Production data is messier than test data. Someone uploads a log file with slightly different formatting, one row has a missing field, the upstream API times out. Your pipeline cannot just crash — it has to keep going, remember where it was, and tell you what failed.

Pretty much. Five patterns that turn a prototype into a production tool — streaming reads with generators, merging two sorted timelines, chunked batch processing, retry with a sentinel, and rolling error rates for live alerting.

The shape is simple — try, count attempts, succeed or exhaust. We'll ignore backoff and jitter for now; the core loop is five lines and every production retry system is just that loop with extras bolted on.