diff --git a/litellm/caching/dual_cache.py b/litellm/caching/dual_cache.py index bef04a5c23c..8542d3364df 100644 --- a/litellm/caching/dual_cache.py +++ b/litellm/caching/dual_cache.py @@ -150,13 +150,23 @@ class DualCache(BaseCache): def set_cache(self, key, value, local_only: bool = False, **kwargs): # Update both Redis and in-memory cache try: + # ``kwargs`` is this call's own dict, so injecting a ttl into it is safe. What must not + # happen is one tier's injected ttl reaching the other, so note whether the caller + # supplied one before either tier writes back. + ttl_given: Final = "ttl" in kwargs if self.in_memory_cache is not None: - if "ttl" not in kwargs and self.default_in_memory_ttl is not None: + if not ttl_given and self.default_in_memory_ttl is not None: kwargs["ttl"] = self.default_in_memory_ttl self.in_memory_cache.set_cache(key, value, **kwargs) if self.redis_cache is not None and local_only is False: + if not ttl_given: + redis_ttl = self.default_redis_ttl + if redis_ttl is None: + redis_ttl = self.default_in_memory_ttl + if redis_ttl is not None: + kwargs["ttl"] = redis_ttl self.redis_cache.set_cache(key, value, **kwargs) except Exception as e: print_verbose(e) @@ -505,12 +515,19 @@ class DualCache(BaseCache): async def async_set_cache(self, key, value, local_only: bool = False, **kwargs): print_verbose(f"async set cache: cache key: {key}; local_only: {local_only}; value: {value}") try: + ttl_given: Final = "ttl" in kwargs if self.in_memory_cache is not None: - if "ttl" not in kwargs and self.default_in_memory_ttl is not None: + if not ttl_given and self.default_in_memory_ttl is not None: kwargs["ttl"] = self.default_in_memory_ttl await self.in_memory_cache.async_set_cache(key, value, **kwargs) if self.redis_cache is not None and local_only is False: + if not ttl_given: + redis_ttl = self.default_redis_ttl + if redis_ttl is None: + redis_ttl = self.default_in_memory_ttl + if redis_ttl is not None: + kwargs["ttl"] = redis_ttl await self.redis_cache.async_set_cache(key, value, **kwargs) except Exception as e: log_redis_failure( @@ -537,7 +554,10 @@ class DualCache(BaseCache): effective_ttl: Final = self.default_in_memory_ttl if ttl is None else ttl if self.in_memory_cache is not None: await self.in_memory_cache.async_set_cache(key, value, ttl=effective_ttl) - return batch.set(key, value, effective_ttl) + redis_ttl = self.default_redis_ttl if ttl is None else ttl + if redis_ttl is None: + redis_ttl = effective_ttl + return batch.set(key, value, redis_ttl) # async_batch_set_cache async def async_set_cache_pipeline( @@ -548,12 +568,19 @@ class DualCache(BaseCache): """ print_verbose(f"async batch set cache: cache keys: {cache_list}; local_only: {local_only}") try: + ttl_given: Final = "ttl" in kwargs if self.in_memory_cache is not None: - if "ttl" not in kwargs and self.default_in_memory_ttl is not None: + if not ttl_given and self.default_in_memory_ttl is not None: kwargs["ttl"] = self.default_in_memory_ttl await self.in_memory_cache.async_set_cache_pipeline(cache_list=cache_list, **kwargs) if self.redis_cache is not None and local_only is False: + if not ttl_given: + redis_ttl = self.default_redis_ttl + if redis_ttl is None: + redis_ttl = self.default_in_memory_ttl + if redis_ttl is not None: + kwargs["ttl"] = redis_ttl await self.redis_cache.async_set_cache_pipeline( cache_list=cache_list, ttl=kwargs.pop("ttl", None), **kwargs ) diff --git a/litellm/proxy/common_utils/user_api_key_cache.py b/litellm/proxy/common_utils/user_api_key_cache.py index 8c830816207..c59412d65f6 100644 --- a/litellm/proxy/common_utils/user_api_key_cache.py +++ b/litellm/proxy/common_utils/user_api_key_cache.py @@ -71,17 +71,21 @@ class UserApiKeyCache(DualCache): default_redis_ttl: float | None = None, key_object_in_memory_cache: InMemoryCache | None = None, ) -> None: + # The auth cache contract is that Redis mirrors the in-memory TTL when no + # explicit Redis TTL is configured (see #43187). Compute the effective + # Redis TTL here so DualCache stays generic and does not special-case us. + effective_redis_ttl: Final = default_redis_ttl if default_redis_ttl is not None else default_in_memory_ttl super().__init__( in_memory_cache=in_memory_cache, redis_cache=redis_cache, default_in_memory_ttl=default_in_memory_ttl, - default_redis_ttl=default_redis_ttl, + default_redis_ttl=effective_redis_ttl, ) self.key_object_cache: Final = DualCache( in_memory_cache=key_object_in_memory_cache or InMemoryCache(), redis_cache=redis_cache, default_in_memory_ttl=default_in_memory_ttl, - default_redis_ttl=default_redis_ttl, + default_redis_ttl=effective_redis_ttl, ) def in_memory_cache_for(self, key: str) -> InMemoryCache: diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index ed2a5b89a32..776f39f650d 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -5003,10 +5003,13 @@ def _attach_redis_usage_cache(redis_cache: RedisCache, enable_redis_auth_cache: default_redis_ttl=CLI_SSO_SESSION_TTL_SECONDS, ) if enable_redis_auth_cache is True: - user_api_key_cache.attach_redis_cache( - redis_cache, - default_redis_ttl=litellm.default_redis_ttl, - ) + # The auth cache mirrors its own in-memory TTL on the Redis tier (see + # UserApiKeyCache.__init__ and #43187). Do not pass the global + # default_redis_ttl here: that would override the auth cache's shorter + # TTL whenever default_redis_ttl is set before Redis is attached. + # Operators who want a longer auth-cache TTL should set + # general_settings.user_api_key_cache_ttl. + user_api_key_cache.attach_redis_cache(redis_cache) verbose_proxy_logger.info( "enable_redis_auth_cache=True: attached Redis to " "user_api_key_cache — virtual-key lookups are now " diff --git a/tests/unit/caching/test_dual_cache.py b/tests/unit/caching/test_dual_cache.py index 46600e0bf60..74c12c0c58e 100644 --- a/tests/unit/caching/test_dual_cache.py +++ b/tests/unit/caching/test_dual_cache.py @@ -1022,3 +1022,82 @@ async def test_async_batch_reads_of_missing_keys_hit_redis_once_per_expiry_windo dual_cache.last_redis_batch_access_time.update({key: time.time() - 61 for key in keys}) await dual_cache.async_batch_get_cache(keys) assert redis_cache.async_batch_get_cache.await_count == 2 + +async def _write_without_a_ttl(dual_cache: DualCache, write_path: str) -> None: + """One of the three write paths, called the way a caller that passes no ``ttl`` calls it.""" + if write_path == "set_cache": + dual_cache.set_cache("ttl_key", "v") + elif write_path == "async_set_cache": + await dual_cache.async_set_cache("ttl_key", "v") + else: + await dual_cache.async_set_cache_pipeline([("ttl_key", "v")]) + + +async def _write_with_a_ttl(dual_cache: DualCache, write_path: str, ttl: float) -> None: + """The same three paths with an explicit ``ttl``, which must reach both tiers unchanged.""" + if write_path == "set_cache": + dual_cache.set_cache("ttl_key", "v", ttl=ttl) + elif write_path == "async_set_cache": + await dual_cache.async_set_cache("ttl_key", "v", ttl=ttl) + else: + await dual_cache.async_set_cache_pipeline([("ttl_key", "v")], ttl=ttl) + + +def _ttl_recording_redis() -> MagicMock: + """A Redis tier that records what it was written with, whichever of the two writes it uses.""" + mock_redis = MagicMock() + mock_redis.async_set_cache = AsyncMock() + mock_redis.async_set_cache_pipeline = AsyncMock() + return mock_redis + + +def _written_redis_ttl(mock_redis: MagicMock, write_path: str) -> object: + """The ttl the recording Redis tier was written with, on whichever of its writes that path uses.""" + return getattr(mock_redis, write_path).call_args.kwargs["ttl"] + + +async def _tier_ttls( + write_path: str, ttl: float | None +) -> tuple[object, float]: + """Write one key through one path and report the Redis ttl it was written with, plus the + in-memory expiry. ``ttl=None`` means the caller supplied no ``ttl`` at all.""" + in_memory_cache = InMemoryCache(default_ttl=600) + mock_redis = _ttl_recording_redis() + dual_cache = DualCache( + in_memory_cache=in_memory_cache, + redis_cache=mock_redis, + default_in_memory_ttl=60, + default_redis_ttl=3600, + ) + if ttl is None: + await _write_without_a_ttl(dual_cache, write_path) + else: + await _write_with_a_ttl(dual_cache, write_path, ttl) + return _written_redis_ttl(mock_redis, write_path), in_memory_cache.ttl_dict["ttl_key"] + + +@pytest.mark.asyncio +@pytest.mark.parametrize("write_path", ["set_cache", "async_set_cache", "async_set_cache_pipeline"]) +async def test_dual_cache_writes_the_redis_tier_with_the_configured_default_redis_ttl(write_path: str): + """ + Regression for #43187: the Redis tier was written with default_in_memory_ttl, so a configured + default_redis_ttl never took effect. The in-memory tier still takes its own default. + """ + before = time.time() + written, expiry = await _tier_ttls(write_path, None) + after = time.time() + + assert written == 3600 + assert before + 60 <= expiry <= after + 60 + + +@pytest.mark.asyncio +@pytest.mark.parametrize("write_path", ["set_cache", "async_set_cache", "async_set_cache_pipeline"]) +async def test_dual_cache_gives_an_explicit_ttl_to_both_tiers_unchanged(write_path: str): + """An explicit ``ttl`` still overrides both tier defaults rather than being replaced by them.""" + before = time.time() + written, expiry = await _tier_ttls(write_path, 99.0) + after = time.time() + + assert written == 99.0 + assert before + 99 <= expiry <= after + 99 diff --git a/tests/unit/caching/test_request_redis_batch_pre_call.py b/tests/unit/caching/test_request_redis_batch_pre_call.py index cee7bfd8c65..69f7a58deee 100644 --- a/tests/unit/caching/test_request_redis_batch_pre_call.py +++ b/tests/unit/caching/test_request_redis_batch_pre_call.py @@ -1054,6 +1054,19 @@ async def test_a_pipelined_management_write_without_a_ttl_expires_in_redis_like_ assert [(c[0], c[1], c[3]) for c in client.pipelines[0].commands] == [("SET", "team_id:t1", 5)] +@pytest.mark.asyncio +async def test_a_pipelined_write_without_a_ttl_uses_the_configured_default_redis_ttl(): + """Regression for #43187: the pipelined Redis SET was written with default_in_memory_ttl, so a + configured default_redis_ttl never reached it.""" + client = FakeClient(_lua_ok_replies) + redis_cache = FakeRedisCache(client) + cache = DualCache(redis_cache=redis_cache, default_in_memory_ttl=5, default_redis_ttl=3600) + with request_redis_batch_scope() as request: + await cache.async_set_cache_pre_call("k1", {"v": 1}, None) + await request.flush_all() + assert [(c[0], c[1], c[3]) for c in client.pipelines[0].commands] == [("SET", "k1", 3600)] + + @pytest.mark.asyncio async def test_identity_prefetch_is_one_mget_after_which_hits_and_misses_alike_cost_no_read(): client = FakeClient(replies) diff --git a/tests/unit/proxy/test_redis_auth_cache_flag.py b/tests/unit/proxy/test_redis_auth_cache_flag.py index cb600bbb5fd..a2ce7a1028d 100644 --- a/tests/unit/proxy/test_redis_auth_cache_flag.py +++ b/tests/unit/proxy/test_redis_auth_cache_flag.py @@ -191,3 +191,27 @@ class TestRedisAuthCacheFlag: ps._attach_redis_usage_cache(fake_redis, enable_redis_auth_cache=False) assert limiter_cache.redis_cache is fake_redis assert ps.user_api_key_cache.redis_cache is None + +def test_auth_cache_redis_tier_keeps_in_memory_ttl_when_default_redis_ttl_is_set(monkeypatch): + """ + #43187: user_api_key_cache mirrors its in-memory TTL on Redis. A global + default_redis_ttl, from litellm_settings or cache_params, must not override + that when Redis is attached. + """ + from types import SimpleNamespace + + from litellm.proxy.common_utils.user_api_key_cache import UserApiKeyCache + + monkeypatch.setattr(litellm, "default_redis_ttl", 3600) + auth_cache = UserApiKeyCache(default_in_memory_ttl=60) + with ( + patch.object(ps, "user_api_key_cache", auth_cache), + patch.object(ps, "spend_counter_cache", DualCache()), + patch.object(ps, "cli_sso_session_cache", DualCache()), + patch.object(ps, "litellm_config_cache", SimpleNamespace(redis_cache=None)), + ): + ps._attach_redis_usage_cache(_FakeRedisCache(), enable_redis_auth_cache=True) + + assert auth_cache.redis_cache is not None + assert auth_cache.default_redis_ttl == 60 + assert auth_cache.key_object_cache.default_redis_ttl == 60