## Summary
The standalone S3 accounting worker can exit before its scan begins
because database startup inherits the relay’s three-second connection
budget. A cold DNS probe took about five seconds.
Give `buzz-admin storage-snapshot` its own pool: one connection, no
minimum idle connections, and a 30-second acquisition budget. Retry
startup timeouts and transport failures at most twice, after two and
five seconds. Total startup wait is bounded to 97 seconds.
Configuration, authentication, TLS, and protocol errors fail
immediately.
Log each startup attempt’s elapsed time and error category without
connection credentials. Relay defaults stay unchanged. Retries stop
before lock acquisition; the same lock-owning database session still
publishes the completed snapshot.
### Related issue
Follow-up to #7543. No duplicate startup fix found.
### Testing
Archived-command check: built a nextest archive, temporarily moved
`target/debug/buzz-admin` out of the checkout, and ran
`scripts/postgres-test-run.sh --archive-file <archive>`. Both tests
failed before the path fix and passed afterward. The tests use nextest’s
relocated executable path, with the original Cargo path retained for
local runs.
Manually restored the old startup call and ran
`scripts/postgres-test-run.sh -p buzz-admin --test
postgres_storage_snapshot -E 'test(storage_snapshot_retries_then)'`. It
failed before reaching S3; restoring the fix passed. This confirms the
test exercises the actual worker command.
Local `just ci` passed all stages before mobile tests, which stopped at
SDK discovery in the `objective_c` native hook. The default Xcode tools
report an unaccepted license. Two existing PostgreSQL tests hit
connection timeouts in the parallel run and passed individually.
After merge, build a worker image and repeat the cold-pod staging smoke
test through snapshot publication before changing relay metrics mode.
Generated with Codex
---------
Signed-off-by: Ravneet Arora <rarora@squareup.com>
a929532a04