diff --git a/litellm/proxy/common_utils/user_api_key_cache.py b/litellm/proxy/common_utils/user_api_key_cache.py index f076dfc2ff9..a2272c3f73d 100644 --- a/litellm/proxy/common_utils/user_api_key_cache.py +++ b/litellm/proxy/common_utils/user_api_key_cache.py @@ -71,6 +71,7 @@ class UserApiKeyCache(DualCache): default_in_memory_ttl: float | None = None, default_redis_ttl: float | None = None, key_object_in_memory_cache: InMemoryCache | None = None, + budget_windows_absent_in_memory_cache: InMemoryCache | None = None, ) -> None: super().__init__( in_memory_cache=in_memory_cache, @@ -84,8 +85,10 @@ class UserApiKeyCache(DualCache): default_in_memory_ttl=default_in_memory_ttl, default_redis_ttl=default_redis_ttl, ) - self.budget_windows_absent_cache: Final = InMemoryCache( - default_ttl=DEFAULT_MANAGEMENT_OBJECT_IN_MEMORY_CACHE_TTL + self.budget_windows_absent_cache: Final = DualCache( + in_memory_cache=budget_windows_absent_in_memory_cache + or InMemoryCache(default_ttl=DEFAULT_MANAGEMENT_OBJECT_IN_MEMORY_CACHE_TTL), + default_in_memory_ttl=default_in_memory_ttl, ) def in_memory_cache_for(self, key: str) -> InMemoryCache: @@ -101,10 +104,14 @@ class UserApiKeyCache(DualCache): self.key_object_cache.update_cache_ttl( default_in_memory_ttl=default_in_memory_ttl, default_redis_ttl=default_redis_ttl ) + self.budget_windows_absent_cache.update_cache_ttl( + default_in_memory_ttl=default_in_memory_ttl, default_redis_ttl=None + ) def update_in_memory_max_size(self, max_size: int | None) -> None: super().update_in_memory_max_size(max_size) self.key_object_cache.update_in_memory_max_size(max_size) + self.budget_windows_absent_cache.update_in_memory_max_size(max_size) def attach_redis_cache( self, redis_cache: RedisCache | None = None, *, default_redis_ttl: float | None = None diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index 07776a2dd6f..99968151451 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -26,6 +26,7 @@ from collections.abc import ( Collection, Container, Iterable, + Iterator, Mapping, MutableMapping, Sequence, @@ -3701,7 +3702,68 @@ async def _increment_spend_counters_batched( cost: Final[float] = response_cost - async def _key_scope(key_token: str) -> tuple[PendingSpendIncrement | BaseException, ...]: + async def _window_increment( + entity_type: Litellm_EntityType, entity_id: str, window: object + ) -> PendingSpendIncrement | None: + duration: Final = ( + window["budget_duration"] if isinstance(window, dict) else getattr(window, "budget_duration", None) + ) + window_reset_at: Final = ( + window.get("reset_at") if isinstance(window, dict) else getattr(window, "reset_at", None) + ) + window_counter: Final = f"spend:{entity_type.value}:{entity_id}:window:{duration}" + window_start: Final = get_budget_window_start(window) + pending_window: Final = ( + await _prepare_window_spend_counter_increment( + counter_key=window_counter, + entity_type=entity_type.value.capitalize(), + entity_id=entity_id, + window_duration=duration, + window_start=window_start, + increment=cost, + ) + if window_counter not in reserved_counter_keys + else None + ) + await _enqueue_window_spend_row_update( + entity_type=entity_type, + entity_id=entity_id, + reset_at=window_reset_at, + window_duration=duration, + window_start=window_start, + increment=cost, + request_started_at=request_started_at, + ) + return pending_window + + async def _window_increments( + entity_type: Litellm_EntityType, entity_id: str, budget_limits: object + ) -> tuple[PendingSpendIncrement | BaseException, ...]: + if not isinstance(budget_limits, list): + return () + window_pending: Final = await asyncio.gather( + *(_window_increment(entity_type, entity_id, window) for window in budget_limits), return_exceptions=True + ) + return tuple(item for item in window_pending if item is not None) + + async def _cached_window_increments( + entity_type: Litellm_EntityType, entity_id: str, cache_key: str + ) -> tuple[_SpendScopeItem, ...]: + cached: Final[object] = _UNTYPED_VALUE.validate_python(await user_api_key_cache.async_get_cache(key=cache_key)) + if cached is None: + return (_UncachedBudgetWindows(entity_type=entity_type, entity_id=entity_id),) + return await _window_increments(entity_type, entity_id, _cached_budget_limits(cached)) + + async def _stored_window_increments( + uncached: _UncachedBudgetWindows, + ) -> tuple[PendingSpendIncrement | BaseException, ...]: + return await _window_increments( + uncached.entity_type, + uncached.entity_id, + await _stored_budget_limits(uncached.entity_type, uncached.entity_id), + ) + + async def _key_scope(key_token: str) -> tuple[_SpendScopeItem, ...]: # key_token arrives pre-hashed from metadata["user_api_key"] (auth flow # hashes raw "sk-..." keys before they reach the callback). The # startswith("sk-") check is a safety net matching update_cache — @@ -3724,54 +3786,9 @@ async def _increment_spend_counters_batched( ) ) - async def _key_window_increment(window: object) -> PendingSpendIncrement | None: - duration = ( - window["budget_duration"] if isinstance(window, dict) else getattr(window, "budget_duration", None) - ) - key_window_reset_at = ( - window.get("reset_at") if isinstance(window, dict) else getattr(window, "reset_at", None) - ) - key_window_counter: Final = f"spend:key:{hashed_token}:window:{duration}" - key_window_start = get_budget_window_start(window) - pending_window: Final = ( - await _prepare_window_spend_counter_increment( - counter_key=key_window_counter, - entity_type="Key", - entity_id=hashed_token, - window_duration=duration, - window_start=key_window_start, - increment=cost, - ) - if key_window_counter not in reserved_counter_keys - else None - ) - await _enqueue_window_spend_row_update( - entity_type=Litellm_EntityType.KEY, - entity_id=hashed_token, - reset_at=key_window_reset_at, - window_duration=duration, - window_start=key_window_start, - increment=cost, - request_started_at=request_started_at, - ) - return pending_window + return key_pending + await _cached_window_increments(Litellm_EntityType.KEY, hashed_token, hashed_token) - key_obj: Final[object] = _UNTYPED_VALUE.validate_python( - await user_api_key_cache.async_get_cache(key=hashed_token) - ) - key_budget_limits: Final = ( - _cached_budget_limits(key_obj) - if key_obj is not None - else await _stored_budget_limits(Litellm_EntityType.KEY, hashed_token) - ) - if not isinstance(key_budget_limits, list): - return key_pending - window_pending: Final = await asyncio.gather( - *(_key_window_increment(window) for window in key_budget_limits), return_exceptions=True - ) - return key_pending + tuple(item for item in window_pending if item is not None) - - async def _team_scope(scope_team_id: str) -> tuple[PendingSpendIncrement | BaseException, ...]: + async def _team_scope(scope_team_id: str) -> tuple[_SpendScopeItem, ...]: team_counter_key: Final = f"spend:team:{scope_team_id}" team_pending: Final[tuple[PendingSpendIncrement, ...]] = ( () @@ -3785,52 +3802,9 @@ async def _increment_spend_counters_batched( ) ) - async def _team_window_increment(window: object) -> PendingSpendIncrement | None: - duration = ( - window["budget_duration"] if isinstance(window, dict) else getattr(window, "budget_duration", None) - ) - team_window_reset_at = ( - window.get("reset_at") if isinstance(window, dict) else getattr(window, "reset_at", None) - ) - team_window_counter: Final = f"spend:team:{scope_team_id}:window:{duration}" - team_window_start = get_budget_window_start(window) - pending_window: Final = ( - await _prepare_window_spend_counter_increment( - counter_key=team_window_counter, - entity_type="Team", - entity_id=scope_team_id, - window_duration=duration, - window_start=team_window_start, - increment=cost, - ) - if team_window_counter not in reserved_counter_keys - else None - ) - await _enqueue_window_spend_row_update( - entity_type=Litellm_EntityType.TEAM, - entity_id=scope_team_id, - reset_at=team_window_reset_at, - window_duration=duration, - window_start=team_window_start, - increment=cost, - request_started_at=request_started_at, - ) - return pending_window - - team_obj: Final[object] = _UNTYPED_VALUE.validate_python( - await user_api_key_cache.async_get_cache(key=f"team_id:{scope_team_id}") + return team_pending + await _cached_window_increments( + Litellm_EntityType.TEAM, scope_team_id, f"team_id:{scope_team_id}" ) - team_budget_limits: Final = ( - _cached_budget_limits(team_obj) - if team_obj is not None - else await _stored_budget_limits(Litellm_EntityType.TEAM, scope_team_id) - ) - if not isinstance(team_budget_limits, list): - return team_pending - window_pending: Final = await asyncio.gather( - *(_team_window_increment(window) for window in team_budget_limits), return_exceptions=True - ) - return team_pending + tuple(item for item in window_pending if item is not None) async def _team_member_scope( scope_user_id: str, scope_team_id: str @@ -3901,22 +3875,30 @@ async def _increment_spend_counters_batched( # return_exceptions so a failing scope does not leave its siblings running # as orphaned tasks that race the caller's reservation-counter invalidation; # all scopes settle, then the first error propagates as before. - scope_results: Final = await asyncio.gather(*scope_coros, return_exceptions=True) - scope_errors: Final = tuple( - item - for scope in scope_results - for item in (scope if isinstance(scope, tuple) else (scope,)) - if isinstance(item, BaseException) + scope_items: Final = tuple(_flatten_spend_scopes(await asyncio.gather(*scope_coros, return_exceptions=True))) + uncached_windows: Final = tuple(item for item in scope_items if isinstance(item, _UncachedBudgetWindows)) + await _apply_spend_counter_increments( + pending=reservation_update.pending + + tuple(item for item in scope_items if isinstance(item, PendingSpendIncrement)) ) - pending: Final = tuple( - item - for scope in scope_results - if not isinstance(scope, BaseException) - for item in scope - if not isinstance(item, BaseException) + stored_window_items: Final = ( + tuple( + _flatten_spend_scopes( + await asyncio.gather( + *(_stored_window_increments(uncached) for uncached in uncached_windows), return_exceptions=True + ) + ) + ) + if uncached_windows + else () ) - await _apply_spend_counter_increments(pending=reservation_update.pending + pending) + stored_window_pending: Final = tuple( + item for item in stored_window_items if isinstance(item, PendingSpendIncrement) + ) + if stored_window_pending: + await _apply_spend_counter_increments(pending=stored_window_pending) stamp_budget_reservation_actual_cost(budget_reservation=budget_reservation, actual_cost=response_cost) + scope_errors: Final = tuple(item for item in scope_items + stored_window_items if isinstance(item, BaseException)) if scope_errors: raise scope_errors[0] @@ -3924,6 +3906,25 @@ async def _increment_spend_counters_batched( budget_reservation["finalized"] = True +@dataclass(frozen=True, slots=True) +class _UncachedBudgetWindows: + entity_type: Litellm_EntityType + entity_id: str + + +_SpendScopeItem: TypeAlias = PendingSpendIncrement | BaseException | _UncachedBudgetWindows + + +def _flatten_spend_scopes( + scope_results: Sequence[tuple[_SpendScopeItem, ...] | BaseException], +) -> Iterator[_SpendScopeItem]: + for scope in scope_results: + if isinstance(scope, BaseException): + yield scope + continue + yield from scope + + @dataclass(frozen=True, slots=True) class _ReservationCounterUpdate: """The reserved counters the direct increment must skip, and the adjustments that settle them on the actual diff --git a/tests/unit/proxy/common_utils/test_user_api_key_cache.py b/tests/unit/proxy/common_utils/test_user_api_key_cache.py index 09523cd1901..577e5e5a389 100644 --- a/tests/unit/proxy/common_utils/test_user_api_key_cache.py +++ b/tests/unit/proxy/common_utils/test_user_api_key_cache.py @@ -1,6 +1,6 @@ import hashlib import json -from typing import Any +from typing import Any, Final import pytest @@ -11,6 +11,7 @@ from litellm.constants import DEFAULT_MANAGEMENT_OBJECT_IN_MEMORY_CACHE_TTL from litellm.proxy._types import UserAPIKeyAuth from litellm.proxy.common_utils.user_api_key_cache import ( UserApiKeyCache, + budget_windows_absent_cache_key, end_user_cache_key, get_management_object_ttl, is_user_key_cache_key, @@ -32,6 +33,17 @@ class CapturingInMemoryCache(InMemoryCache): super().set_cache(key, value, **kwargs) +class _ManualClock: + def __init__(self) -> None: + self.seconds = 1000.0 + + def __call__(self) -> float: + return self.seconds + + def advance(self, seconds: float) -> None: + self.seconds += seconds + + class FakeRedisCache(RedisCache): """ In-memory fake that enforces the UserApiKeyCache Redis payload contract. @@ -432,6 +444,29 @@ class TestUserKeyObjectPartition: assert first_key.token == tokens[0] assert cache.get_cache(end_user_cache_key("u0")) == {"user_id": "u0"} + def test_update_in_memory_max_size_applies_to_no_window_notes(self): + cache = UserApiKeyCache() + cache.update_in_memory_max_size(1) + + cache.budget_windows_absent_cache.set_cache(key=budget_windows_absent_cache_key("key", "first"), value=True) + cache.budget_windows_absent_cache.set_cache(key=budget_windows_absent_cache_key("key", "second"), value=True) + + assert cache.budget_windows_absent_cache.get_cache(key=budget_windows_absent_cache_key("key", "first")) is None + assert cache.budget_windows_absent_cache.get_cache(key=budget_windows_absent_cache_key("key", "second")) is True + + def test_update_cache_ttl_applies_to_no_window_notes(self): + clock: Final = _ManualClock() + cache: Final = UserApiKeyCache( + default_in_memory_ttl=60, budget_windows_absent_in_memory_cache=InMemoryCache(clock=clock) + ) + cache.update_cache_ttl(default_in_memory_ttl=5, default_redis_ttl=5) + note: Final = budget_windows_absent_cache_key("key", "no-windows") + + cache.budget_windows_absent_cache.set_cache(key=note, value=True) + clock.advance(6) + + assert cache.budget_windows_absent_cache.get_cache(key=note) is None + def test_update_in_memory_max_size_none_resets_key_object_partition_to_default(self): cache = UserApiKeyCache(key_object_in_memory_cache=InMemoryCache(max_size_in_memory=1)) cache.update_in_memory_max_size(None) diff --git a/tests/unit/proxy/management_endpoints/test_key_management_endpoints.py b/tests/unit/proxy/management_endpoints/test_key_management_endpoints.py index 73bb93f7387..b453f985512 100644 --- a/tests/unit/proxy/management_endpoints/test_key_management_endpoints.py +++ b/tests/unit/proxy/management_endpoints/test_key_management_endpoints.py @@ -6,7 +6,7 @@ from contextlib import ExitStack from datetime import datetime, timedelta, timezone from types import SimpleNamespace from typing import Final -from unittest.mock import AsyncMock, MagicMock, patch +from unittest.mock import AsyncMock, MagicMock, call, patch from uuid import UUID, uuid4 import pytest @@ -51,7 +51,11 @@ from litellm.proxy.auth.auth_checks import ( ) from litellm.proxy.auth.user_api_key_auth import UserAPIKeyAuth from litellm.proxy.common_utils.auth_cache_invalidation_pubsub import AuthCacheInvalidationSubscriber -from litellm.proxy.common_utils.user_api_key_cache import UserApiKeyCache, project_cache_key +from litellm.proxy.common_utils.user_api_key_cache import ( + UserApiKeyCache, + budget_windows_absent_cache_key, + project_cache_key, +) from litellm.proxy.db.db_lookup_gate import db_lookup_stall_tracker from litellm.proxy.management_endpoints.key_management_endpoints import ( KeyProjectBindingError, @@ -8911,7 +8915,10 @@ async def test_delete_cache_key_object_broadcasts_invalidation(monkeypatch): # Real, observable state: the cache object itself no longer holds the entry. assert real_user_api_key_cache.get_cache(key="hashed-broadcast-key") is None - mock_publish.assert_any_await(cache_key="hashed-broadcast-key") + assert mock_publish.await_args_list == [ + call(cache_key="hashed-broadcast-key"), + call(cache_key=budget_windows_absent_cache_key("key", "hashed-broadcast-key")), + ] @pytest.mark.asyncio diff --git a/tests/unit/proxy/management_endpoints/test_team_endpoints.py b/tests/unit/proxy/management_endpoints/test_team_endpoints.py index e7868103c5a..2f4dad0ccd7 100644 --- a/tests/unit/proxy/management_endpoints/test_team_endpoints.py +++ b/tests/unit/proxy/management_endpoints/test_team_endpoints.py @@ -2684,7 +2684,7 @@ async def test_team_block_unblock_reach_warmed_auth_on_every_worker( from litellm.proxy._types import BlockTeamRequest from litellm.proxy.auth.auth_checks import get_team_object from litellm.proxy.auth.team_grants import team_model_aliases - from litellm.proxy.common_utils.user_api_key_cache import UserApiKeyCache + from litellm.proxy.common_utils.user_api_key_cache import UserApiKeyCache, budget_windows_absent_cache_key from litellm.proxy.management_endpoints import team_endpoints team_id: Final = "team-kill-switch" @@ -2721,7 +2721,11 @@ async def test_team_block_unblock_reach_warmed_auth_on_every_worker( seen_by_auth: Final = await get_team_object(team_id=team_id, prisma_client=prisma_client, user_api_key_cache=cache) assert seen_by_auth.blocked is blocked assert team_model_aliases(seen_by_auth) == {"fast": "gpt-4o"} - assert redis.broadcast_keys[:2] == [f"team_id:{team_id}", "team_alias:kill-switch"] + assert redis.broadcast_keys == [ + f"team_id:{team_id}", + "team_alias:kill-switch", + budget_windows_absent_cache_key("team", team_id), + ] assert (response.blocked, response.litellm_model_table, response.object_permission) == (blocked, None, None) diff --git a/tests/unit/proxy/test_proxy_server_endpoints_and_startup.py b/tests/unit/proxy/test_proxy_server_endpoints_and_startup.py index b48b02ccb32..645646388ac 100644 --- a/tests/unit/proxy/test_proxy_server_endpoints_and_startup.py +++ b/tests/unit/proxy/test_proxy_server_endpoints_and_startup.py @@ -14485,6 +14485,40 @@ async def test_late_requests_from_many_keys_with_windows_do_not_push_out_a_known assert no_window_key_reads() == reads_after_the_first_request +@pytest.mark.asyncio +async def test_a_slow_window_row_read_does_not_hold_back_the_key_spend_counter(): + import litellm.proxy.proxy_server as ps + from litellm.proxy.proxy_server import increment_spend_counters + + prisma_client: Final = _prisma_with_rows() + read_started: Final = asyncio.Event() + finish_read: Final = asyncio.Event() + + async def slow_row_read(where: Mapping[str, str]) -> _StoredRow: + read_started.set() + await finish_read.wait() + return _StoredRow([_window("7d", 5.0)]) + + prisma_client.db.litellm_verificationtoken.find_unique = slow_row_read + + with _window_spend_enqueue_env({}, prisma_client=prisma_client) as queue: + ps.spend_counter_cache.set_cache(key="spend:key:hashed-token", value=0.0) + ps.spend_counter_cache.set_cache(key="spend:key:hashed-token:window:7d", value=1.0) + charge: Final = asyncio.create_task( + increment_spend_counters(token="hashed-token", team_id=None, user_id=None, response_cost=0.25) + ) + assert await asyncio.wait_for(read_started.wait(), timeout=5) + key_spend_while_the_row_read_waits: Final = ps.spend_counter_cache.get_cache(key="spend:key:hashed-token") + finish_read.set() + await charge + window_spend: Final = ps.spend_counter_cache.get_cache(key="spend:key:hashed-token:window:7d") + charged: Final = await _drain(queue) + + assert key_spend_while_the_row_read_waits == pytest.approx(0.25) + assert window_spend == pytest.approx(1.25) + assert _charged_windows(charged) == [("hashed-token", "7d", pytest.approx(0.25))] + + @pytest.mark.asyncio async def test_a_late_request_that_misses_the_updated_team_charges_windows_the_team_did_not_have_before(): """Another worker's /team/update only evicts this worker's cached team, so its next late cost reads the row."""