Why

Stacked on top of #26838. We want to be able to answer three questions from a single CircleCI log:

1. Which Redis instance is each xdist worker actually talking to?
2. Are SET calls succeeding (or silently failing — noeviction will start refusing writes once memory fills)?
3. Which tests result in cache hits vs misses vs no recording at all?

Right now the persister logs nothing, so all three of those are guesses.

What

Two diagnostic surfaces, both namespaced under the litellm.vcr.persister logger and prefixed with [VCR] so they're trivially grep-able (grep '\[VCR\]' build.log):

1. log_redis_target_banner() — fires once per worker at pytest_recording_configure. Logs the resolved Redis URL (password masked), maxmemory_policy, used_memory_human, used_memory_peak_human, evicted_keys, and the current count of keys under the litellm:vcr:cassette: prefix.

2. _RedisPersister.{load,save}_cassette — one [VCR] line per call: hit / miss / persist / persist-failed, with key, payload size, and episode count.

Sample output

Smoke-tested against the actual CI Redis (redis-19853.crce262.us-east-1-1.ec2.cloud.redislabs.com:19853):

[VCR] target=redis://default:***@redis-19853.crce262.us-east-1-1.ec2.cloud.redislabs.com:19853
[VCR] server maxmemory_policy=noeviction
[VCR] server used_memory_human=7.31M
[VCR] server used_memory_peak_human=1.7G
[VCR] server evicted_keys=0
[VCR] existing vcr keys under 'litellm:vcr:cassette:': 0
[VCR] persist key=litellm:vcr:cassette:tests/llm_translation/diagnostic_smoke/test_one bytes=334 episodes=1
[VCR] hit  key=litellm:vcr:cassette:tests/llm_translation/diagnostic_smoke/test_one bytes=334
[VCR] miss key=litellm:vcr:cassette:tests/llm_translation/diagnostic_smoke/missing

That output already confirms two things worth knowing about the target instance independent of this change:

• maxmemory_policy=noeviction — once we hit maxmemory, every SET (VCR or otherwise) will start failing with OOM. There is nothing automatic that prunes our keys; the only pruning is the per-key 24h TTL set by save_cassette, and a successful SET resets that TTL.
• used_memory_peak_human=1.7G — the instance has filled in the past, almost certainly during a previous llm_translation run.

Tests

tests/llm_translation/test_vcr_redis_persister.py (the existing 16 unit tests) still pass.

Type

🚄 Infrastructure
✅ Test

Slack Thread