Skip to main content

Results at a glance

Every cell below is one library driving one screen-shaped task. It was verified against a plain-Map reference model before timing, then timed in its own processes beside a hand-written anchor. Each row is one scenario, named as in its own table below, and the fastest library's cell carries a 🏆 and its time; every other cell says how many times slower that library is than the fastest one: 1.3×, 12×. The two baseline-* columns are plain Map code with no library; they are shown for scale and never win a row.

Most rows are hand-written collections on top of a library

Most libraries here ship no collections, group indexes or sorted windows. In those rows the collections are written by hand on top of the library's primitives, without OIMDB, so the row measures that code, not a feature the library ships. Such a column is marked "+ hand-written collections".

scenariobaseline-arraybaseline-indexedoimdboimdb-sorted-indexoimdb-3.12.1tinybasesignaldbtanstack-dbtanstack-db-eagertanstack-db-indexedlokijsmobx+ hand-written collectionsjotai+ hand-written collectionslegend-state+ hand-written collectionseffector+ hand-written collectionsnanostores+ hand-written collectionsnanostores-per-row+ hand-written collectionsredux-toolkit+ hand-written collectionspreact-signals+ hand-written collectionsalien-signals+ hand-written collectionszustand+ hand-written collectionsvaltio+ hand-written collectionsxstate-store+ hand-written collectionsvue-reactivity+ hand-written collectionssolid-js+ hand-written collectionstanstack-store+ hand-written collections
batch insert0.74×3.2×3.9×1.7×17×68×72×19×40×41×6.0×25×14×51×2.0×2.3×3.2×16×3.5×4.0×1.1×44×🏆 998.13 µs3.4×5.1×1.8×
insert into a populated store3.1×3.2×2.7×2.1×7.1×14×66×23×25×20×🏆 113.67 µs7.9×5.1×18×7.4×67×2.9×45×2.7×2.8×20×10×21×2.6×3.7×2.6×
the population keeps growing1.3×0.77×1.2×🏆 1.082 ms7.7×28×102×16×25×24×1.4×12×8.9×28×5.7×22×2.8×48×1.6×2.3×20×33×20×1.8×3.4×1.1×
cold start to first render0.71×1.8×2.1×1.6×8.7×68×68×24×38×38×5.7×20×14×50×2.0×2.1×3.0×16×3.6×3.7×1.1×44×🏆 1.036 ms4.1×4.7×1.7×
tear the screen down1.4×2.0×2.4×1.9×18×5517×101×26×39×35×16×72×110×21×4.3×5.0×7.9×2.6×3.6×3.0×1.1×65×🏆 2.75 µs3.7×5.3×4.0×
patch one column28×2.7×1.1×🏆 133.52 µs1.1×4.2×1397×11×11×11×2.7×1.7×40×111×4.3×2.8×2.2×784×1.8×1.2×524×21×524×1.9×2.6×3.3×
patch a row off screen3216×1.9×2.4×2.6×5.5×258×150113×632×663×678×309×5.6×205×13227×492×319×257×91450×🏆 1.14 µs1.1×61887×2385×61757×1.1×1.2×372×
patch an unread column3166×0.85×3.3×1.9×112×193×160005×921×925×960×298×5.4×231×73×481×312×254×88726×🏆 1.20 µs1.1×60108×2308×58957×1.1×1.2×360×
batch patch2.2×1.0×1.4×1.4×2.9×30×1409×42×51×48×6.5×5.0×18×43×9.6×225×4.7×62×1.4×1.0×33×7.4×32050×🏆 69.54 µs1.7×2.2×
batch patch an unread column2.7×0.67×1.2×1.1×3.5×35×1720×53×60×60×7.9×4.6×15×32×12×279×5.9×77×1.2×🏆 56.26 µs42×9.4×40231×1.0×1.2×2.7×
most rows change every frame0.70×0.70×1.4×1.5×2.7×27×1226×50×56×55×4.8×4.0×12×26×3.6×184×3.8×21×1.3×🏆 901.46 µs6.3×6.1×25019×1.2×1.4×1.5×
one frame, every kind of change2.0×1.00×1.5×1.2×4.9×9.7×110×11×13×13×23×5.2×7.2×18×1.2×7.6×1.6×57×1.1×1.0×25×4.2×304×🏆 94.79 µs1.3×1.2×
row moves between groups9.2×2.8×1.7×🏆 101.76 µs2.8×9.2×522×6.4×6.3×6.4×1.8×18×23×76×3.1×5.8×1.9×270×1.2×1.3×177×13×178×1.3×1.5×2.1×
a row arrives mid-order6.0×1.9×1.6×1.1×3.4×18×328×11×11×12×2.1×13×20×60×2.3×7.9×1.9×179×🏆 80.79 µs1.1×111×10×220×1.1×1.5×1.6×
sort key moves21×2.0×1.8×1.1×4.4×10×1146×13×19×19×2.8×15×26×86×3.3×2.1×1.6×615×🏆 43.82 µs1.1×410×16×407×1.5×1.8×2.4×
the whole window re-sorted6.9×0.71×1.3×1.7×28×14×614×37×41×41×1.8×6.2×17×33×1.3×1.7×1.2×204×1.1×1.7×136×6.2×4376×1.5×2.2×🏆 132.34 µs
a consumer per group385×1.5×2.1×1.0×3.8×18×3474×8.2×9.3×9.3×4.3×17×31×71×2.6×4.7×2.0×440×🏆 454.18 µs1.2×325×14×313×1.2×1.5×1.9×
targeted column delivery1.8×0.94×1.0×1.0×2.1×3.4×3832×28×28×29×1.9×1.5×3.5×7.0×13×9.0×6.4×2883×🏆 18.69 µs1.0×1939×1.3×1930×1.1×1.4×1.1×
delete the visible window4.5×0.69×1.7×1.5×15×23×342×49×42×45×114×8.3×14×32×1.0×1.7×1.5×148×🏆 34.25 µs1.3×68×4.0×67×1.0×2.0×1.2×
delete one visible row19×1.9×1.9×1.4×5.4×22×810×13×14×15×190×18×22×85×2.9×9.2×2.0×693×🏆 57.46 µs1.2×315×12×315×1.2×1.8×2.5×
the group empties under the view3.5×1.0×1.9×1.8×16×52×474×59×68×69×17×8.5×22×46×🏆 78.35 µs1.3×1.5×109×1.5×1.5×57×14×57×1.2×1.9×1.5×
re-send unchanged list68×7.0×1.9×1.9×1.5×48×5240×20×21×20×14×8.1×83×24×12×12×10×1325×1.1×2.4×1331×12×1342×🏆 1.69 µs1.8×10×
mount and unmount one subscriber1.1×·1.0×🏆 2.58 µs2.6×14×11967×250×264×255×1.7×6.1×5.2×19×5.8×1.1×1.1×20829×1.7×1.2×13860×3.2×13956×2.7×3.2×1.2×
N subscribers on one root, oldest out first0.96×·1.0×1.0×1.9×8.4×2296×61×66×61×2.2×13×4.9×24×12×🏆 19.47 µs1.0×2725×3.2×2.1×1834×4.6×1838×9.1×6.7×1.7×
N subscribers on one root, newest out first1.0×·🏆 20.12 µs1.0×1.8×8.0×2276×54×60×52×1.6×13×4.7×23×12×1.1×1.1×2650×5.2×2.0×1772×4.4×1777×6.9×6.0×1.6×
N×8 subscribers on one root, oldest out first1.1×·1.0×🏆 147.19 µs1.9×7.5×✗50×54×49×6.9×12×4.6×23×12×1.2×1.2×367×5.7×2.2×243×4.6×245×7.7×6.7×1.9×
N×8 subscribers on one root, newest out first1.4×·🏆 146.71 µs1.0×1.7×7.5×✗46×49×47×2.3×12×4.5×23×12×1.5×1.4×367×5.4×2.2×244×4.6×247×7.8×8.8×1.9×
one subscriber over many sources1.9×·1.0×🏆 13.41 µs5.1×20×2422×38×39×37×1.9×25×28×27×50×3190×3197×3968×7.6×4.8×2656×4.6×2652×9.5×20×1.3×
N subscribers on one column0.93×·1.1×1.1×2.0×8.3×2260×✗✗✗·8.3×36×28×73×1.5×1.5×2629×1.5×🏆 20.32 µs1754×2.5×1762×4.8×3.6×2.7×
N subscribers on N rows, a list1.1×·1.0×🏆 26.88 µs2.5×16×1.9×1.1×
N subscribers over a derivation chain1.7×·2.0×2.0×✗✗·13×19×348×5.7×🏆 153.48 µs·2.7×2.5×
a write through standing chains0.54×·🏆 51.71 µs1.0×2.2×✗·7.1×22×1055×1.2×1.1×·1.4×2.5×
N subscribers over a diamond1.2×·1.8×1.9×11×·2.6×🏆 110.04 µs·2.8×3.7×
a write through standing diamonds0.27×·1.0×1.0×4.2×·1.2×🏆 78.31 µs·1.8×3.6×
churn against a standing population0.71×·1.0×🏆 93.90 µs1.6×1.1×
an application screen fills1.1×·🏆 22.08 µs1.0×8.8×1.1×
an application screen in use0.65×·1.0×🏆 341.73 µs1.7×7.7×2.5×1.9×5.2×5.0×
one setting moves0.86×·🏆 4.68 µs1.0×1.4×1.2×1.9×2.2×
read a page by key0.98×0.98×1.0×1.0×🏆 36.11 µs1.2×1.3×
scroll the window21×14×1.2×1.3×2.0×🏆 42.36 µs1.0×1.2×1.5×
scroll deep into the group21×4.8×1.3×2.3×1.8×1.1×🏆 48.67 µs1.3×1.6×
the view switches group19×1.6×2.3×2.3×1.5×1.0×🏆 51.30 µs1.2×1.6×
a view filtered on two axes13×3.3×2.2×1.1×5.1×🏆 34.15 µs1.1×1.1×1.7×
a write to a related entity2.0×·1.0×🏆 7.78 µs1.5×1.5×1.5×1.6×1.8×
a record keyed by two identities1.9×·1.0×🏆 8.94 µs1.6×1.7×1.6×1.8×2.1×
write one value0.85×·1.1×1.1×1.8×1.1×🏆 5.87 µs
re-write one value21×·5.1×5.1×1.1×🏆 80 ns

How to read a cell

  • Colour runs from green to red: dark green is the fastest library, light green is a little slower, yellow is about twice as slow, and red is slower still, darkest at an order of magnitude and beyond. The thresholds are in BenchMatrix.tsx. The value is always printed; colour only helps the eye find the pattern.
  • Hatched cells are stale (the code, a library version or the sampling plan changed since they were measured) or flagged (the batch's control pair could not resolve its own known difference, the machine was busy, a process was lost). Tick only fresh, unflagged cells to hide them.
  • · means the library declined that wiring, with a written reason in the scenario tables. ✗ means it ran and either disagreed with the reference model or crashed; it gets no number. An empty dashed cell has not been measured yet.
  • Hover a cell for the absolute time, drift, render counts and date.
  • The six floors (scripts with no store, measured once to bound the harness's own share) are instruments, not rows to compare, so they appear only in the scenario tables.

The matrix ranks nothing on its own: a place is given only in the scenario tables, and only where a row stands clear of every row above it by more than both rows' measured drift.

Published 2026-10-31 19:36 UTC from arm64 8× Apple M1, Node v24.16.0.

How much code each task took​

Time is one cost; the code a developer writes to get it is another. Each cell is the number of code lines (comments and blanks excluded) reachable from the participant's wiring for that task's shape: its builder, every helper it calls and the constants it reads, but not the methods only another shape needs. The source is first re-printed in one canonical layout (the TypeScript printer, every block and object literal one member per line, comments dropped), so joining calls onto one long line or splitting them changes nothing. The count is computed from the adapter source by npm run lines, needs no measurement, and is current on every publish. Each row is one wiring, shared by the tasks it names; hover it for the list. The library with the fewest lines in a row gets a 🏆, and the colour runs from green (fewest) to red (ten times as many or more). The two baseline-* columns are plain Map code, shown grey for scale. Fewer lines is not better on its own: read it beside the time, never instead of it.

lines of code to wirebaseline-arraybaseline-indexedoimdboimdb-sorted-indexoimdb-3.12.1tinybasesignaldbtanstack-dbtanstack-db-eagertanstack-db-indexedlokijsmobx+ hand-written collectionsjotai+ hand-written collectionslegend-state+ hand-written collectionseffector+ hand-written collectionsnanostores+ hand-written collectionsnanostores-per-row+ hand-written collectionsredux-toolkit+ hand-written collectionspreact-signals+ hand-written collectionsalien-signals+ hand-written collectionszustand+ hand-written collectionsvaltio+ hand-written collectionsxstate-store+ hand-written collectionsvue-reactivity+ hand-written collectionssolid-js+ hand-written collectionstanstack-store+ hand-written collections
Board screen 21 tasks123246285285216214271249249249212200305🏆 182314216216290298328245212297332322244
A consumer per group 1 task117240279279210208265243243243206194299🏆 176308210210284292322239206291326316238
Subscribe and unsubscribe 10 tasks151·155155117🏆 95210189189189172152162114179104104225143141215116228172140130
One subscriber over many sources 1 task151·155155117🏆 95210189189189172152162114179104104225143141215116228172140130
Subscribers on one column 1 task148·157157101🏆 89208199199199·1431511051699191220133131210103215172140127
Subscribers over a derivation chain 2 tasks167·229229133112201219219219·154158116174🏆 101🏆 101224144142214·226178147130
Subscribers over a diamond 2 tasks161·229229123110193212212212·146150108172🏆 94🏆 94224137135214·219176145129
Read a page by key 1 task89161193193168194206206206206🏆 150175243171252175175281284300240164309308291210
Scroll a window 2 tasks83155189189162177202210210210🏆 144168236164242170170271276293237156302301284203
Switch the viewed group 1 task84156194194164180203203203203🏆 146169236167242170170271279296237159302302287206
Filter on two axes 1 task79151183183157176198196196196🏆 140165233161241164164270273290229153298298281200
Related entities 2 tasks154·214214184145286273273273223210220🏆 137226140140267209191253165302213176169
One value 2 tasks43·929266🏆 5910214414414486728870686060857076666384997261

Where the rest is​

pagewhat it answers
Scenario tablesevery scenario's full table: time, ratio, drift, render counts, flags, who declined and why
Participantseach library's version, what it stores, and the audit of how idiomatically its adapter uses it
Scenarioswhat each script models, what decides it, and the audit of how representative it is
Methodologyverification, the anchor, the controls, the ranking rule, the load gate, staleness
Running the suitemeasuring in small pieces, a budget at a time, without giving the machine a day