fix(rds): connect_timeout on RdsServiceGateway so connects fail fast

[Kydoimos97]

Fixes

• Connection/pool construction no longer hangs forever on an unreachable or saturated database. RdsServiceGateway built its ThreadedConnectionPool (and the reconnect() pool rebuild and single-connection reconnect) with a bare DSN and no connect_timeout, so libpq waited indefinitely for each connect. A consumer whose worker builds the pool at startup could hang forever with no recovery. This is the library-level root cause behind Datadog BUG-96: a transient prod DB CPU saturation left AiAxis hung on pool construction and boot-looping (HTTP 502) for ~40 minutes.
• Adds a connect_timeout parameter (default 10s, floored to the libpq minimum of 2s) threaded through all three connection sites. Connects now fail fast so consumers can surface the error and self-heal.

Tests

• Updated the reconnect assertions to expect connect_timeout, and added coverage for default / custom / floored values on the pool constructor. Full tests/ suite green (159 passed), ruff clean.

Impact

Backward compatible — new optional parameter with a safe default; every existing consumer immediately gains fail-fast connects with no code change. Consumers that want a tighter bound (e.g. AiAxis) can pass a smaller value.

References

• Datadog case BUG-96. Consumed by: AiAxis bump (forthcoming), ai-pipeline bump (follow-up).

[Wrench-Review-Bot]

PR Review — APPROVE

Adds connect_timeout (default 10 s, floored to libpq minimum 2 s) to RdsServiceGateway and threads it through the three psycopg2 connection sites the class owns: pool construction, pool rebuild in reconnect(), and single-connection rebuild in reconnect(). Targeted, backward-compatible fix for BUG-96.

Nits

• Nit: WrenchCL/Connect/AwsClientHub.py:179 — The initial single-connection path (non-multithreaded __init__) goes through AwsClientHub._rds_handle_configuration → psycopg2.connect(host=..., port=..., ...) with no connect_timeout. A saturated DB would still hang the first connect in non-multithreaded mode. This is pre-existing ambient debt in AwsClientHub, not introduced here; the pool / reconnect paths (the BUG-96 scenario) are correctly fixed. Worth a follow-up PR on AwsClientHub._rds_handle_configuration.
• Nit: WrenchCL/Connect/RdsServiceGateway.py:39 — type annotation is int but the assignment max(2, int(connect_timeout)) silently accepts floats. Union[int, float] would be more accurate, or add a # type: ignore if the narrower hint is intentional.

Context

Diff: 2 files, +55/−5 · Prior reviews: none · Related: Datadog BUG-96; consumer bump PRs forthcoming in AiAxis and ai-pipeline per PR body

— Claude

[Wrench-Review-Bot]

PR Review — APPROVE

Adds connect_timeout (default 10 s, floored to libpq minimum 2 s) to RdsServiceGateway and threads it through the three psycopg2 connection sites the class owns: pool construction, pool rebuild in reconnect(), and single-connection rebuild in reconnect(). Targeted, backward-compatible fix for BUG-96.

Nits

• Nit: WrenchCL/Connect/AwsClientHub.py:179 — The initial single-connection path (non-multithreaded __init__) goes through AwsClientHub._rds_handle_configuration → psycopg2.connect(host=..., port=..., ...) with no connect_timeout. A saturated DB would still hang the first connect in non-multithreaded mode. This is pre-existing ambient debt in AwsClientHub, not introduced here; the pool / reconnect paths (the BUG-96 scenario) are correctly fixed. Worth a follow-up PR on AwsClientHub._rds_handle_configuration.
• Nit: WrenchCL/Connect/RdsServiceGateway.py:39 — type annotation is int but the assignment max(2, int(connect_timeout)) silently accepts floats. Union[int, float] would be more accurate, or add a # type: ignore if the narrower hint is intentional.

Context

Diff: 2 files, +55/−5 · Prior reviews: none · Related: Datadog BUG-96; consumer bump PRs forthcoming in AiAxis and ai-pipeline per PR body

— Claude

[Wrench-Review-Bot]

✅ PR Review complete — APPROVED. Full review ↗