Seven `test-go-pg` failures on `main` since 2026-08-31 share one
signature: unrelated tests in different packages expire their contexts
within the same few seconds while the process keeps logging normally.
Only filesystem writes and database round trips stall. This points at
the runner's EBS-backed root disk, not CPU.
- Mount an 8 GB tmpfs on GOTMPDIR for Linux jobs using the test-go-pg
action. It holds test binaries and per-test temp dirs. Windows and macOS
already use RAM-backed storage; Linux was the only platform on disk.
- Stop `log_statement=all` on the CI test Postgres container. Nothing in
the test jobs reads it, and with 64 concurrent tests it is the largest
continuous disk writer. The gen job keeps the default.
- Sample `vmstat` and `iostat` during the test step and upload on
failure, so the next occurrence can confirm or refute the disk theory.
Depot documents the 8-core Linux runner as an EBS volume at 8000 IOPS
and 250 MB/s behind an 8 GB in-memory write buffer, which fits a stall
that hits every writer at once.
Refs coder/internal#1365