Repository navigation
Logging: Reduce INFO log noise #1243
Description
Describe the current behavior
The backend ships every INFO log to Sentry Logs. Over 7 days the kaapi-production project received about 335k INFO logs. A large share is routine per-request or per-task chatter that is never read and adds cost and noise when debugging.
Unique logger.info call sites in the codebase (excluding tests and migrations), by module:
| Module | Files | info | warning | error/exception |
|---|---|---|---|---|
| Knowledge base (collections, documents) | 31 | 59 | 34 | 34 |
| Assessment (API-client + submissions) | 11 | 21 | 17 | 14 |
| LLM call, chain, config & version | 29 | 53 | 105 | 50 |
| Evals 1.0 only (batch, Langfuse, embeddings) | 8 | 42 | 31 | 22 |
| Evals 2.0 only (judge, iteration loop) | 14 | 4 | 9 | 6 |
| Evals shared by 1.0 + 2.0 (core, cron, fast, dataset) | 21 | 47 | 22 | 21 |
| Onboarding (credentials, project, org, user-project) | 12 | 25 | 11 | 9 |
| Analytics (analytics + daily stats) | 5 | 0 | 1 | 0 |
Whole app: 558 logger.info call sites.
Actual production INFO volume in Sentry, 7 days, by module:
| Module | INFO logs | Main loggers |
|---|---|---|
| LLM call / chain | ~77,800 | services.llm.jobs 50.7k, crud.llm 16.9k, OpenAI provider 6.5k, Google AI Studio provider 3.7k |
| Evals (almost all shared code) | ~15,000 | crud.evaluations.cron_utils 13.4k, crud.evaluations.core 0.6k, services.evaluations.evaluation 0.5k |
| Assessment (API-client + submissions) | ~6,500 | crud.assessment.api 4.1k, services.assessment.api.batch 1.4k, callbacks / submission / submission_store 0.3k each |
| Knowledge base | ~1,500 | collection_job, crud.rag.open_ai, create_collection |
| Onboarding | negligible | |
| Analytics | 0 |
Cross-cutting loggers outside these modules, also worth reviewing:
| Logger | INFO logs | What it is |
|---|---|---|
http_request_logger |
86.7k | one access line per request |
app.celery.celery_app |
18.5k | [pool] status before and after every task |
app.api.routes.cron |
10.8k | cron polling every 5 minutes |
celery.worker.strategy |
10.4k | Celery "task received" |
app.celery.tasks.job_execution |
9.2k | task wrapper |
app.utils |
8.6k | mostly callback success lines |
app.celery.utils |
8.4k | enqueue lines |
app.core.langfuse.langfuse |
8.2k | tracing client |
Describe the enhancement you'd like
Go module by module (one PR per module) and for each INFO call site:
- Keep it if it marks a state change someone would search for (job created, job finished, provider called, callback sent).
- Downgrade to DEBUG if it narrates internal steps of a single request or task.
- Remove it if it duplicates another log, a span or a metric (for example per-task pool status, per-request access lines already covered by HTTP metrics and spans).
Suggested order by volume: LLM call/chain, cross-cutting Celery/cron loggers, Evals, Assessment, Knowledge base, Onboarding.
Why is this enhancement needed?
Lower Sentry Logs volume and cost, and faster debugging because the remaining INFO lines are the ones that matter.
Additional context
- Assessment scope is the API-client BATCH path and submissions only (
api/routes/assessment/api.py,services/assessment/api/,crud/assessment/api.py,crud/assessment/submission.py,services/assessment/submission.py). The legacy RUN pipeline (cron poller, UI routes,services/assessment/{service,stages,tasks}.py,crud/assessment/{cron,processing}.py) is outdated and out of scope. - Out of scope here but found while measuring: 46k
uvicorn.accesslogs inkaapi-productioncome from kaapi-guardrails development containers using the production DSN (health-check lines). That is a guardrails configuration fix, not a log cleanup. - Counts are from 2026-10-10; re-measure in Sentry Logs (group by
logger.name, filterseverity:info) before and after each module.
Activity
[-]Reduce unnecessary INFO logs, module by module[/-][+]Logging: Reduce INFO log noise[/+]