Skip to content

Document post-#55 jemalloc rejection #63

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
aksOps merged 4 commits into from
Jul 30, 2026
Merged

Document post-#55 jemalloc rejection #63

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
05bb6b5
Use jemalloc allocator on non-MSVC targets
aksOps Jul 29, 2026
62318ca
Reject jemalloc after measured RSS regression
aksOps Jul 29, 2026
274a1c6
Merge origin/main into Stream A
aksOps Jul 30, 2026
dde621f
Document post-#55 jemalloc rejection
aksOps Jul 30, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
41 changes: 41 additions & 0 deletions docs/decisions.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,6 +4,47 @@ Newest first. Each entry records what was decided, why, and what was rejected.

---

## 2026-07-30 — jemalloc is rejected on the post-#55 reference build

**Decided: keep the system allocator.** Stream A measured the non-MSVC
`tikv-jemallocator` 0.7.0 build (`05bb6b5`) against its parent
`035821a` on the read-only Kubernetes tree at
`7fbcad5e8ec387e23061602eb2d4b110b049410c` (17,873 Go files,
5,353,211 lines). Both release binaries ran three cold scans with one clean
store per run under `taskset -c 0,1` and `/usr/bin/time -v`; no cargo, rustc,
or arthron process was live before the batch.

| build | cold wall (s) | peak RSS (kB) | median wall / RSS |
|---|---|---|---|
| system allocator | 132.20, 135.66, 122.42 | 285,832, 285,324, 284,296 | 132.20 / 285,324 |
| jemalloc | 123.33, 115.65, 101.82 | 322,504, 313,964, 320,808 | 115.65 / 320,808 |

Jemalloc improved median wall time by 16.55 seconds (12.52%), but increased
median peak RSS by 35,484 kB (12.44%). The acceptance requires both metrics to
improve or hold, so the RSS regression rejects the allocator. All six stores
had the same SHA256,
`e9b6c3e4633e3aa99054d3b2a4e8c26e9cf7505ba4a1b30aa3c620cf31ad955b`;
the result is a resource regression, not a graph difference. Three Flask cold
runs also had higher jemalloc RSS.

This does not erase the earlier 793,548 kB / 63.6 s versus 832,740 kB /
70.6 s measurement below. That comparison was taken before #55 changed the
walk's retained-reference memory model, whereas this one is against the
post-#55 285 MB baseline. The measurements therefore do not describe the same
binary memory model; this record makes no unmeasured claim that their corpora
differed. The N-033 ruling requires Stream B to re-measure both allocators on
the pinned Node tree before its final RSS gate.

*Rejected: carrying the older jemalloc result forward as a current claim.* Its
code footing predates the structural change that changed the observed allocator
behavior. A past saving is evidence to re-measure, not permission to ignore a
current regression.

*Rejected: landing jemalloc for wall time alone.* The cold-RSS ceiling is hard;
the time target cannot trade it away.

---

## 2026-07-30 — Stream C's Java measurements are rebased with attributed aliases

**Decided: rebase the two Java baselines and target pins from the signed
Expand Down
Loading