Check for existing issues

• I have searched the existing issues and checked that my issue is not a duplicate.

What happened?

Found a bug that broke our spend tracking in production. When you set cache: true with cache_params that has literally any key in it (like mode: default_off), the proxy completely ignores Redis environment variables and falls back to in-memory cache instead.

This is a problem because spend tracking between pods relies on Redis to sync counters. Without it, each pod tracks usage independently and the whole thing falls apart.

Took me a while to debug but the issue is in proxy_server.py around line 3138:

if (
    cache_type == "redis" or cache_type == "redis-semantic"
) and len(cache_params.keys()) == 0:
    cache_host = get_secret("REDIS_HOST", None)
    # ...

The len(cache_params.keys()) == 0 check means it only loads Redis from env vars when cache_params is completely empty. If you add mode or any other key, this condition fails and Redis config from env never gets added.

Then later when the code checks isinstance(litellm.cache.cache, RedisCache) to set up spend_counter_cache, it returns False because we got InMemoryCache instead.

Workaround: Adding type: redis explicitly fixes it:

cache_params:
  type: redis
  mode: default_off

But this isn't obvious and people using mode: default_off probably won't realize they need to also specify the type.

Possible fix: Instead of checking if cache_params is empty, check if Redis env vars actually exist and use them when type/host aren't already specified in cache_params. Something like:

if cache_type == "redis" and "host" not in cache_params:
    cache_host = get_secret("REDIS_HOST", None)
    if cache_host is not None:
        cache_params.update({
            "type": "redis",
            "host": cache_host,
            "port": get_secret("REDIS_PORT", None),
        })
        # etc

This bit us pretty hard in production because usage tracking silently failed and we only noticed when budgets weren't enforced correctly. Worth fixing since multi-pod deployments are probably pretty common.

Steps to Reproduce

How to reproduce:

We have Redis configured via env vars:

REDIS_HOST=our-redis-host
REDIS_PORT=6379
REDIS_PASSWORD=xxx

And this config:

litellm_settings:
  cache: true
  cache_params:
    mode: default_off

Running multiple pods, making requests through different ones, and spend tracking is completely wrong.

What I expected: Redis should connect using the env vars and sync spend counters across pods.

What actually happens:

• Cache gets initialized as InMemoryCache instead of RedisCache
• spend_counter_cache.redis_cache stays None
• Each pod has its own local counters
• Usage tracking doesn't work properly

Relevant log output

What part of LiteLLM is this about?

Proxy

What LiteLLM version are you on ?

v1.28.0

Twitter / LinkedIn details

No response