Generate outputs for images and pipelines (Anthropic)

Enumerate constraints first: input format (jpg, png, multi-page tiff), model I/O (CPU vs GPU), expected cardinality (100k images), target SLA. Standard skeleton:

1. A globbing iterator that yields (path, stable_id) to preserve deterministic ordering even after sort.
2. A thread or process pool sized to min(cpu_count, model_concurrency).
3. A producer-consumer queue decoupling disk I/O from inference.
4. An output writer that flushes per shard to avoid data loss on crash.
5. An optional retry queue keyed by stable_id, not path (filenames can collide).

paths = sorted(glob("*.jpg"))
pool = ThreadPoolExecutor(max_workers=GPU_CONCURRENCY)
for path, result in zip(paths, pool.map(run_pipeline, paths)):
    write(stable_id(path), result)

Gotchas that separate strong from weak: blocking the event loop on numpy decode; hashing by path when timestamps change; holding the entire batch in memory; not handling intermittent CUDA OOM.