Key the stage cache by the node's input names, not the flow's

A cache key held qualified input names, so the same node reading the same
values through two flows keyed differently and only a node with no inputs
could ever hit across one. The fingerprint beside the key already says what
the node is, and it has been flow-agnostic since it moved ahead of
assign_flow — the names were the last thing tying an entry to one flow.

Inputs now reduce by the node's own name for them; a name belonging to
another flow keeps its prefix, since reading it is part of what the
execution is. Every stored entry misses once and is re-run.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-08-26 22:41:35 +02:00
co-authored by Claude Opus 5
parent 647644ebbd
commit 001ec7b282
2 changed files with 27 additions and 6 deletions
+11 -3
View File
@@ -151,16 +151,24 @@ class RunCacheLookup(Protocol):
"""What an equal execution produced, or None when there is no entry."""
def run_cache_key(fingerprint: str, inputs: dict[str, Any]) -> str:
def run_cache_key(fingerprint: str, inputs: dict[str, Any], flow: str = "") -> str:
"""What this node, with these inputs, is known by.
An artifact input counts as its digest: the reference carries a name and a
size beside it, and the same bytes under another name are the same input.
A value JSON cannot carry cannot be part of a key, and a node reading one
is simply not cacheable.
Inputs are keyed by the *node's* name for them, not the flow's: the same
node reading the same values through ``study`` and through ``quick`` did
the same work, and the fingerprint beside it already says what the node is.
A name belonging to some other flow keeps its prefix — reading
``other.metric`` is part of what makes this execution what it is.
"""
reduced = {
name: value["digest"] if is_reference(value) else value
(name.removeprefix(f"{flow}.") if flow else name): (
value["digest"] if is_reference(value) else value
)
for name, value in inputs.items()
}
try:
@@ -932,7 +940,7 @@ class Pipeline:
key = ""
if self.run_cache is not None and node.fingerprint:
key = run_cache_key(node.fingerprint, inputs)
key = run_cache_key(node.fingerprint, inputs, node.flow)
if key:
hit, restored = self._from_cache(node, key, state, entry_id)
if hit:
+16 -3
View File
@@ -406,7 +406,7 @@ def test_a_cache_hit_restores_the_outputs_without_running_the_node():
node, calls = counting_node()
state = MemoryState()
seen: list[NodeOutcome] = []
key = run_cache_key("fp-train", {"study.lr": 0.5})
key = run_cache_key("fp-train", {"study.lr": 0.5}, "study")
cache = FakeCache({key: {"study.loss": 99.0}})
Pipeline(nodes=[node], state=state, observer=seen.append, run_cache=cache).run(
@@ -428,7 +428,7 @@ def test_a_hit_from_another_flow_restores_under_this_flow_s_names():
node, calls = counting_node()
state = MemoryState()
seen: list[NodeOutcome] = []
key = run_cache_key("fp-train", {"study.lr": 0.5})
key = run_cache_key("fp-train", {"study.lr": 0.5}, "study")
# Recorded by a run of "other", which is what a node shared between two
# flows gets a hit from — the values are the same, the namespace is not.
cache = FakeCache({key: {"other.loss": 99.0}}, flow="other")
@@ -455,7 +455,7 @@ def test_a_miss_runs_the_node_and_carries_what_would_be_stored():
).run(seed_values(flow, {"lr": 0.5}))
assert calls == [1]
assert cache.asked == [run_cache_key("fp-train", {"study.lr": 0.5})]
assert cache.asked == [run_cache_key("fp-train", {"study.lr": 0.5}, "study")]
assert not seen[0].cached
assert seen[0].cache_key and seen[0].output_values == {"study.loss": 1.0}
@@ -467,6 +467,19 @@ def test_the_key_follows_the_inputs():
assert first == run_cache_key("fp", {"lr": 0.5})
def test_the_key_does_not_follow_the_flow_an_input_hangs_in():
# The same node reading the same value through two flows did the same work,
# so a run of one is a hit for the other.
assert run_cache_key("fp", {"study.lr": 0.5}, "study") == run_cache_key(
"fp", {"quick.lr": 0.5}, "quick"
)
# A name belonging to neither keeps its prefix: reading another flow's
# message is part of what this execution is.
assert run_cache_key("fp", {"other.lr": 0.5}, "study") != run_cache_key(
"fp", {"other.lr": 0.5}, "other"
)
def test_an_artifact_input_counts_as_its_digest():
digest = "sha256:" + "0" * 64
# The same bytes under another name, of a size recorded differently, are