fix(router): scope tag-based routing to prefixed tags only once tag_routing_prefix is configured

Once tag_routing_prefix is set, an unrelated request tag a caller happens to send (e.g. a
logging or attribution tag with no routing intent) was still treated as a routing signal,
causing no_deployments_with_tag_routing whenever no deployment in the group carried tags.
Scopes matching, requiring, and excluding to tags the caller explicitly marked with the prefix,
while key/team/project policy tags in inherited_tags keep applying regardless of prefix, since
they are not caller-controlled.

Fixes #39901
This commit is contained in:
Deepanshu 2026-09-05 08:49:29 -04:00
parent 6c69dd0f72
commit 4a18a022e2
2 changed files with 278 additions and 1 deletions

View file

@ -434,6 +434,49 @@ def _inherited_constraint_sets(
return frozenset(inherited_required), frozenset(inherited_excluded)
def _inherited_full_tags(inherited_tags: object, routing_prefix: str) -> frozenset[str]:
# Every policy-contributed tag's exact, marker-preserving rewritten form --
# used to exempt inherited tags from the request-tag noop filter below, since
# key/team/project policy is not caller-controlled and merges its tags into
# the same "tags" list a caller's own tags land in (see
# litellm_pre_call_utils.py). Matching on the full tag (not just its bare
# value, as an earlier version of this helper did) matters: bare-value-only
# matching would let a caller smuggle an unprefixed tag past the filter just
# by sharing its bare value with a policy tag under a different "&"/"!"
# marker (e.g. policy requires "&region:eu"; caller separately, unprefixed,
# sends "!region:eu" -- bare-value matching would wrongly exempt it too).
if not isinstance(inherited_tags, (list, tuple)):
return frozenset()
return frozenset(_strip_routing_prefix(inherited_tags, routing_prefix)[0])
def _scope_to_confirmed_and_inherited_tags(
rewritten_tags: tuple[str, ...],
original_tags: Sequence[str],
inherited_tags: object,
routing_prefix: str,
) -> tuple[str, ...]:
# Once a routing prefix is configured, an unprefixed request tag (e.g. an
# unrelated attribution tag like "user_id:234") must never be treated as a
# routing signal at all -- not matched, not required, not excluded, and
# never a cause of no_deployments_with_tag_routing. Only tags the caller
# explicitly marked with the prefix participate in tag-based routing.
# Exemption is keyed on each tag's exact, marker-preserving rewritten form
# (not _strip_routing_prefix's own `confirmed` return, which is bare values
# only, used by _chain_allows_fail_open's "known required tag" check for a
# different purpose): bare-value matching would let a caller smuggle an
# unprefixed tag past this filter just by sharing its bare value with a
# genuinely confirmed or inherited tag under a different "&"/"!" marker.
# _inherited_full_tags exempts key/team/project policy's own tags
# unconditionally: policy isn't caller-controlled and was never expected to
# carry the prefix, so it must keep applying regardless.
confirmed_full_tags: Final = frozenset(
rewritten for rewritten, original in zip(rewritten_tags, original_tags) if original.startswith(routing_prefix)
)
inherited_full_tags: Final = _inherited_full_tags(inherited_tags, routing_prefix)
return tuple(t for t in rewritten_tags if t in confirmed_full_tags or t in inherited_full_tags)
def _tag_known_to_group(
llm_router_instance: LitellmRouter,
model: str,
@ -532,7 +575,14 @@ async def get_deployments_for_tag(
# caller-declared routing directive, exempt from the "maybe foreign to this
# group" heuristics that unprefixed tags still go through unchanged below.
rewritten_tags, routing_confirmed = _strip_routing_prefix(request_tags or [], routing_prefix)
required_tags, positive_tags, excluded_patterns = _split_tags(rewritten_tags)
scoped_tags: Final = (
_scope_to_confirmed_and_inherited_tags(
rewritten_tags, request_tags or (), metadata.get("inherited_tags"), routing_prefix
)
if routing_prefix
else rewritten_tags
)
required_tags, positive_tags, excluded_patterns = _split_tags(scoped_tags)
inherited_required_set, inherited_excluded_set = _inherited_constraint_sets(
metadata.get("inherited_tags"), routing_prefix
)

View file

@ -3160,3 +3160,230 @@ async def test_chat_request_carrying_litellm_metadata_still_routes_on_proxy_merg
)
assert deployment["model_info"]["id"] == "team-a-deployment"
# --- unprefixed request tags must be a no-op once tag_routing_prefix is configured
# (GitHub issue #39901): an unrelated attribution tag (e.g. "user_id:234") the
# caller happens to send must never be treated as a routing signal the moment an
# operator opts into tag_routing_prefix, since the prefix's entire point is to
# scope which tags carry routing intent. ---
def _untagged_single_deployment_router(tag_routing_prefix: str = "route:"):
return litellm.Router(
model_list=[
{
"model_name": "chat",
"litellm_params": {
"model": "gpt-4o",
"api_base": "https://exampleopenaiendpoint-production.up.railway.app/",
},
"model_info": {"id": "only-deployment"},
},
],
enable_tag_filtering=True,
tag_routing_prefix=tag_routing_prefix,
)
@pytest.mark.asyncio()
async def test_unprefixed_request_tag_is_noop_when_prefix_configured():
# No deployment in the group carries any tags at all -- a perfectly normal,
# unadopted state. An unrelated tag the caller happens to send must not be
# treated as a routing signal, matched, or cause no_deployments_with_tag_routing,
# once tag_routing_prefix is configured.
router = _untagged_single_deployment_router()
response = await router.acompletion(
model="chat",
messages=[{"role": "user", "content": "hi"}],
metadata={"tags": ["user_id:234"]},
mock_response="hi",
)
assert response._hidden_params["model_id"] == "only-deployment"
@pytest.mark.asyncio()
async def test_no_tags_still_noop_when_prefix_configured():
# Baseline regression guard: a request with no tags at all already worked
# before this fix and must keep working identically after it.
router = _untagged_single_deployment_router()
response = await router.acompletion(
model="chat",
messages=[{"role": "user", "content": "hi"}],
metadata={"tags": []},
mock_response="hi",
)
assert response._hidden_params["model_id"] == "only-deployment"
@pytest.mark.asyncio()
async def test_prefixed_tag_still_raises_when_unresolvable():
# A genuinely prefixed tag remains a real, trusted routing directive: if it
# can't be resolved against any deployment, this must still raise. Proves the
# fix scopes out unprefixed noise without disabling tag routing entirely.
router = _untagged_single_deployment_router()
with pytest.raises(Exception, match='Not allowed to access model due to tags configuration\\.') as exc_info:
await router.acompletion(
model="chat",
messages=[{"role": "user", "content": "hi"}],
metadata={"tags": ["route:quality:nonexistent"]},
mock_response="hi",
)
from litellm.types.router import RouterErrors
assert RouterErrors.no_deployments_with_tag_routing.value in str(exc_info.value)
@pytest.mark.asyncio()
async def test_unprefixed_tag_still_raises_when_no_prefix_configured():
# Backward-compat guard: with tag_routing_prefix left at its default empty
# string, an unprefixed tag must behave exactly as it did before this fix --
# the new filtering only activates once a prefix is actually configured.
router = _untagged_single_deployment_router(tag_routing_prefix="")
with pytest.raises(Exception, match='Not allowed to access model due to tags configuration\\.') as exc_info:
await router.acompletion(
model="chat",
messages=[{"role": "user", "content": "hi"}],
metadata={"tags": ["user_id:234"]},
mock_response="hi",
)
from litellm.types.router import RouterErrors
assert RouterErrors.no_deployments_with_tag_routing.value in str(exc_info.value)
@pytest.mark.asyncio()
async def test_inherited_required_tag_unaffected_by_prefix_scoping():
# An unprefixed admin/team-policy "&" requirement (merged into "tags" by
# litellm_pre_call_utils.py and snapshotted into inherited_tags) is not
# caller-controlled and must keep constraining deployment choice even once
# tag_routing_prefix is configured -- the new scoping applies only to what the
# caller themselves declared, not to what policy already contributed.
router = _eu_region_router()
router.update_settings(tag_routing_prefix="route:")
response = await router.acompletion(
model="chat",
messages=[{"role": "user", "content": "hi"}],
metadata={"tags": ["&region:eu"], "inherited_tags": ["&region:eu"]},
mock_response="hi",
)
assert response._hidden_params["model_id"] == "eu-1"
@pytest.mark.asyncio()
async def test_inherited_excluded_tag_unaffected_by_prefix_scoping():
# Same guarantee for an unprefixed admin/team-policy "!" exclusion, exercised
# without exhausting the pool (no fail-open floor involved) -- this proves the
# primary (non-fail-open) candidate computation itself still honors it, not
# just the trusted-only fail-open fallback.
router = _eu_region_router()
router.update_settings(tag_routing_prefix="route:")
response = await router.acompletion(
model="chat",
messages=[{"role": "user", "content": "hi"}],
metadata={"tags": ["!region:us"], "inherited_tags": ["!region:us"]},
mock_response="hi",
)
assert response._hidden_params["model_id"] == "eu-1"
@pytest.mark.asyncio()
async def test_foreign_tag_ignored_alongside_inherited_constraint_when_prefix_configured():
# Realistic combined scenario: the caller's own unrelated attribution tag
# rides alongside an inherited policy requirement in the same request. The
# foreign tag must be ignored while the inherited requirement still routes
# correctly. Deliberately not using _eu_region_router() here: both of its
# deployments set allow_fail_open=True, so a leaked "user_id:234" positive
# tag would still resolve to "eu-1" via the fail-open floor's own
# inherited-tags intersection, masking a regression in the filter itself.
# No allow_fail_open here means a leaked foreign tag has no safety net to
# hide behind -- it must show up as an outright raise, not a lucky match.
router = litellm.Router(
model_list=[
{
"model_name": "chat",
"litellm_params": {
"model": "gpt-4o",
"api_base": "https://exampleopenaiendpoint-production.up.railway.app/",
"tags": ["region:eu"],
},
"model_info": {"id": "eu-1"},
},
{
"model_name": "chat",
"litellm_params": {
"model": "gpt-4o-mini",
"api_base": "https://exampleopenaiendpoint-production.up.railway.app/",
"tags": ["region:us"],
},
"model_info": {"id": "us-1"},
},
],
enable_tag_filtering=True,
)
router.update_settings(tag_routing_prefix="route:")
response = await router.acompletion(
model="chat",
messages=[{"role": "user", "content": "hi"}],
metadata={"tags": ["&region:eu", "user_id:234"], "inherited_tags": ["&region:eu"]},
mock_response="hi",
)
assert response._hidden_params["model_id"] == "eu-1"
@pytest.mark.asyncio()
async def test_unprefixed_tag_cannot_piggyback_via_bare_value_collision_with_confirmed_tag():
# Regression: exemption from the noop filter must be keyed on a tag's exact,
# marker-preserving form, not just its bare value. A caller sending both a
# genuinely confirmed positive tag ("route:region:eu") and an unrelated,
# unprefixed exclusion sharing the same bare value ("!region:eu") must not
# have that exclusion smuggled through just because "region:eu" happens to
# be confirmed under a different marker -- the exclusion is caller noise and
# must stay inert, same as any other unprefixed tag.
router = litellm.Router(
model_list=[
{
"model_name": "chat",
"litellm_params": {
"model": "gpt-4o",
"api_base": "https://exampleopenaiendpoint-production.up.railway.app/",
"tags": ["region:eu"],
},
"model_info": {"id": "eu-1"},
},
{
"model_name": "chat",
"litellm_params": {
"model": "gpt-4o-mini",
"api_base": "https://exampleopenaiendpoint-production.up.railway.app/",
"tags": ["region:us"],
},
"model_info": {"id": "us-1"},
},
],
enable_tag_filtering=True,
tag_routing_prefix="route:",
)
response = await router.acompletion(
model="chat",
messages=[{"role": "user", "content": "hi"}],
metadata={"tags": ["route:region:eu", "!region:eu"]},
mock_response="hi",
)
assert response._hidden_params["model_id"] == "eu-1"