Exploration: TrackedArray #storages Map → Array

[johanrd]

Exploration PR — change from NVP's emberjs/ember.js#21221.

Replace TrackedArray's #storages: Map<number, Tag> with #storages: Array<Tag | undefined>. Deliberately without NVP's INDEX_STORAGE_THRESHOLD = 10000 cap — that would silently drop granular invalidation past index 10k.

Isolated microbench (bin/tracked-array-storages.bench.mjs)

pattern	n	Map	Array	delta
sequential read	1000	25.06 µs	4.71 µs	−81 %
sequential read	5000	232.98 µs	24.34 µs	−90 %
write-read cycle	1000	41.33 µs	7.90 µs	−81 %
sparse hole-then-fill	1000	24.61 µs	3.48 µs	−86 %
sparse random access	1000	27.33 µs	3.92 µs	−86 %

Krausest tracerbench (80-fid)

The microbench win does NOT translate to Krausest.

phase	Δms	Δ%
duration	+25	+1.20 % (regression)
render1000Items1End	+1	+1.47 %
render5000Items2End	+10	+1.54 %
clearManyItems2End	+1	+1.92 %
render1000Items3End	+3	+6.91 %
no phases improved

Saved to .bench/c-storages-array-80fid.json.

Why the divergence

The microbench simulates the data-structure swap in isolation using simple objects. In the real TrackedArray:

• #storages is accessed through a private class field, which adds a slot lookup V8 handles uniformly for both Map and Array values
• #dirtyCollection() sets this.#storages.length = 0 on every array mutation — this empties the array but keeps it allocated. Subsequent sequential re-fills trigger ElementsKind transitions that Map doesn't have to pay
• The TrackedArray Proxy's get trap calls convertToInt → #readStorageFor → consumeTag per numeric access — the inner storage lookup is a smaller fraction of the trap total than the microbench measured in isolation
• V8 may already have tuned Map for small-integer-keyed monomorphic access in the glimmer runtime's specific call patterns

Verdict

Closing. Microbench says Array should be faster; real-world Krausest says it's a regression. The isolated mechanism doesn't survive contact with the surrounding code's access pattern and V8's actual optimizations.

Good reminder that microbench speedups don't guarantee production speedups — the same lesson in reverse of #6/#7 where microbench showed huge wins that didn't register on Krausest either.

[github-actions]

📊 Package size report   0.01%↑

File	Before (Size / Brotli)	After (Size / Brotli)
dist/dev/packages/@glimmer/validator/index.js	40 kB / 8.4 kB	0.9%↑40.3 kB / 2%↑8.6 kB
dist/prod/packages/@glimmer/validator/index.js	31.8 kB / 6.4 kB	1%↑32.1 kB / 3%↑6.6 kB
Total (Includes all files)	5.4 MB / 1.3 MB	0.01%↑5.4 MB / 0.03%↑1.3 MB
Tarball size	1.2 MB	0.03%↑1.2 MB

_{🤖 This report was automatically generated by pkg-size-action}