Commit graph

2 commits

Author SHA1 Message Date
c29f4aef57 fix(llm): cap серверного Retry-After через _MAX_BACKOFF_S=30s (#1209)
Some checks are pending
Deploy / build-frontend (push) Blocked by required conditions
Deploy / deploy (push) Blocked by required conditions
Deploy / build-backend (push) Blocked by required conditions
Deploy / build-worker (push) Blocked by required conditions
Deploy / changes (push) Successful in 7s
В complete-loop'е min(...,30) применялся ТОЛЬКО к exponential backoff'у
(else-ветка). Серверный Retry-After уходил в time.sleep как есть.
_parse_retry_after принимает любое число секунд ("86400".isdigit() → True),
а provider'ы шлют Retry-After до 86400с при quota-exhaustion (OpenAI,
CDN-503). Эффект: time.sleep блокирует поток anyio-threadpool на часы
(до 48ч при llm_max_retries=2). Async-консьюмер (chat.py) мостится через
run_in_threadpool — поток держит токен пула (~40 потоков, общий с sync
Depends(get_db)) и DB-сессию → пул исчерпывается → стопор приложения.

Patch: вынес кап на module-level _MAX_BACKOFF_S=30s, применяю к ОБЕИМ
веткам (Retry-After И exp.backoff). raw_wait логируется отдельно для
наблюдаемости (видно когда провайдер просил больше).

Новый юнит-тест test_rate_limited_retry_after_capped: provider шлёт
retry_after=86400 → time.sleep вызывается с 30 (не 86400). 14/14
client тестов + 55/55 LLM-suite зелёные. ruff clean.

Closes #1209
2026-06-13 06:07:46 +00:00
4af7ba5a40 feat(llm): foundational LLM infra package with §19 redaction + deterministic fallback (#960)
All checks were successful
CI / changes (pull_request) Successful in 5s
CI / backend-tests (push) Successful in 6m25s
CI / backend-tests (pull_request) Successful in 6m20s
Deploy / changes (push) Successful in 5s
Deploy / build-frontend (push) Has been skipped
Deploy / build-backend (push) Successful in 1m38s
Deploy / build-worker (push) Successful in 2m42s
Deploy / deploy (push) Successful in 1m13s
CI / changes (push) Successful in 5s
CI / frontend-tests (push) Has been skipped
CI / frontend-tests (pull_request) Has been skipped
Optional external-OpenAI layer over the deterministic forecasting engine. Gated by
llm_enabled (default False) so prod makes no network calls until deliberately enabled.
Allowlist-first SafePayload contract + is_confidential hard-block + RU-PII regex scrub
(mandatory on the external path). Abstract LLMProvider seam (is_external) for a future
RU-hosted provider. Sync httpx core (Celery-friendly); tool/function-calling pass-through;
timeout + bounded 429/5xx retry + per-request call cap, all degrading to fallback.
Raw httpx (no openai SDK -> no pyproject/lock drift). 47 tests, ruff + mypy clean.

Refs #960
2026-06-08 15:44:16 +05:00