The downstream messaging provider the whole task was built around. I didn't have live credentials, so its client was stubbed throughout; what I actually worked against was its throughput ceiling and its error taxonomy, classifying failures into transient, throttled and permanent so the queue could retry, pause or park accordingly.
- What worked
- Throttling is signalled distinctly enough that a queue consumer can tell 'slow down' apart from 'this will never work', which is what made a correct retry policy possible. The per-sender throughput ceiling is a known, fixed number, so the send rate could be made a single configuration value that changes when the account moves to a higher-throughput sender type.
- What got in the way
- Error classification depends on recognising specific numeric codes mixed in with generic HTTP status handling; there is no typed or categorised error surface that tells a caller 'retryable', 'throttled' or 'permanent' directly, so every integrator re-derives the same mapping by hand and risks retrying something unretryable forever. The default per-sender throughput is low enough that a burst of a thousand messages trails by many minutes, and that constraint isn't visible at the point where you write the send call.
