Host ports
Use a run handle to send workflow input and read workflow output. The scheduler,
run types, and OutputCursor are exported from duron. For a complete example,
see workflow and host ports.
Send and close inputs
| Operation | Behavior |
|---|---|
await run.port(Ports.items).send(value, idempotency_key=None) |
Durably enqueue input without waiting for the workflow to accept it. |
await run.port(Ports.items).close() |
Enqueue closure after earlier values. |
The scheduler delivers queued input to the local run owner in insertion order. A successful send confirms that the message was durably enqueued. Messages for closed inputs or terminal runs are discarded, and processing rejections are logged. An interrupted send may already have committed.
To retry a send, reuse its idempotency_key and encoded value. An identical
retry is deduplicated, including after input closure. Reusing the key with a
different value raises PostConflictError. Without a key, each send creates a
new message.
Use handles from the single active scheduler that owns the database file.
Ownership covers all namespaces; a second active scheduler on that file raises
StorageError. Sending input does not acquire or replace the run's write claim.
Read outputs
| Operation | Behavior |
|---|---|
run.port(Ports.events).reader(after=None, prefetch=64) |
Create an independent output reader, optionally resuming from a cursor. |
async for value in run.port(Ports.events) |
Iterate a fresh output reader. |
Readers drain committed values before ending. They do not propagate workflow
failures; await run.result() to observe the result or error. A blocked run or a
scheduling failure without a workflow outcome leaves outputs open, so a reader
may wait indefinitely after draining the available values.
prefetch must be a positive integer. It bounds reader buffering without
applying backpressure to the producer.
Resume from a cursor
Persist the reader's cursor alongside the consumer's committed work. Store both
its opaque stream identity and its offset. After reconnecting, reconstruct
an OutputCursor and pass it to run.port(Ports.events).reader(after=cursor).
A cursor belongs to one persisted run and output port. It remains valid across reconnects. The same run ID in another namespace or database identifies a different run and cannot use that cursor.
Reader and scheduler lifetime
Dropping a reader leaves execution running and workflow ports open. Closing the scheduler stops local execution. Open a new scheduler context, reopen the run, and recreate its readers to continue reading. If an unfinished effect blocks recovery, its committed output is still available.