A batch run skips the nodes waiting on a failed one and ends error; a run whose nodes cannot load fails with the load error before anything executes; cancelling a run stops its pipeline without announcing a pause of the live flow; batch_issues refuses a delay node's interval; the client accepts an engine URL with or without /api/v1; RedisState's compare-and-swap records the names it writes

This commit is contained in:
2026-11-04 21:38:21 +01:00
parent eb17528ce6
commit d028bf326a
11 changed files with 221 additions and 32 deletions
+18 -3
View File
@@ -147,6 +147,8 @@ class NodeOutcome(BaseModel):
artifacts: dict[str, dict[str, Any]] = {}
#: Restored from an earlier run rather than executed.
cached: bool = False
#: Not executed, because a node it waits on failed or was skipped itself.
skipped: bool = False
#: Which run the restored values came from — and, because a restored row
#: holds no series of its own, which run this node's metrics live in.
cached_from: str = ""
@@ -1071,7 +1073,8 @@ class Pipeline:
overrides: dict[str, Any] | None = None,
inputs: dict[str, Any] | None = None,
) -> dict[str, Any] | None:
"""Run one node and record its outputs. Never raises.
"""Run one node and record its outputs. Never raises: None means it
failed, and a node that published nothing answers ``{}``.
``inputs`` is what the readiness check already read — the same keys, a
moment earlier. Reading them again here was the second of two MGETs
@@ -1106,7 +1109,7 @@ class Pipeline:
if key:
hit, restored = self._from_cache(node, key, state, entry_id)
if hit:
return restored
return restored or {}
with logs.capture(collected):
result = node.execute(inputs)
@@ -1169,7 +1172,7 @@ class Pipeline:
output_values=result,
)
)
return result
return result or {}
except Exception as exc:
# One failing node must not take the rest of the graph down.
error = self.publish_error(node, exc, collected, entry_id)
@@ -1431,6 +1434,11 @@ class Pipeline:
ready_nodes = deque(n for n, count in in_degree.items() if count == 0)
blocked: set[Node] = set()
# A batch run starts from an empty state, so a node whose producer
# failed would be called without that input. It is skipped instead,
# and so is everything waiting on it. A live wave still has the last
# values to run on.
failed: set[Node] | None = set() if self.run_id else None
gate_revision = self._gate_revision
node_futures: dict[Future[dict[str, Any] | None], Node] = {}
finished: SimpleQueue[Future[dict[str, Any] | None]] = SimpleQueue()
@@ -1469,6 +1477,11 @@ class Pipeline:
if n.passive:
complete(n)
continue
if failed is not None and not failed.isdisjoint(deps[n]):
failed.add(n)
self._observe(NodeOutcome(node=n.id, ok=False, skipped=True))
complete(n)
continue
if fresh is not None and untouched(n):
complete(n)
continue
@@ -1497,6 +1510,8 @@ class Pipeline:
future = finished.get()
n = node_futures.pop(future)
result = future.result()
if result is None and failed is not None:
failed.add(n)
if result is not None and fresh is not None:
fresh.update(result)
# Silent, failed and throttled producers still release their