fix(proxy): stop CacheCodec dropping null fields on cache round-trip

[ryan-crabbe-berri]

Relevant issues

Linear ticket

Resolves LIT-3277
Resolves LIT-3427

Pre-Submission checklist

• I have added meaningful tests
• My PR passes all CI/CD checks (e.g., lint, format, unit tests)
• 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

CacheCodec.serialize wrote cached Pydantic models with model_dump(exclude_none=True), which drops any None-valued key, while deserialize does a strict model_validate. When a cached model has a required-but-nullable field (Optional[X] with no default), a None value is written as an absent key and then fails model_validate on read with Field required, so the entry can never be read back

LiteLLM_ManagedVectorStoresTable is such a model, so this is live today on any deployment that resolves a managed vector store (the cache read in get_managed_vector_store_rows_by_uuids runs on the request path)

Repro config (Redis-backed cache + DB):

general_settings:
  master_key: os.environ/LITELLM_MASTER_KEY
  database_url: os.environ/DATABASE_URL
  store_model_in_db: true
litellm_settings:
  cache: true
  cache_params: { type: redis, host: localhost, port: 6379 }

Before (current litellm_internal_staging)

# 1) create a managed vector store whose optional columns are null
curl -s $BASE/vector_store/new -H "Authorization: Bearer $MASTER_KEY" -H "Content-Type: application/json" \
  -d '{"vector_store_id":"vs_lit3277_repro","custom_llm_provider":"openai"}'

# 2) resolve it 5x (the search reads the store from cache before anything else)
for i in 1 2 3 4 5; do
  curl -s $BASE/v1/vector_stores/vs_lit3277_repro/search -H "Authorization: Bearer $MASTER_KEY" \
    -H "Content-Type: application/json" -d '{"query":"hello world"}' -o /dev/null -w "search #$i -> HTTP %{http_code}\n"
done

# 3) how many cache reads failed to deserialize?
grep -c "CacheCodec.deserialize: validation failed" proxy.log

search #1 -> HTTP 500
search #2 -> HTTP 500
search #3 -> HTTP 500
search #4 -> HTTP 500
search #5 -> HTTP 500
4

Every read after the first logged:

ERROR ... UserApiKeyCache.async_get_cache failed to deserialize cached value for
  key='managed_vector_store_id:vs_lit3277_repro' model_type=LiteLLM_ManagedVectorStoresTable
WARNING ... CacheCodec.deserialize: validation failed for LiteLLM_ManagedVectorStoresTable (5 validation errors ...
  vector_store_name / vector_store_description / vector_store_metadata / litellm_credential_name / team_id
  Field required [type=missing] ...)

The HTTP 500 is the downstream provider call (there is no real openai vector store behind this id) and is unrelated; the cache read runs first, fails, and forces a DB re-query on every request so the cache never serves the row

After (this PR)

Same flow, same searches, count of deserialize failures is 0:

search #1 -> HTTP 500
search #2 -> HTTP 500
search #3 -> HTTP 500
search #4 -> HTTP 500
search #5 -> HTTP 500
0

Type

🐛 Bug Fix

Changes

• litellm/proxy/common_utils/cache_pydantic_utils.py: CacheCodec.serialize no longer passes exclude_none=True, so serialize and deserialize are a lossless pair and None fields are written as null. This is the root fix and covers every cached model, not just the one below
• litellm/models/managed_files.py: give the nine Optional fields on LiteLLM_ManagedVectorStoresTable a None default, matching every other cached model; this is the one cached model still declared required-but-nullable
• tests/test_litellm/proxy/common_utils/test_cache_codec.py: flip the two tests that asserted None keys were dropped, and add regression tests that fail under the old exclude_none behavior (a required-nullable None field must survive serialize then deserialize, and a real LiteLLM_ManagedVectorStoresTable with all optional fields null round-trips without a validation warning)

Out of scope here to keep this isolated: three readers still rebuild a cached model from the raw dict without the typed codec (budget_reservation.py, the budget path in auth_checks.py, mcp_server_manager.py). After this fix the cached dict is complete so they no longer break, but routing them through the typed codec would make a future schema drift degrade to a cache miss instead of an uncaught error; I can do that in a separate PR

[greptile-apps]

Greptile Summary

This PR fixes a silent data-loss bug in CacheCodec.serialize where model_dump(exclude_none=True) was stripping None-valued fields before writing to cache, while deserialize used a strict model_validate — causing any model with a required-but-nullable field to fail deserialization on every cache read after the first write.

• Root fix (cache_pydantic_utils.py): drops exclude_none=True from all four model_dump calls so None fields are serialized as JSON null, making serialize/deserialize a lossless pair for all cached models.
• Secondary fix (managed_files.py): adds = None defaults to the nine Optional fields on LiteLLM_ManagedVectorStoresTable, making old cache entries (written without those keys) still deserialize correctly after the upgrade.
• Tests (test_cache_codec.py): existing assertions are updated to match the corrected behavior, and three new regression tests cover the required-nullable round-trip and the LiteLLM_ManagedVectorStoresTable full round-trip.

Confidence Score: 5/5

Safe to merge — the change is minimal, well-scoped, and the two complementary fixes together ensure both old and new cache entries deserialize correctly.

The serialize change is a one-line mechanical removal of exclude_none=True applied consistently across all four call sites. The defaults added to LiteLLM_ManagedVectorStoresTable make the fix backward-compatible with entries already in cache. Test updates correctly reflect the new (fixed) behavior rather than masking a regression, and three new regression tests specifically exercise the previously broken round-trip scenario.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/proxy/common_utils/cache_pydantic_utils.py	Removes exclude_none=True from all model_dump calls, making serialize/deserialize a lossless round-trip that preserves None-valued fields as JSON null instead of silently dropping them.
litellm/models/managed_files.py	Adds = None defaults to the nine required-but-nullable Optional fields on LiteLLM_ManagedVectorStoresTable, making them genuinely optional and enabling backward-compatible deserialization of old cache entries that were serialized without those keys.
tests/test_litellm/proxy/common_utils/test_cache_codec.py	Updates two existing assertions to reflect the corrected behavior (None fields now preserved, not dropped) and adds three regression tests — including a full round-trip for LiteLLM_ManagedVectorStoresTable with all optional fields null.

_{Reviews (1): Last reviewed commit: "fix(proxy): stop CacheCodec dropping nul..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!