perf: externalise @ioredis/commands, and identify what actually holds the memory

Tested the hypothesis that cached base64 images were bloating TanStack Query and
Redis. They are not, measured on a loaded 4-tab session:

  base64 data-URLs, server heap   0.01 MiB (2 strings)
  base64 blobs, browser heap      0.01 MiB (7 strings)
  Redis total                     2.17 MB
  image bytes in Redis            none - image-proxy stores URL + headers only

What the server heap is actually made of (141 MiB used): 41.4 MiB is bundled
JavaScript source retained as strings, ~29% of the heap. The memory is the code, not
cached data — which is why every cache-tuning idea in the fix catalog measured at
zero, and why serverExternalPackages was the only lever that moved idle memory.

7.86 MiB of that is literal duplicate copies: the same module inlined into several
Turbopack chunks and retained once per chunk. @ioredis/commands' command table was in
10 chunks and resident 6 times (2.07 MiB wasted), so it is now external — ioredis is
already external and owns it as a direct dependency, so it stays traced.

Recorded as bundle hygiene, NOT a memory win: idle 89.7 -> 89.4 MiB is inside noise,
and board-loaded/peak deltas were smaller than the run-to-run spread (one "before" run
read 220.6 MiB against 169-173 for identical code).

got and @octokit/request-error duplicate just as badly but were deliberately left
bundled: they are only transitive dependencies, and a check inside the image showed
got is not traced into node_modules, so externalising it would have turned a bundled
module into a require that resolves to nothing — a possible MODULE_NOT_FOUND for zero
measured gain.

analyze-heap-strings.mjs buckets a snapshot's strings by what they are, which is what
made "37 MiB of strings" actionable instead of just alarming.

Verified: 28/28 widgets render, auth control passes, zero module-not-found in the
container log, typecheck clean. pnpm test 1794/1795 with two different Testcontainers
flakes across runs (pi-hole, sabnzbd); both pass 18/18 in isolation.
This commit is contained in:
Thomas Camlong committed 2026-08-09 16:07:26 +02:00
1 parent 933a54ead9
commit f707f64333
3 files changed
+129

No files matched your search

+55
View File
@@ -301,8 +301,63 @@ is a correctness fix — idle commits stayed within their 73–96 range afterwar
the app-root churn dominates and this hook is not what drives it. Recorded as *not* a
measured win.
## What is actually hogging the memory
Tested directly against the hypothesis that cached base64 images were bloating
TanStack Query / Redis. **They are not** — measured on a loaded 4-tab session:
| suspect | measured | verdict |
| --- | --- | --- |
| base64 data-URLs in server heap | **0.01 MiB** (2 strings) | not a factor |
| base64 blobs in browser heap | **0.01 MiB** (7 strings) | not a factor |
| Redis total | **2.17 MB** | not a factor |
| image bytes in Redis | **none** — `image-proxy` stores only URL + headers | not a factor |
| React Query `gcTime` | already 5 min, not 24 h | already addressed |
What the server heap is actually made of (141 MiB used, 259 MiB RSS):
| bucket | size | share of heap |
| --- | --- | --- |
| **bundled JS source retained as strings** | **41.4 MiB** | **~29%** |
| other strings (translations, ids, misc) | 10.3 MiB | 7% |
| URLs | 2.0 MiB | 1% |
| everything else (objects, code, maps) | ~87 MiB | 62% |
**The memory is the code, not cached data.** That is why every cache-tuning idea in the
fix catalog measured at zero, and why `serverExternalPackages` — which removes code from
the boot-loaded bundle — was the only lever that moved idle memory.
Within that, **7.86 MiB is literal duplicate copies** of identical chunk sources: the same
module inlined into several Turbopack chunks and retained once per chunk. Worst offenders
were `@ioredis/commands`' command table (in 10 chunks, resident 6×, 2.07 MiB wasted) and
`got`/`@octokit/request-error` (2.1 MiB each).
`@ioredis/commands` is now external, which removes it from all 10 chunks. Recorded as
**bundle hygiene, not a memory win**: idle went 89.7 → 89.4 MiB, inside noise, and
board-loaded/peak differences were smaller than the run-to-run spread (one "before" run
read 220.6 MiB against 169–173 for identical code).
`got` and `@octokit/request-error` were deliberately **not** externalised. They are only
transitive dependencies here, and a check inside the image confirmed `got` is **not
traced** into `node_modules` — externalising it would have turned a bundled module into a
runtime `require` that resolves to nothing, trading zero measured gain for a possible
`MODULE_NOT_FOUND` on a code path the benchmarks never hit.
### Where the remaining headroom is
Nothing further is available from caches, Redis, or process trimming. The two real levers
left are both large:
1. **Ship less server JavaScript.** ~29% of the heap is retained bundle source. Every
package moved out of the boot graph pays twice — smaller chunks and less retained
source. This is the same lever that produced the −37% idle result.
2. **Stop the App Router re-rendering every second** (see above). That governs the
browser side, which is where the 650 MB peak lives.
## Measured but not pursued
| lever | measured cost | why not pursued |
| --- | --- | --- |
| in-image Redis (F12) | **8.1 MiB RSS** (6.3 PSS) | The catalog estimated 10–40 MB; the real figure is 8. Removing it means replacing the pub/sub bus for single-instance — real risk for ~5% of the footprint. |