2.0.0-rc7 | possible regression from rc0 with many stores single write and a flush

[mizulu]

Describe the bug

RC.0
https://mizulu.users.s.olid.uk/solid2-store-write-benchmark/r/2/

RC.7
https://mizulu.users.s.olid.uk/solid2-store-write-benchmark/r/1/

Your Example Website or App

.

Steps to Reproduce the Bug or Issue

.

Expected behavior

.

Screenshots or Videos

No response

Platform

• OS: [e.g. macOS, Windows, Linux]
• Browser: [e.g. Chrome, Safari, Firefox]
• Version: [e.g. 91.1]

Additional context

import {
  createStore,
  flush
} from "solid-js"

const STORE_COUNT = 100000
const WRITES_PER_STORE = 1

const stores = Array.from(
  { length: STORE_COUNT },
  () =>
    createStore({
      value: 0
    })
)

const start =
  performance.now()

for (
  let write = 0;
  write < WRITES_PER_STORE;
  write++
) {
  for (
    let i = 0;
    i < STORE_COUNT;
    i++
  ) {
    const [, setStore] =
      stores[i]

    setStore(store => {
      store.value =
        write + 1
    })
  }
}

flush()

console.log(
  {
    stores:
      STORE_COUNT,

    writesPerStore:
      WRITES_PER_STORE,

    totalWrites:
      STORE_COUNT *
      WRITES_PER_STORE,

    ms:
      (
        performance.now() -
        start
      ).toFixed(3)
  }
)

I am not sure how much of this benchmark is related to the fixes #3044
or the benchmarks from that issue, because I don't see this numbers in the reproductions in the 3044

[ryansolid]

Fixed on next in two parts:

• perf(signals): narrow store writes clone by spread; overlay only for wide owned backings (#3360 part 1) #3367 (0f144302): plain-data containers clone by spread instead of a descriptor walk, and the prototype overlay is reserved for wide (>32 keys) store-owned backings — the overlay's Object.create(v) was turning every narrow raw into a prototype, which is where most of the rc.0 → rc.7 regression came from. 629 → 340 ns per one-key write.
• perf(store): stamp owned backings with $OWNER instead of weak-collection registrations #3368 (4935c7dd): store-owned backings carry their owner under an internal enumerable symbol stamp instead of registering in two weak collections per draft; the identity-hash/ephemeron cost of those registrations on a fresh object was the remaining floor. 340 → ~178 ns.

Net on your reporter shape (100k one-key stores, prod build): rc.7 ~230 ms for the first write + flush → ~51 ms; steady-state single-key writes now sit below rc.0 (214 ns). CodSpeed on #3368 shows the two new write-floor benches at +47% / +29% with no regressions elsewhere; those benches pin the floor going forward.

Thanks for the precise reproduction — it made the two costs easy to separate.

— Claude via Cursor