.305966125 (0.025445312; 0.0263125)	5 · original
Duplicate patterns	1.871252625 (0.017622375; 0.054755188)	3.55315625 (0.02338525; 0.03960425)	1.631329437 (0.016760406; 0.011783875)	5 · original
Early hit	1.580188812 (0.033903625; 0.009697906)	13.1050625 (0.1155415; 0.08775)	2.129585938 (0.015182313; 0.073263)	7 · recheck
Late hit	1.568641938 (0.013188812; 0.041936187)	12.9421045 (0.0714375; 0.1565215)	2.169085938 (0.027036438; 0.052179688)	7 · recheck
10k patterns	0.819513016 (0.022947266; 0.032115234)	3.840359375 (0.053541625; 0.099510375)	0.598176438 (0.006532563; 0.005749344)	5 · original
Long text	12.49685425 (0.09892725; 0.2773645)	60.378291 (1.155417; 0.773166)	9.545948 (0.07855225; 0.21780225)	5 · original
Chinese	2.33740625 (0.043601625; 0.076854188)	4.956802 (0.06201575; 0.082036375)	2.350195313 (0.0244115; 0.056028625)	5 · original
Emoji	Not comparable	1.875125 (0.031598937; 0.0247005)	0.809717438 (0.003578125; 0.009024719)	5 · original
ZWJ sequences	Not comparable	0.700683594 (0.004184891; 0.005352203)	0.331774742 (0.002745773; 0.001072594)	5 · original
Combining marks	Not comparable	2.134664063 (0.026484375; 0.01678125)	1.042622375 (0.008112; 0.008574188)	5 · original
CRLF	Not comparable	2.288255187 (0.019752563; 0.003966125)	0.570280609 (0.004173156; 0.010173844)	5 · original

Native v1/v2 search returns grouped output; v3 returns independent match objects. The native timings below measure different output work and should not be read as equivalent throughput.

Native output timings · median ms (MAD; between-round MAD)

Scenario	v1 · ms	v2 · ms	Optimized v3 · ms	Rounds / source
Ordinary text	0.479438141 (0.007430984; 0.004863266)	1.701596375 (0.018164062; 0.020921937)	0.610501953 (0.004858719; 0.014290375)	7 · recheck
No matches	0.423757812 (0.005669266; 0.006125)	2.1667995 (0.017763; 0.02735675)	0.598205719 (0.007033203; 0.028513016)	7 · recheck
Shared prefixes	0.436317703 (0.004397125; 0.000654312)	1.302658875 (0.008703156; 0.014859406)	0.392411781 (0.002458328; 0.011767258)	5 · original
Dense suffixes	0.007156077 (0.000054026; 0.000077382)	0.051514242 (0.000309977; 0.000531006)	2.313322875 (0.032333313; 0.012132875)	7 · recheck
Duplicate patterns	0.327894531 (0.00320475; 0.001456055)	0.943231781 (0.005602844; 0.011592437)	1.623438813 (0.014154938; 0.016690125)	7 · recheck
Early hit	1.584460937 (0.019053406; 0.014001313)	6.58317725 (0.04205225; 0.0798855)	2.144432313 (0.017877625; 0.087416625)	7 · recheck
Late hit	1.571467438 (0.013057313; 0.020328125)	6.478395875 (0.0583645; 0.109989625)	2.182192687 (0.015388; 0.026875)	7 · recheck
10k patterns	0.65314975 (0.004081359; 0.016722016)	1.886552125 (0.0406745; 0.010877562)	0.599035156 (0.005404297; 0.011255859)	5 · original
Long text	8.896427 (0.0968125; 0.1152395)	29.96625 (0.276334; 0.093042)	9.651927 (0.06138525; 0.2726145)	7 · recheck
Chinese	1.906877563 (0.02800525; 0.02590625)	2.318304687 (0.052122375; 0.033828125)	2.413601563 (0.051817688; 0.018005188)	7 · recheck
Emoji	Not comparable	1.018970031 (0.008847656; 0.015799438)	0.819953125 (0.005609375; 0.018988281)	5 · original
ZWJ sequences	Not comparable	0.383055344 (0.00359375; 0.000662766)	0.331748703 (0.002734055; 0.001452789)	5 · original
Combining marks	Not comparable	1.077026031 (0.004104156; 0.013281219)	1.052846375 (0.010964875; 0.005266937)	5 · original
CRLF	Not comparable	1.088936187 (0.006821625; 0.011173188)	0.584257156 (0.007115234; 0.009666672)	5 · original

A seven-round recheck replaces a complete scenario/operation group across all three versions only after source, environment, runner and corpus validation. Remaining groups keep five rounds. Memory measurements stay with the original report.

benchmarks-versions-optimization.json · 2026-09-23 · 5 independent rounds · dd97451

Environment and full provenance

node: v24.18.0
pnpm: pnpm@12.5.1
platform: darwin
arch: arm64
cpu: Apple M4 Max
icu: 78.3
unicode: 17.0
Measured at: 2026-09-23T08:22:10.112Z
Source revision: dd97451efdce9586bbe7d3f442a410ed77d07853
Working tree changes: true
Implementation SHA-256: c7af20de2e8a6a0064b3755ef656d653b0e6d469429ac614a7f0c03d5bc32a6c
Runner SHA-256: e557a2c3f4c207fb5a8f2dbf24a9da15572830e94a005716559218cb6323f7ea
Corpus SHA-256: 8dbac3966d9f23c4047c59576d55779948d57a39c7603acdf7b748506f701d4d
Report SHA-256: f9df675eeca635f032eb77e18f107bd6ba7d8031194ee4a4fcf7dcb2d6dcd9f3
v1 · 1.1.0: 92e364bded18733ea20c25d2ba2773677970600be6d65da05fa325267a19d2c1
v2 · 2.0.4: 8de77ed31903160adb53c77cf4aff80032144f685710c719c5490e592c9d7397
current · 3.2.0: c7af20de2e8a6a0064b3755ef656d653b0e6d469429ac614a7f0c03d5bc32a6c

benchmarks-versions-optimization-recheck.json · 2026-09-23 · 7 independent rounds · dd97451

Environment and full provenance

node: v24.18.0
pnpm: pnpm@12.5.1
platform: darwin
arch: arm64
cpu: Apple M4 Max
icu: 78.3
unicode: 17.0
Measured at: 2026-09-23T08:31:59.230Z
Source revision: dd97451efdce9586bbe7d3f442a410ed77d07853
Working tree changes: true
Implementation SHA-256: c7af20de2e8a6a0064b3755ef656d653b0e6d469429ac614a7f0c03d5bc32a6c
Runner SHA-256: e557a2c3f4c207fb5a8f2dbf24a9da15572830e94a005716559218cb6323f7ea
Corpus SHA-256: 8dbac3966d9f23c4047c59576d55779948d57a39c7603acdf7b748506f701d4d
Report SHA-256: ef2bb327ec120146210728cadba8d622c0498530b18a4072086db5daace5f1c2
v1 · 1.1.0: 92e364bded18733ea20c25d2ba2773677970600be6d65da05fa325267a19d2c1
v2 · 2.0.4: 8de77ed31903160adb53c77cf4aff80032144f685710c719c5490e592c9d7397
current · 3.2.0: c7af20de2e8a6a0064b3755ef656d653b0e6d469429ac614a7f0c03d5bc32a6c

Cold construction at scale

At 1,000,000 patterns, cold construction time changed by -59% and peak RSS by -71.3%. These are seven-process medians; cold process construction is distinct from repeated warm construction.

Construction time

Peak process RSS

Patterns on the horizontal axis. Peak RSS includes the process runtime and temporary construction state; it is not the retained dictionary size. The historical baseline loads modules through data URLs, so module-loader storage also differs within this whole-process measurement.

Exact measurements and variation

Patterns	Before · ms	After · ms	Time change	Before MAD · ms	After MAD · ms	Before RSS · MiB	After RSS · MiB	RSS change	Before RSS MAD · MiB	After RSS MAD · MiB	Before heap · MiB	After heap · MiB	Before buffers · MiB	After buffers · MiB
10,000	31.9375	20.737292	-35.07%	0.640625	1.149459	106.296875	71.625	-32.62%	0.640625	0.453125	0.54612	0.541451	1.602516	1.602516
100,000	186.384875	92.04775	-50.61%	2.611084	2.525708	407.109375	140.375	-65.52%	0.625	0.21875	3.495079	3.490295	16.022072	16.022072
1,000,000	1,945.30275	797.322209	-59.01%	67.384125	20.274958	2,300.21875	660.15625	-71.3%	2.453125	1.234375	34.217422	34.212784	160.217628	160.217628

benchmarks-builder-final.json · 2026-09-23 · 7 independent rounds · dd97451

Environment and full provenance

node: v24.18.0
platform: darwin
arch: arm64
cpu: Apple M4 Max
icu: 78.3
unicode: 17.0
Measured at: 2026-09-23T08:20:48.462Z
Source revision: dd97451efdce9586bbe7d3f442a410ed77d07853
Working tree changes: true
Baseline: b01c9f2
Implementation SHA-256: c7af20de2e8a6a0064b3755ef656d653b0e6d469429ac614a7f0c03d5bc32a6c
Baseline SHA-256: 18cd42aabc95d9d3dfcd94ae9983fdd8d1ad7e2908860f26cc36f38136e28dbf
Runner SHA-256: 05d461798498b3528241cf8dc5e088b4297f93da826d6caeeab01939ba2c7b85
Report SHA-256: ed7f42a9f1abb3eef12296b678b40edeead71419619cba954d7371fc20f2832a

Memory retained by the dictionary

The dedicated retention run disables timing and averages multiple live dictionary copies. The bars compare retained JS heap; ArrayBuffer storage is reported separately in bytes. Buffer measurements are unchanged in 18 of 18 scenarios.

After / before · baseline = 1× · zero-based scale to avoid exaggerating small differences

Exact measurements and variation

Scenario	Before heap · B	After heap · B	Heap change	Before buffers · B	After buffers · B
Tiny ASCII	3,929.904	3,902.088	-0.71%	0	0
CRLF	4,366.952	4,368.912	+0.04%	0	0
ASCII + combining	4,192.584	4,191.728	-0.02%	0	0
Unicode at start	4,148.256	4,146.648	-0.04%	0	0
ASCII prefix	4,081.648	4,066.56	-0.37%	0	0
Early Unicode	3,907.472	3,900.928	-0.17%	0	0
Ordinary text	13,854.672	13,817.168	-0.27%	8,156	8,156
No matches	29,062.84	29,109.44	+0.16%	20,356	20,356
Mixed Unicode	9,098.68	9,102.552	+0.04%	7,260	7,260
Shared prefixes	55,875.84	55,757.2	-0.21%	264,568	264,568
Dense suffixes	8,500.048	8,497.984	-0.02%	3,868	3,868
High-output, no hit	5,992.08	6,041.376	+0.82%	328	328
High-output, sparse	5,992.08	6,041.392	+0.82%	328	328
Duplicate patterns	14,492.592	14,432.592	-0.41%	1,600	1,600
Early hit	3,907.472	3,900.928	-0.17%	0	0
Late hit	3,907.472	3,900.928	-0.17%	0	0
10k patterns	498,346.4	496,666.4	-0.34%	1,680,360	1,680,360
Long text	4,845.152	4,823.632	-0.44%	0	0

benchmarks-retained-optimization.json · 2026-09-23 · 7 independent rounds · 0d10f03

Environment and full provenance

node: v24.18.0
platform: darwin
arch: arm64
cpu: Apple M4 Max
icu: 78.3
unicode: 17.0
Measured at: 2026-09-23T08:13:54.285Z
Source revision: 0d10f03bade9609faf6bbe086b0dd4666452c4e0
Working tree changes: true
Baseline: b01c9f2
Implementation SHA-256: 6716d18808b34e066dc30eaaaa8b90d773adc567feda47150f18e51b0283296c
Baseline SHA-256: 68cd0ffc3df077980b83ce045256e4627b9c006a0d086f334c32492bb13ffb18
Corpus SHA-256: 4afa3eee0768ace1840771227bd94753d446ab82a09bbaf11d2acf1834462c5e
Report SHA-256: 71bc3027f7f0ff0101008e492526541a79f895ec517aff726c158a9e87a32ad0

10,000-pattern dictionary across versions

Stacked components share the same unit: 1 MiB = 1,048,576 bytes. This excludes returned result objects and process peak RSS.

Dictionary memory by version

Version	JS heap · B	ArrayBuffer · B	Combined · MiB
v1	7,366,382.4	0	7.02513
v2	7,366,560	0	7.025299
Optimized v3	496,567.2	1,680,360	2.07608

benchmarks-versions-optimization.json · 2026-09-23 · 5 independent rounds · dd97451

Environment and full provenance

node: v24.18.0
pnpm: pnpm@12.5.1
platform: darwin
arch: arm64
cpu: Apple M4 Max
icu: 78.3
unicode: 17.0
Measured at: 2026-09-23T08:22:10.112Z
Source revision: dd97451efdce9586bbe7d3f442a410ed77d07853
Working tree changes: true
Implementation SHA-256: c7af20de2e8a6a0064b3755ef656d653b0e6d469429ac614a7f0c03d5bc32a6c
Runner SHA-256: e557a2c3f4c207fb5a8f2dbf24a9da15572830e94a005716559218cb6323f7ea
Corpus SHA-256: 8dbac3966d9f23c4047c59576d55779948d57a39c7603acdf7b748506f701d4d
Report SHA-256: f9df675eeca635f032eb77e18f107bd6ba7d8031194ee4a4fcf7dcb2d6dcd9f3
v1 · 1.1.0: 92e364bded18733ea20c25d2ba2773677970600be6d65da05fa325267a19d2c1
v2 · 2.0.4: 8de77ed31903160adb53c77cf4aff80032144f685710c719c5490e592c9d7397
current · 3.2.0: c7af20de2e8a6a0064b3755ef656d653b0e6d469429ac614a7f0c03d5bc32a6c

Long streams with continuous consumption

The caller consumes every output batch immediately. Points show post-GC heap growth relative to a baseline taken after creating the dictionary and input. This excludes their preexisting memory and caller-retained outputs, as well as native ICU allocations and peak RSS. Straight lines only connect measured input sizes; runtime and JIT noise remain included.

all

leftmost-longest

Before EOF, optimized leftmost-longest had already emitted 79,997 of 80,000 matches. The table also shows retained memory with the handle still held after finish and cancel.

Exact measurements and variation

Strategy / snapshot	UTF-16	Heap growth · KiB	Emitted before EOF	Final matches	Closed heap growth · KiB
all · Before	40,000	2,031.015625	11,997	120,000	14,666.757813
all · Before	120,000	3,848.695313	35,997	120,000	14,666.757813
all · Before	400,000	14,707.84375	119,997	120,000	14,666.757813
all · After	40,000	263.6875	11,997	120,000	158.46875
all · After	120,000	268.117188	35,997	120,000	158.46875
all · After	400,000	264.007813	119,997	120,000	158.46875
leftmost-longest · Before	40,000	2,899.570313	0	80,000	23,179.203125
leftmost-longest · Before	120,000	6,388.929688	0	80,000	23,179.203125
leftmost-longest · Before	400,000	23,208.25	0	80,000	23,179.203125
leftmost-longest · After	40,000	271.140625	7,997	80,000	158.4375
leftmost-longest · After	120,000	275.679688	23,997	80,000	158.4375
leftmost-longest · After	400,000	271.09375	79,997	80,000	158.4375

benchmarks-stream-optimization.json · 2026-09-23 · 3 independent rounds · b6ebdce

Environment and full provenance

node: v24.18.0
platform: darwin
arch: arm64
cpu: Apple M4 Max
icu: 78.3
unicode: 17.0
Measured at: 2026-09-23T06:23:33.062Z
Source revision: b6ebdce7f04a8c511060a9af9a65ff64e9c9c04f
Working tree changes: true
Baseline: b01c9f22b2566b3a6058b41b05564b214899c8d7
Implementation SHA-256: 593af04bb53db03a3dd84c0be75c30ec04dc51cbd6e863af74db75d6a0e1fb7a
Baseline SHA-256: 8c4bb499ac67a4b7917f264e0f84b36c5d88a3ba7cd0f5cdb0c75fd87b9b3566
Report SHA-256: 0da839a7d97d632f9646508c5cc4a96da35371bb9e689a4013659235b573d389

Backend choices

Start with the default compact backend. /fast opts into double-array storage; /unicode-fast combines that backend with full folding. The default import does not load the optional compiler or folding data.

import FastAhoCorasick from 'modern-ahocorasick/fast'

const matcher = new FastAhoCorasick(['cat'])
matcher.search('cat')[0]?.start // 0
matcher.getStats().backend // 'double-array'

Measured tradeoffs

/fast is an explicit double-array backend, not a universal speed guarantee. It can reduce ArrayBuffer storage while increasing JavaScript heap usage; cursor overhead may dominate scans. In the snapshot before the streaming, scanner and builder optimizations, on Node 24.18.0 / Apple M4 Max, the 10,000-pattern ASCII comparison measured 28.23 ms construction and 4.34 ms search for /fast, versus 13.61 ms and 2.00 ms for the default backend. Retained heap was 3.52 MiB versus 0.47 MiB; buffers were 1.18 MiB versus 1.60 MiB. These are three-round medians, not measurements of the optimized build or a prediction for other dictionaries. Boundaries and folding use the general cursor, so filtered count() does enumerate accepted outputs. The default exact counting and presence paths retain their allocation-light implementation.

The feature comparison targets @monyone/aho-corasick@1.5.10 and @tanishiking/aho-corasick@0.0.1. Their index, duplicate and Unicode contracts differ; this package preserves its own ranges and semantics rather than copying defects or method aliases. Run pnpm benchmark:external for the printable-ASCII comparison, including normalized independent ranges, native results, build costs and memory.

Run pnpm benchmark:versions to compare pinned npm v1.1.0 and v2.0.4 against the current local default entry: construction, native search, conversion to independent ranges and retained memory. Historical Unicode cases with incorrect results receive no speed ratios. The optimized-build version comparison includes the current measurements; the earlier v1/v2/v3 report remains a historical snapshot.

In the optimized build's five-round comparison, independent range output was faster than v2 in all 14 fixtures. Seven-round rechecks confirmed v1 still scanned some ASCII inputs faster, and historical native grouped output has different allocation costs. The 10,000-pattern dictionary retained about 2.08 MiB of heap plus buffers, versus 7.03 MiB for either older version. These results support choosing by the output contract and workload, rather than a universal version ranking.

The streaming, scanning and construction report records the subsequent changes against b01c9f2: incremental transformed-stream retention, an 18-scenario scanner comparison, retained-dictionary measurements, and cold construction at 10,000–1,000,000 patterns. It distinguishes repeated warm construction, retained heap, ArrayBuffer storage and construction peak RSS, with raw process results, implementation fingerprints and targeted rechecks.

Full enumeration uses native grapheme segmentation for dictionaries that can emit many matches at one state. This avoids extra overhead in dense output, but those dictionaries also use that path when an input has few or no matches. ASCII acceleration therefore depends on the dictionary as well as the text; measure representative no-hit inputs too. Counting and selected-match scans keep their own paths.

Reproduce and interpret

pnpm benchmark
pnpm benchmark:external
pnpm benchmark:versions
pnpm benchmark:scale
pnpm benchmark:stream
pnpm benchmark:docs

Run these commands from the repository. Compare build and scan time, retained heap and buffers separately; getStats().typedArrayBytes is not total heap. Filtering, folding, result arrays and longest-first have costs beyond the default exact scan. The workbench includes debounce, Worker and transfer time, so its display is not a library benchmark.

Sources: Extension benchmarks · v2/v3 benchmarks · ASCII benchmarks · Workbench measurements

Backends and performance ​

Performance in charts ​

Scanning before and after optimization

v1, v2 and optimized v3

Cold construction at scale

Construction time

Peak process RSS

Memory retained by the dictionary

10,000-pattern dictionary across versions

Long streams with continuous consumption

all

leftmost-longest

Backend choices ​

Measured tradeoffs ​

Reproduce and interpret ​

Backends and performance

Performance in charts

Backend choices

Measured tradeoffs

Reproduce and interpret