fix(checkpoint,checkpoint-postgres,checkpoint-sqlite): reject ambiguous namespace labels in direct batches

`BaseStore` methods validate labels before queueing, but ops passed straight
to `batch`/`abatch` on the SQL stores were flattened unchecked, so
`("foo.bar",)` could read, overwrite, or delete `("foo", "bar")`. Validate
each op in `_group_ops` before any SQL runs. Queued ops are already checked
before they are enqueued, so this only fails a direct caller's own batch.

Also document `InvalidNamespaceError` on the read, delete, and listing
methods, and tighten the base validation tests.
This commit is contained in:
byhow
2026-10-06 11:57:59 -07:00
parent 7184d7e3fa
commit 6c2b9568b7
8 changed files with 342 additions and 17 deletions
@@ -25,6 +25,7 @@ from langgraph.store.base import (
BaseStore,
GetOp,
IndexConfig,
InvalidNamespaceError,
Item,
ListNamespacesOp,
Op,
@@ -1386,11 +1387,36 @@ def _group_ops(ops: Iterable[Op]) -> tuple[dict[type, list[tuple[int, Op]]], int
grouped_ops: dict[type, list[tuple[int, Op]]] = defaultdict(list)
tot = 0
for idx, op in enumerate(ops):
_validate_op_namespace(op)
grouped_ops[type(op)].append((idx, op))
tot += 1
return grouped_ops, tot
def _validate_op_namespace(op: Op) -> None:
"""Reject namespace labels that would alias another namespace once dot-joined.
`("foo.bar",)` and `("foo", "bar")` flatten to the same `prefix` text.
`BaseStore` methods check labels before batching, but ops passed directly to
`batch`/`abatch` reach here unchecked.
"""
if isinstance(op, (GetOp, PutOp)):
paths = (op.namespace,)
elif isinstance(op, SearchOp):
paths = (op.namespace_prefix,)
elif isinstance(op, ListNamespacesOp):
paths = tuple(condition.path for condition in op.match_conditions or ())
else:
return
for path in paths:
for label in path:
if not isinstance(label, str) or not label or "." in label:
raise InvalidNamespaceError(
f"Invalid namespace label {label!r} found in {path}. Namespace "
"labels must be non-empty strings without periods ('.')."
)
def _json_loads(content: bytes | orjson.Fragment) -> Any:
if isinstance(content, orjson.Fragment):
if hasattr(content, "buf"):
@@ -13,8 +13,10 @@ import pytest
from langchain_core.embeddings import Embeddings
from langgraph.store.base import (
GetOp,
InvalidNamespaceError,
Item,
ListNamespacesOp,
MatchCondition,
PutOp,
SearchOp,
)
@@ -871,3 +873,62 @@ async def test_omit_expired_search_pagination(store: AsyncPostgresStore) -> None
page2 = await store.asearch(ns, limit=2, offset=2)
assert [i.key for i in page1] == ["a", "b"]
assert [i.key for i in page2] == ["c"]
async def test_abatch_rejects_ambiguous_namespace_labels(
store: AsyncPostgresStore,
) -> None:
await store.aput(("foo", "bar"), "key", {"original": True})
for op in (
GetOp(("foo.bar",), "key"),
PutOp(("foo.bar",), "key", {"changed": True}),
PutOp(("foo.bar",), "key", None),
SearchOp(("foo.bar",)),
ListNamespacesOp((MatchCondition("prefix", ("foo.bar",)),)),
):
with pytest.raises(InvalidNamespaceError):
await store.abatch([op])
item = await store.aget(("foo", "bar"), "key")
assert item is not None and item.value == {"original": True}
async def test_invalid_namespace_only_fails_its_own_call(
store: AsyncPostgresStore,
) -> None:
"""Concurrent calls share one `abatch`, which fails every op if it raises.
Labels are checked before an op is queued, so one caller's bad label cannot
fail another caller's request.
"""
await store.aput(("foo", "bar"), "key", {"original": True})
valid, invalid = await asyncio.gather(
store.aget(("foo", "bar"), "key"),
store.aget(("foo.bar",), "key"),
return_exceptions=True,
)
assert isinstance(valid, Item) and valid.value == {"original": True}
assert isinstance(invalid, InvalidNamespaceError)
async def test_sync_methods_reject_ambiguous_namespace_labels(
store: AsyncPostgresStore,
) -> None:
"""The sync wrappers run off the event loop thread and must validate too."""
await store.aput(("foo", "bar"), "key", {"original": True})
for call in (
lambda: store.get(("foo.bar",), "key"),
lambda: store.search(("foo.bar",)),
lambda: store.delete(("foo.bar",), "key"),
lambda: store.list_namespaces(prefix=("foo.bar",)),
lambda: store.batch([GetOp(("foo.bar",), "key")]),
):
with pytest.raises(InvalidNamespaceError):
await asyncio.to_thread(call)
item = await store.aget(("foo", "bar"), "key")
assert item is not None and item.value == {"original": True}
@@ -11,6 +11,7 @@ import pytest
from langchain_core.embeddings import Embeddings
from langgraph.store.base import (
GetOp,
InvalidNamespaceError,
Item,
ListNamespacesOp,
MatchCondition,
@@ -1164,3 +1165,51 @@ def test_namespace_labels_with_trailing_newline(store) -> None:
assert set(store.list_namespaces(prefix=["users", "alice"], limit=100)) == {
("users", "alice"),
}
@pytest.mark.parametrize("namespace", [("foo.bar",), ("foo", ""), ("foo", 1)])
@pytest.mark.parametrize(
"kind", ["get", "put", "delete", "search", "list_prefix", "list_suffix"]
)
def test_batch_rejects_ambiguous_namespace_labels(
store, namespace: tuple, kind: str
) -> None:
"""Ops passed straight to `batch` must not reach another namespace.
Namespaces are stored dot-joined, so `("foo.bar",)` flattens to the same
text as `("foo", "bar")`. `BaseStore` methods validate labels themselves,
but `batch` takes ops as given.
"""
op = {
"get": GetOp(namespace, "key"),
"put": PutOp(namespace, "key", {"changed": True}),
"delete": PutOp(namespace, "key", None),
"search": SearchOp(namespace),
"list_prefix": ListNamespacesOp((MatchCondition("prefix", namespace),)),
"list_suffix": ListNamespacesOp((MatchCondition("suffix", namespace),)),
}[kind]
store.put(("foo", "bar"), "key", {"original": True})
with pytest.raises(InvalidNamespaceError):
store.batch([PutOp(("valid",), "key", {}), op])
item = store.get(("foo", "bar"), "key")
assert item is not None and item.value == {"original": True}
# The whole batch is rejected before any SQL runs.
assert store.get(("valid",), "key") is None
def test_batch_allows_empty_search_prefix_and_listing_wildcards(
store,
) -> None:
store.put(("foo", "bar"), "key", {"v": 1})
found, listed = store.batch(
[
SearchOp(()),
ListNamespacesOp((MatchCondition("prefix", ("foo", "*")),)),
]
)
assert [item.namespace for item in found] == [("foo", "bar")]
assert listed == [("foo", "bar")]