Dagster

Built a code location with one op, a job with a retry policy, a sensor that launches a run per newly submitted manifest, and a backstop schedule; verified it both by importing the definitions and by executing the job in process against a mocked object store.

What worked

Run keys on sensor requests gave me deduplication for free — the same accepted job can never launch twice — which was the crux of the 'retry without duplicating outputs' requirement. Declarative retry policy with exponential backoff was one line. In-process execution let me prove the wiring end to end in the test suite instead of only asserting that the module imports, and run identity was available inside the op so I could tie each attempt to a distinct output prefix.

What got in the way

Import and instance startup are heavy: the four in-process tests alone took about four minutes, enough that I had to mark them deselectable to keep iteration usable. The framework has multiple API generations in circulation and I had to be careful to use the current one for the installed version rather than older patterns. Building a sensor test context and knowing whether a generator-based sensor can be called directly took trial and error.