Store rewrite: single-home storage model replaces the legacy store - #3019
Merged
Conversation
🦋 Changeset detectedLatest commit: 07d06df The changes in this PR will be included in the next version bump. This PR includes changesets to release 9 packages
Not sure what this means? Click here to learn what changesets are. Click here if you're a maintainer who wants to add another changeset to this PR |
size-limit report 📦
|
Coverage Report for CI Build 32223705775Warning No base build found for commit Coverage: 72.158%Details
Uncovered ChangesNo uncovered changes found. Coverage RegressionsRequires a base build to compare against. How to fix this → Coverage Stats
💛 - Coveralls |
Merging this PR will regress 2 benchmarks
|
| Benchmark | BASE |
HEAD |
Efficiency | |
|---|---|---|---|---|
| ❌ | reconcile: deep tree, single deep() effect |
79.9 ms | 134.6 ms | -40.63% |
| ❌ | omit |
234.6 µs | 308.4 µs | -23.94% |
| ⚡ | hasAllowed |
54.5 µs | 24.6 µs | ×2.2 |
| ⚡ | readAllowed |
32.4 µs | 24.7 µs | +30.89% |
| ⚡ | readAllowed |
32.8 µs | 25.5 µs | +28.59% |
| ⚡ | readAllowed |
32.3 µs | 25.2 µs | +27.88% |
| ⚡ | dbmon partial tick — deep reconcile |
793.8 ms | 626.7 ms | +26.66% |
| ⚡ | dbmon full tick — deep reconcile |
766.9 ms | 605.6 ms | +26.62% |
| ⚡ | read |
34.5 µs | 27.4 µs | +25.92% |
| ⚡ | read |
34.1 µs | 27.3 µs | +24.71% |
| ⚡ | read |
36.8 µs | 29.7 µs | +23.82% |
| ⚡ | readAllowed |
34 µs | 27.7 µs | +22.61% |
| ⚡ | reconcile: deep tree, all ~12k paths subscribed |
80 ms | 65.8 ms | +21.64% |
| ⚡ | construct |
39.7 µs | 33.7 µs | +17.87% |
| ⚡ | construct |
40.1 µs | 34.3 µs | +17.07% |
| ⚡ | hasAllowed |
26.7 µs | 23 µs | +16.01% |
| ⚡ | construct |
41.7 µs | 36.6 µs | +13.92% |
| ⚡ | construct |
62.7 µs | 55.2 µs | +13.54% |
| ⚡ | construct |
51.8 µs | 45.9 µs | +12.86% |
| ⚡ | reconcile: read-once untracked tree, 10 of ~12k paths subscribed |
7 ms | 6.3 ms | +12.23% |
| ... | ... | ... | ... | ... |
ℹ️ Only the first 20 benchmarks are displayed. Go to the app to view all benchmarks.
Tip
Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.
Comparing store-rewrite (8190d55) with next (2a49adc)
…ship, adoption-channel reconcile Plain deep stores and reconcile now serve from src/store/next/ per the INTERNALS-STORE-STATE.md contract (mined from 217 suite-derived rules, all design rulings closed): one committed home (owned-raw backing, CoW privatization with path copying — user sources never mutated, pinned by a __TEST__ oracle), lazy per-property nodes as literal core signals, and a per-target pending backing as the draft surface. Write channel: drafts mutate the pending clone natively (array methods just work); setter exit diffs pb vs committed and setSignals observed keys — write-time notification with commit deferred to node commit, so transition holds, per-leaf isPending, and lanes ride core machinery natively (a plain store write held by downstream async now pends exactly the touched leaves). Reconcile is the adoption channel: eager backing swap per target, keyed/positional matching with mode-dependent reachability pruning (keyed detaches unobserved captures; positional preserves slot identity), kind changes detach, and the identity skip carries its completed proof (incoming === backing && !owned) — fixing FINDING-1's unsound same-reference skip, whose rule test flips to plain `it` here. Snapshot resolves per-object through the registration (multi-parent/DAG writes visible through any path). Core gains two treeshakeable seams: a store commit hook in the scheduler (folds node values into backing at pending commit) and devGuardStoreSetterWrite (owned-scope write guard moved to setter entry). Transitional dispatchers route derived/shallow/optimistic forms to the legacy implementation until their increments land. Suite: 1227 passing / 26 failing, every failure in a named unported bucket (shallow/markRaw, affects-on-stores, projection interop, dev diagnostics, 2 ruled internals). Hot-path gate suites clean. dbmon: mount/remount/sort beat legacy; tick profiled 23.6→14.4ms/tick across two optimization rounds (subscription-keys-only fold diff, scan-once accessor detection). Includes the design doc, rules-mining catalogs, findings log (3 shipped bugs documented, 1 fixed here), rule-derived tests, the vite.next gate config, and the store-size/scan-hangs scripts. Co-authored-by: Cursor <cursoragent@cursor.com>
The rewrite now passes the entire suite (91/91 files, 1252 tests; 1 expected fail owned by the future optimistic increment, 2 ruled skips of legacy-internals pins). Landed in this increment: - Write path finalized per §3: drafts mutate the pending backing; setter exit diffs and setSignals observed keys (write-time notification), so transition holds, per-leaf isPending, and lanes ride core machinery — a plain store write held by downstream async pends exactly the touched leaves. Pending visibility mirrors core read()'s owner-context rule (Root→parentComputed, #2687), with the owned-scope write guard moved to setter entry (devGuardStoreSetterWrite). - Legacy interop via structural field aliasing: next targets expose the legacy field names (v/n/h/d/s intentionally shared; px/u renames avoid the x/p collisions) plus a $PROXY field, so shared machinery — unwrap, affects scope walks, witnessAffectsMark, snapshot resolution through lookupTarget — reads rewrite targets natively. Cleared the projection snapshot and affects buckets wholesale. - markRaw/shallow interop: raw values never wrap (leaves for reconcile), legacy-tracked raws serve their legacy proxy (cross-implementation dedupe), shallow's never-both-wrapped-and-raw dev check sees next targets. Shallow is NOT a port target (Ryan: perf-only, retirement candidate — the dbmon shallow column is its retirement bar). - Snapshot: owned (written) subtrees copy — non-enumerable symbols excluded, copies registered before descent — fixing FINDING-3 (cycle identity on written cyclic objects); unowned subtrees keep identity. - Dev diagnostics: registerGraph on creation, onStoreNodeUpdate full-key diff (dev-only), strictRead untracked-read warnings. Co-authored-by: Cursor <cursoragent@cursor.com>
…ive createProjection bring-up Families (§7b): projection children wrap into a per-projection registry (writes land in the projection, never a source family) and every family node carries the projection computed as its firewall — reads link the derive's status/lifecycle natively. storeSetterNext extracted as the low-level write primitive (no wholesale untrack: the writing flag already prevents draft self-tracking while external signal reads keep tracking — they ARE the derive's dependencies). Write-override interop honors the legacy draft flag so post-await async draft writes land with per-op notification. §6c status gate: an uninitialized async derive throws NotReady from every trap, probed untracked so the gate never coarsens isolation. Replace-mode reconcile: projection roots merge entity changes in place (root proxy stable for life, children not key-matched across the entity change) and displaced raws unregister. Two correctness fixes surfaced by bring-up, benefiting all stores: - notification diff old-side: t.v lags when a recompute runs before the prior fold commits (effects run before commitPendingNodes) — swallowed writes; the node's own current value is the true old side (setSignal's internal equality), pre-compare dropped. - logical-slot node equality: privatization/adoption change raw identity without changing the logical value; node equals now resolves both sides through the (family-aware) registration, eliminating phantom notifies on container writes after child privatization. Default suite green (91/91, 1252 passing). Next-native projections pass basics/selection (27/48 across the projection suites via the gate config); the default build keeps routing createProjection to legacy until async/chained bring-up completes. Co-authored-by: Cursor <cursoragent@cursor.com>
…gate) Chained backing works end to end (§7b): a derive returning a live store proxy adopts the PROXY as backing — reads flow through the inner store's traps (consumers dual-link inner graph + outer swap nodes; outer node values never shadow the chain), switching severs via dynamic dependency rebuild, and snapshot resolves chained backings to the base raw through a multi-level registration loop. Projection draft writes commit eagerly (authoritative-landing semantics, matching legacy overlay resolution). Entity changes swap wholesale (children never matched across an entity change); same-entity keyed merges preserve child identity UNCONDITIONALLY in projection mode while plain keyed reconcile keeps the R18 detach of unobserved captures — identity (re-pointing) and notification depth (d-gate) are now separate concerns, mode-split per the pinned contracts. $REFRESH resolves the projection computed from the trap (refresh()/ isPending). Root propagation bug fixed (replace flag now reaches the descent). Default suite green throughout (91/91). Gate: projection suites 49/58 (from 27/48) — remaining: the isPending transition cluster, generator-yield identity policy, Loading-boundary settle, and two sync stragglers. Co-authored-by: Cursor <cursoragent@cursor.com>
Projection port complete on the single-home model. Firewall gate links tracked readers (NotReady settle wake-up, deps drop post-settle so isolation holds); per-store write scopes extended through draft reads (legacy Writing semantics — chained projections track dependencies, cross-store draft writes work); CAS parent-slot fix on fold (draft splices can't resurrect rows via stale wrap-time indices). RUL-12 unkeyed-merge ruling executed on the two async yield-identity tests. Gates: projection suites 58/58 on next config, default 91/91 files, next-gate store sweep 17/17 files (361 tests). Co-authored-by: Cursor <cursoragent@cursor.com>
… suite) Armed nodes ride the engine wholesale — no store-side layer/backup/owner machinery. Landed: optimistic write channel (visible-view diff, committed raw untouched), membership overlay on armed has-nodes, optimistic-view pb seeding + draft reads (#2951 compose), firewall-transition entanglement + next-shape transitionBlocked half (#2951 hold), landing consumption split (structural consumes / owned values persist — #2719 + rapid-toggle both hold), lane-view reconcile baseline (§6b — optimistic rows recycle), §7b chained-gate pierce for overrides, snapshot/deep optimistic-view compose. Core fixes: createWriteTraps now save/restores projectionWriteActive (was hard-reset per op); untracked node-first reads serve the backing for committed state (O6 read-through). Remaining: one cross-test zombie cascade (documented in INTERNALS decision log) + affects/status clusters. Co-authored-by: Cursor <cursoragent@cursor.com>
The affects machinery reads next targets structurally (field names alias by design) — only node creation dispatches, via injected factories. lookupTarget handles next family maps (targets, not proxies); the declaration walk reads pending backings and composes the optimistic view (injected resolver) so in-motion rows are covered; next nodes inherit live marks at birth and survive unobserve while marked. Fixed "authority wins at reveal": ensurePB seeds the optimistic view ONLY for user drafts — authoritative landings seed committed truth (the override was folding into the committed home on multi-key post-await landings). affects-propagation and question-scoped-pending suites fully green. Co-authored-by: Cursor <cursoragent@cursor.com>
User-context reconcile on optimistic families diffs against the lane view and emits engine overrides per level: flat slot/membership/length overrides ride armed nodes (reverting with their transaction), while key-matched wrappable pairs descend into existing child targets so captured-proxy identity holds. Committed raw is never touched. The FINDING-2 rule test (key added by in-window reconcile reverts at settle) flips from test.fails to passing; reconcile-captured-proxies suite green. Co-authored-by: Cursor <cursoragent@cursor.com>
…ixed Strict-read refetch escalation (PENDING_ASYNC_UNTRACKED_READ on pending firewalls), witnessAffectsMark resolves chained backings, multi-level optimistic-view snapshot composition, presence-override structural classification for landing consumption (post-landing adds consume correctly), projection folds never eager (downstream holds form later in the flush; foldHeld + pending-backing visibility give context-free freshness without breaking hold staleness), ownedWrite store nodes + root-exempt setter guard, live-override-gated transitionBlocked half. Default suite: 1251/1255 — remaining 4 are one in-file zombie cascade (all pass in isolation; unhandled-rejection leak documented in the decision log as next cycle's first item). Co-authored-by: Cursor <cursoragent@cursor.com>
…ixes 91/91 files, 1253 passed, zero unhandled errors; next-gate 362 passed. (1) §6 length-as-view on optimistic arrays: length serves the composed view (backing ± overrides), node kept for tracking only — the torn iteration class (length on the stale rail, indices on pending) is impossible by construction. (2) Landing consumption performs the full legacy node reset (fold committed into _value, clear pending) — a parked transaction's stashed queues stranded subscriber wakes otherwise. (3) $TRACK on chained backings chains to the inner key-set node (§7b structural chaining, #2864): mapArray over wrapper views observes source structural notifications. Co-authored-by: Cursor <cursoragent@cursor.com>
Plain-store adoption notifies inline after the reconcile descent (the reconcile diff IS the fold diff — no foldOlds queue/drain round trip); projections keep deferred folds for downstream-hold semantics. Suite stays green (1253). Type fixes restore `pnpm build`: overload placement in legacy optimistic.ts, Computed<void> shape in next/optimistic.ts, ownKeys cast. Perf checkpoint + profile attribution recorded in the decision log (single-sweep numbers, ABBA required before claims). Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
…ction scanAccessors (full descriptor enumeration per object on first read — 115.7ms self in the uibench profile, the largest store cost) is deleted. Accessor-ness is a per-node creation-time flag via allocation-free own-gated __lookupGetter__/__lookupSetter__ probes; accessor keys serve through Reflect.get with the proxy receiver (R20-correct); folds use the cached flag + one own-gated incoming-side probe (merge-installed getters pinned by suite; prototype getters keep invoke-compare, also pinned). createTarget assigns fields directly on a shared hidden-class chain. uibench: legacy 36.6 → next 27.5 (25% faster; keyed ops 0.67-0.85x). dbmon tick: 14.35 → 11.50ms — legacy parity. Suite green (1253). Co-authored-by: Cursor <cursoragent@cursor.com>
Cached chained-backing flag (no per-read symbol probe), node lookup threaded into serveDataKey (was doubled), interned-string pollution guard, fam-first gate ordering, descend consults legacy interop only on next-lookup miss, object-branch adoption iterates for-in + conditional symbol pass. Suite green (1253). ABBA rounds pin the remaining dbmon delta at ~1.3x with a single structural cause documented in the decision log: the adoption channel's second diff (notifyFold re-walk) vs legacy's fused applyStateFast — the fuse is the next scoped item. Co-authored-by: Cursor <cursoragent@cursor.com>
Size script gains next-only entries (dispatchers bypassed) and a FLOOR variant with legacy modules stubbed. Full-featured next store (plain + projections + optimistic + reconcile + snapshot/deep) floors at 12.8kb gz total = ~4.4kb store attribution vs shipped's 15.5kb — 3.5x smaller with capabilities 1.9 never had. As-is next-only (19.5kb) shows interop imports drag legacy wholesale: the gap is the deletion worklist (port derived createStore form, move createWriteTraps, then delete legacy). Co-authored-by: Cursor <cursoragent@cursor.com>
createStore(fn, seed) routes to projection internals with a recompute-masking setter (core R31). The §6c gate covers errored derives and reduces to the raw-fallthrough guard: tracked reads link through firewall-backed nodes (core read links-then-throws — the node link wakes async-memo readers on landings, legacy parity; a pre-linking gate broke Loading boundaries and was discarded). Has-/key-set nodes carry the firewall arg. Eager folds split by channel: sync-derive drafts defer, post-await landings commit immediately — both spec-async pins hold. Suite green (1253). All non-shallow paths now serve from the rewrite. Co-authored-by: Cursor <cursoragent@cursor.com>
applyAdopt notifies inline per key in both object and array branches (descend-then-notify ordering preserves R9 slot identity silence); deleted keys via counted fast-out on a live node count; shared notifyKeyDiff/notifyFoldTail helpers; direct value writes (no closure per changed key). notifyFold remains for deferred projection folds only. Suite + next-gate green. dbmon next/legacy ≈ 1.2-1.3x (from ~1.35); cool-machine ABBA owed before the shallow ruling — inputs recorded in the decision log. Co-authored-by: Cursor <cursoragent@cursor.com>
…nnel-split probes Tracked node reads in serveDataKey use read()'s hoisted plain-signal fast path (legacy trap parity); the fused walk pre-checks wrappability inline so scalar keys skip the descend call; accessor probes are setter-channel only (reconcile adopts immutable data by contract, R2a — all pinned getter-preservation tests are setter-channel; suite green confirms). Suite 1253 green. Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
… bench Per-key notify body inlined into the fused walk (no helper call per key); reference-identity early-continue with the ownership guard (the sound form of legacy's FINDING-1 skip); values fetched once per side. New in-process A/B bench interleaves next and legacy in one vitest run — full-tick parity confirmed (0.98x mean), partial residual ~7-10%. Suite green (1253). Journal updated. Co-authored-by: Cursor <cursoragent@cursor.com>
Default vitest defines __TEST__ true: next pays per-adoption invariant oracles legacy lacks — prior A/B overstated next. In benchmark mode full tick oscillates around parity (0.81-1.06x across runs), partial 1.0-1.14x; cross-run variance now exceeds effect size. Measurement protocol recorded: benchmark mode only, cool machine for the verdict. Co-authored-by: Cursor <cursoragent@cursor.com>
Symbol compares gated behind one typeof on the hot string path; common serve case inlined (readNodeFast, no helper frame, no FORCE compare for plain nodes, primitive bail before isWrappable). Keyed arrays walk a positional prefix in place (identity skip inline with the FINDING-1 ownership guard) and only build prevByKey for a misaligned remainder — aligned polling ticks never allocate the map. Browser dbmon partial closed from an order-robust 1.24x to 0.97-1.03x; full tick in the parity band. Suite + next-gate green. Sweep-order bias documented. Co-authored-by: Cursor <cursoragent@cursor.com>
Stage 1 validated deep cannot reach shallow-class tick perf (phase-2 edit-script question; retirement may reopen there) — shallow stays and ports: t.s targets serve children verbatim (proxies by reference, #2932), sticky raw-marking at creation/set/adoption (shared R41 invariant), slot-granular positional reconcile with descends gated off. Dispatcher routes plain shallow to next. Suite green (1253) incl the #2932 suite; in-process A/B parity with legacy shallow (15.88 vs 15.48ms mean, p75 favors next). Remaining for wholesale deletion: fam.shallow forms, createWriteTraps move, legacy+dispatch delete. Co-authored-by: Cursor <cursoragent@cursor.com>
…plete fam.shallow wires shallow projections and optimistic stores to the same t.s slot semantics; the serve rule matches #2932 exactly (raw-marked data verbatim, store-proxy slots get boundary wrappers in the shallow family for chain write isolation). Dispatcher shallow routes removed. Suite green (1253) — every public store form now serves from the rewrite; legacy is unreachable except shared utilities and createWriteTraps. Deletion is next. Co-authored-by: Cursor <cursoragent@cursor.com>
The rewrite now serves every public store form, so legacy goes: reconcile/projection/optimistic modules removed, store.ts gutted to shared machinery (~1600 → ~570 lines: symbols, raw-marking, isWrappable, write-override, next-only affects walk, StoreNode as a structural view of StoreNextTarget), utils.ts to merge/omit, the transitional dispatchers and test shims removed — store/index exports next directly. createWriteTraps lives in next/projection; next's optimistic hooks are self-contained (engine + own clear + blocked half). Orphan sweep: storeLookup, symbolKeyedRecords (+ skipped legacy-pinned test), registerTransientStoreNode, mergedOverlay. Suite 1253 green; full monorepo build + tests green; in-process dbmon bench in family. Size vs main (gzip): full 24162 → 23259 (−3.7%), store attribution 15572 → 14803. Co-authored-by: Cursor <cursoragent@cursor.com>
reconcile and setter writes store function leaves by identity, never invoking them — the rewrite is immune by construction (every node write is thunk-wrapped through setSignal). Optimistic case pins that function leaves behave identically to scalars: draft serves by identity, never invoked, ambient write reverts at settle. Tests only — no changeset. Co-authored-by: Cursor <cursoragent@cursor.com>
Interleaved thermal-fair A/B (sign test over alternating samples) exposed a +1.4ms (~10%) dbmon tick regression that block-ordered and dev-mode benches masked: raw-as-truth nodes serve raw values, so every tracked object read paid wrapNext's WeakMap lookup + isWrappable where legacy nodes stored pre-wrapped values. Nodes now cache the served proxy + the raw it wrapped (px/pxv); a pointer compare replaces both costs, a replaced child fails the compare and re-wraps. Measured: tick +1.4ms → +0.3-0.5ms, tickPartial parity, sort now WINS (68/23). Negative result kept as a comment: notifyKeyValue's pre-compare is load-bearing (setSignal parks pending before commit-time equality, RUL-1) — removing it made identity-preserved adopted containers pay full write machinery (93/5 partial-tick regression, reverted). Co-authored-by: Cursor <cursoragent@cursor.com>
descend resolves the tracked target FIRST: a lookup hit implies the old side was wrappable and never raw-marked (only wrappables acquire targets), deleting per-pair isWrappable/isRawValue on that side; the new side still validates fully (frozen/platform/markRaw'd values stay leaves, R42). The keyed walk's prefix/remainder alignment checks are routing heuristics — both routes notify identically — so they drop to typeof gates with descend as the one validator. isWrappable left the hot profile. Interleaved n=100 ×3: tick +0.0/+0.3/+0.4ms (statistical parity), partial parity, sort next-wins every run; block sweep tick median next-faster (12.3 vs 12.8, best 10.4 vs 11.5). Suite green. Co-authored-by: Cursor <cursoragent@cursor.com>
Moves the optimistic-only machinery (notifyOptimisticWrites, consumeOverridesNext, optimisticView, applyTentative) out of the plain store/reconcile modules into next/optimistic.ts behind an injection table installed in createOptimisticStore's once-guard (no module-scope side effects — the module shakes). Sound by construction: every call site is fam?.opt-gated and optimistic families are only mintable by createOptimisticStore. Dead-export sweep rides along. Trade-off ruled (Ryan): plain store entry −481 gzip (13710 → 13229), full-import bundle +181 (headline eaten); package remains −722 vs main. Perf guards: dbmon tick next-wins 41/16, optimistic write-storm 0.94x, suite + full monorepo green. Co-authored-by: Cursor <cursoragent@cursor.com>
Rebase onto next brought effect-phase-writes coverage: a store write inside onSettled must park (immediate read returns the settled value, signal parity). Core's read() gained the rule in #3006 — CHILDREN_ FORBIDDEN execution scopes get committed visibility — but the store's inOwnerContext() mirror predated it. Now checks the resolved owner's config flag exactly like core. Suite green (1282). Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
…es (same public contract, new implementation) Co-authored-by: Cursor <cursoragent@cursor.com>
+createStore 13.5 -> 12.2 KB (measured 11.98 — legacy deletion + tree-shakeable optimistic channel took −1.26 KB out of the scenario); hydrating+stores 23.65 -> 23.3 KB (measured 22.84). Both at the ~2% headroom convention the config documents. Tooling only — no changeset. Co-authored-by: Cursor <cursoragent@cursor.com>
ryansolid
force-pushed
the
store-rewrite
branch
from
August 19, 2026 05:07
9bba377 to
8190d55
Compare
CodSpeed flagged 2 regressions on the PR. omit-mixed(100,25) reproduced at parity locally (0.6%) — CodSpeed environment noise per its own runtime-environment warning. The deep() one was real (+43% local): deepNext created/read a node PER PATH on every effect re-run where legacy subscribed one $TRACK per record. Targets now carry a lazy deep-witness node (dk): deep() reads key-set + witness per record and walks targets directly (no per-child proxy round-trip); the adoption walk, setter notify, fold, and optimistic channels bump dk only when it exists — value changes on un-noded keys included, in-place adoptions excluded (targetsEqual after descend). affects() declaration walks mark k/dk like property nodes and dk inherits live marks at birth, keeping isPending-over-deep() probes working (3 optimistic suite pins). Bench: 21.7ms → 15.18ms vs next's 15.24ms — parity; the sibling rows stay branch-favored (+21%/+9%); dbmon in family. Suite 1282 green; full monorepo build + tests green. Co-authored-by: Cursor <cursoragent@cursor.com>
brenelz
added a commit
to brenelz/solid
that referenced
this pull request
Aug 19, 2026
The store rewrite (solidjs#3019) replaced the legacy store with the store/next implementation; the attribution-gated store node naming moves from the legacy getNode to store/next getNode (same semantics: name only when the engine is installed, one null check otherwise). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Ground-up rewrite of the store subsystem on a single-home storage model: raw objects are committed truth (CoW-owned, privatize-on-first-write), per-property nodes are real core signals created lazily on subscription, writes park in copy-on-write pending backings that fold at flush, and transactions/optimism ride core lanes — the store-side transactional subsystem (override layers, layer clears, duplicate entanglement) is deleted wholesale. Design + decision log in
packages/solid-signals/INTERNALS-STORE-STATE.md; semantics were mined from ~17k lines of existing store tests into 217 pinned rules before implementation (rules-mining/).The legacy implementation is fully deleted (−3k lines): every public form — plain, shallow, derived, projections, optimistic — serves from the rewrite, with
store.tsreduced to shared machinery and no transitional dispatchers.Performance (vs shipped
next, prod builds)Size
next(bundlephobia-style, 22,845 vs 23,367)createOptimisticStore; all call sitesfam?.opt-gated)Fixes shipped bugs
snapshotbreaking cycle identity on written cyclic objects (FINDING-3)Test plan
Made with Cursor