[Fix] Refresh Redis TTL on counter writes, skip stale in-memory in Redis

[Michael-RZ-Berri]

Relevant issues

Budgets can still be bypassed on multi-pod / multi-worker setups because of how caching is organized, and pods not communicating spending to eachother. This PR does a few things to enforce budgets more correctly on multi-pod:

• get_current_spend and SpendCounterReseed.coalesced skip stale per-pod in-memory on a clean Redis miss; reseed from DB instead.

• RedisCache.async_increment gains refresh_ttl opt-in. Spend counters now use it so the one minute TTL extends / resets when the counter is being hit, other keys keep the previous set-once behavior. This is for performance reasons (keys accessed often avoid pinging the DB too much, especially in multi-pod).

• ResetBudgetJob invalidates the Redis counter alongside every DB row reset (keys, users, teams, team members, budgets-linked keys), so refreshed counters still get reset.

Linear ticket

Resolves LIT-2466.

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have Added testing in the tests/test_litellm/ directory, Adding at least 1 test is a hard requirement - see details
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

Screenshots / Proof of Fix

Before:

After:

Type

🐛 Bug Fix
🚄 Infrastructure
✅ Test

Changes

dual_cache, redis_cache, reset_budget_job, spend_counter_reseed, proxy_server

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

Michael Riad Zaky seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[Michael-RZ-Berri]

@greptile-ai

[Michael-RZ-Berri]

@greptile-ai

[codecov]

Codecov Report

❌ Patch coverage is 87.75510% with 6 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
litellm/proxy/common_utils/reset_budget_job.py	89.74%	4 Missing ⚠️
litellm/caching/redis_cache.py	75.00%	1 Missing ⚠️
litellm/proxy/db/spend_counter_reseed.py	83.33%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[greptile-apps]

Greptile Summary

This PR fixes multi-pod budget bypass by (1) skipping stale per-pod in-memory on a clean Redis miss and reseeding from DB instead, (2) adding refresh_ttl=True so spend-counter TTLs are extended on every write rather than only at first creation, and (3) correctly ordering Redis counter invalidation after the DB row reset in ResetBudgetJob for all entity types.

The DB-write-first ordering is now consistently applied across keys, users, teams, team members, and linked-budget keys. Two minor P2 findings: the find_many exception in reset_budget_for_litellm_team_members is silently swallowed, which can leave stale Redis counters after a successful DB reset in transient-failure scenarios; and the sequential _invalidate_spend_counter loop could be parallelised with asyncio.gather for large batches.

Confidence Score: 5/5

Safe to merge — no security bypasses or data-loss paths introduced; both findings are P2.

All P1 concerns from prior review rounds (DB-before-Redis ordering, explicit TTL on reset writes) have been addressed. The two remaining findings are P2: a silent exception swallow that causes over-enforcement (not bypass) in a transient DB failure, and a sequential loop that could be batched. No P0 or P1 issues are present.

litellm/proxy/common_utils/reset_budget_job.py — the find_many exception handling in reset_budget_for_litellm_team_members is worth a second look.

Important Files Changed

Filename	Overview
litellm/caching/redis_cache.py	Adds refresh_ttl flag to async_increment; when True, calls expire() unconditionally on every write instead of only on initial creation. Logic is correct and backward-compatible.
litellm/caching/dual_cache.py	Threads refresh_ttl kwarg through async_increment_cache to the underlying Redis layer. Minimal, correct change.
litellm/proxy/proxy_server.py	Introduces redis_clean_miss flag in get_current_spend to skip stale per-pod in-memory on a clean Redis None; passes refresh_ttl=True on all spend counter increments. Correctly handles single-pod (no Redis) and multi-pod cases.
litellm/proxy/db/spend_counter_reseed.py	Mirrors the redis_clean_miss pattern in coalesced and uses refresh_ttl=True when warming the cache from DB. Correctly skips stale in-memory on a clean Redis miss.
litellm/proxy/common_utils/reset_budget_job.py	Introduces _invalidate_spend_counter and correctly orders DB-write-then-Redis-invalidate for keys, users, teams, linked-budget keys, and team members. The find_many exception swallow in reset_budget_for_litellm_team_members can leave stale Redis counters after a successful DB reset; sequential invalidation loops could be batched.
tests/test_litellm/caching/test_redis_cache.py	Adds two clean unit tests for refresh_ttl behaviour — one verifying expire() is called unconditionally, one verifying it is skipped on the default path. All mock-only, no real network calls.
tests/test_litellm/proxy/common_utils/test_reset_budget_job.py	Adds five counter-invalidation tests covering team-member, key, user, team, and linked-budget paths. All use mocks, no real network calls.
tests/test_litellm/proxy/test_proxy_server.py	Adds two tests verifying the redis_clean_miss behaviour in get_current_spend: clean Redis None triggers DB reseed, Redis error falls back to in-memory. Both are well-structured and mock-only.

_{Reviews (2): Last reviewed commit: "Reorder counter invalidation to run afte..." | Re-trigger Greptile}

[Michael-RZ-Berri]

@greptile-ai

[ryan-crabbe-berri]

lgtm