## Summary
Pins musl thread stacks at 8 MiB on every link, and fixes a crash in
**native Alpine builds** (`make build`, `make test`, or `go test` on an
Alpine host).
**The released images are not affected.** An earlier version of this
description said public endpoints could crash the shipped image. That
was wrong, and review caught it: see "Who is affected".
## Root cause
- musl sizes every non-main thread's stack from the ELF `PT_GNU_STACK`
size, with a floor of 128 KiB and a cap of 8 MiB (glibc defaults to 8
MiB).
- cgo calls run on the calling thread's stack, and SQLite's
`multiSelect` recurses about 850 bytes per `UNION ALL` term.
- `lookupPrefixes` (up to 400 terms) and `sharedEightBytePrefixes` (up
to 450 terms) overflow a 128 KiB stack and SIGSEGV the whole process.
## Who is affected
| Link | `PT_GNU_STACK` without the flag | Thread stack | Wide-UNION
tests |
|---|---|---|---|
| Alpine gcc + GNU ld (`make build` / `make test` / `go test` on Alpine)
| `0` | 128 KiB | **SIGSEGV** |
| zig 0.16.0, x86_64-linux-musl (Dockerfile, `crossbuild`, release job)
| `0x1000000` (16 MiB) | 8 MiB | pass 3/3 |
| zig 0.16.0, aarch64-linux-musl | `0x1000000` (16 MiB) | 8 MiB | pass
3/3, run on an arm64 node |
So for the released images the flag pins a value that zig's default
already provides. It stops being a property of a linker default, and a
build check now fails if it is lost. For native Alpine links it is the
fix: with the flag those tests pass.
Control: a zig arm64 link with `-Wl,-z,stack-size=131072` gets 128 KiB
threads and the same tests SIGSEGV in cgo on the arm64 node.
## Fix
- Every musl link passes `-Wl,-z,stack-size=8388608`: the `Dockerfile`,
the `Makefile` (`build` and `test` on Linux, and `crossbuild`) and the
release job in `deploy.yml`.
- This reserves address space only. Memory is committed as pages are
touched.
- **The artifacts are checked.** The Dockerfile's builder stage and the
release job's "Verify release binaries" step read `PT_GNU_STACK` with
`readelf` and fail unless it is exactly 8 MiB. Verified: the builder
stage passes for `linux/amd64` and `linux/arm64`, and fails with
`--build-arg STACK_LDFLAG=` (`/corescope-server: PT_GNU_STACK is not 8
MiB`, memsz `0x1000000`).
## Testing
- `TestMuslThreadStackWideUnionQueries`
(`cmd/server/musl_thread_stack_linux_test.go`, Linux only) runs both
queries off the main thread and asserts exact results, so a failed chunk
cannot pass as "nothing matched". Under `golang:1.27-alpine` with gcc:
SIGSEGV without the flag, pass with it. On CI's glibc runners it passes
either way, which is why the artifact checks above exist.
- `tests/unit/test-issue-1956-release-routing.js`: the release verify
step must check both binaries, and a `0x000000` or `0x1000000` header
must fail the release.
- `tests/unit/test-dockerfile-buildplatform.js`: all three `go build`
lines carry `STACK_LDFLAG`, and the `readelf` check covers all three
binaries.
- `sh test-all.sh` passes. `go vet` passes for `cmd/server` on Linux
with gofmt clean.
- Not verified: a Windows `go vet`. The fix there is the
`_linux_test.go` file name; no Windows machine was available to run it.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---
## Merge order
These four PRs were split out of #2396 and are **independent — they can
merge in any order.** Verified: every pair is conflict-free, and all
four merged together pass `go vet`, `go test ./...` (all 16 modules) and
`sh test-all.sh`. (That combined run predates the review fixes on #2409
and #2410.)
Suggested order, none of it urgent:
1. **#2408** — never lose rows during startup / a failed poll.
2. **#2407** — explicit `observers.idx` primary key.
3. **#2409** — content-hash migration moved into the ingestor.
4. **#2410** — pin musl thread stacks at 8 MiB. Released images are not
affected; it fixes native Alpine builds.
After these land, #2396 (Postgres) will be rebased onto master.
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: efiten <erwin.fiten@gmail.com>