Results, cancellation, and reruns

Calling .run() starts a workflow and returns an awaitable ava.RunHandle immediately:

run = document_flow().run(executor=ava.LocalExecutor())
print(run.run_id)
result = run.result()

# In async code:
result = await document_flow().run(executor=ava.LocalExecutor())

A workflow may return JSON-compatible values, Pydantic models, ava.File, or ava.Workspace, including those values nested in models and containers.

​

Embedded and operator results

In embedded execution, .result() returns the original Python value. In an operator-managed run, retrieve a successful result through the client or the CLI:

uv run ava result "$RUN_ID" \
  --connect localhost:7433 \
  --wait \
  --timeout 300 \
  --output-dir ./run-result

The output directory must not already exist. The command writes result metadata and downloaded file or workspace contents beneath it.

​

Cancellation

run.cancel()

Cancellation is cooperative. Avalanche stops submitting later nodes after a cancellation request, but an already running local function or Ray task is not forcibly interrupted.

​

Rerun a section of a workflow

Give durable rerun targets stable slugs:

@ava.step(slug="chunk-docs")
def chunk_documents(documents: list[str]) -> list[str]:
    return documents


result = document_flow().run(
    executor=ava.LocalExecutor(),
    run_id="rerun_002",
    rerun=ava.Rerun(run_id="run_001", start=["chunk-docs"], mode="autorun"),
).result()

mode="autorun" runs each selected node and everything downstream. mode="lazy" runs only the selected nodes. Supply input= again when a rerun needs workflow input; Avalanche does not restore it from the source run.

For data-backed incremental execution, see streams, cursors, and lineage.