hono-bun: size the Postgres pool to the process count

[MDA2AV]

site/static/logs/crud/4096/hono-bun.log came out of #1242 at 54MB / 1.9M
lines, where every other log in that directory is under 11KB. 72,031 of those
lines are Postgres answering sorry, too many clients already.

It is not cosmetic: the published crud-4096 carries status_5xx=23308.
Every other JS entry publishes 0 there, and hono-bun's 218k rps trails
bun/node/deno at 310–347k. It cleared the 5% error gate, so it was published
anyway.

Cause

The pool was a fixed max: 8 per process. entrypoint.sh starts one process per
CPU, and the crud profile hands the container the cpuset 1-31,65-95 — so
nproc reports 62 on the benchmark box:

62 processes x max 8 = 496 connections
postgres         max_connections=256, minus superuser_reserved_connections

The comment above it claimed "64 cores × 8 = 512 to match aspnet-minimal's
Npgsql pool for fair cross-framework comparison", which is over the limit on its
own terms. Both the pool and the SQLite block predate #1242 (March/April); crud
was already in meta.json, so what that PR changed was only that the profile got
re-run and wrote this log.

Fix

The pool divides DATABASE_MAX_CONN by the process count, which entrypoint.sh
now exports, leaving headroom for superuser_reserved_connections.

Verified on a 32-CPU box, driving /crud/items/:id at 1024 connections:

	connections held	too many clients
before	256 asked (32 × 8)	yes on the bench box (62 × 8 = 496)
after	226 (32 × 7, +2)	0

Two other things from the same log

• No pool.on("error") handler. node-pg re-emits connection failures as
  'error' events; with nothing attached, each printed a full stack trace, which
  is what turned a connection problem into a 54MB artifact. Every sibling entry
  attaches one.
• Unguarded SQLite open — 186 SQLITE_CANTOPEN stack traces (62 processes ×
  3 attempts). The harness mounts /data/dataset.json and /data/static only,
  never /data/benchmark.db, so the file is absent on every benchmarked profile.
  Guarded with existsSync, matching express and ultimate-express.

Validation

62 passed, 0 failed.

Worth a separate look

hono-bun is not alone in publishing 5xx — 31 profile-runs across the results
set do, and it is not the worst. The top of that list:

php-fpm                baseline-h2-1024   5xx=1462506   rps=294864
symfony-spawn-franken  baseline-h2-256    5xx=381252    rps=28243
roadrunner             crud-4096          5xx=86954     rps=189233
vanilla-epoll          crud-4096          5xx=66317     rps=317372
vanilla-io_uring       crud-4096          5xx=63843     rps=338676

All under the 5% gate, all published as if clean.

🤖 Generated with Claude Code