-
[OPIK-3754] [BE] Add Optimization Studio config defaults (#4756)
frostbyte_neo 发布于
2026-01-16 11:35:22 +00:00 - [OPIK-3754] [BE] Add Optimization Studio config defaults for LLM max tokens and RQ failure TTL
- Add OPTSTUDIO_LLM_MAX_TOKENS (default: 8192) to prevent LLM response truncation
when generating structured JSON outputs (fixes Hierarchical + JSON validation errors) - Add RQ_WORKER_TTL_FAILURE (default: 86400 = 1 day) for failed job retention in Redis
Tested locally with parallel optimization runs (5 concurrent workers):
- Confirmed parallel execution works correctly
- Memory usage remains stable (~200MB per worker)
- Max tokens default fixes truncation errors with Hierarchical optimizer + JSON validation
-
Update Helm documentation
-
Revision 2: Fix comment inconsistency (default 8192) and add env var usage in Python code
-
Revision 3: Defensive env var parsing and fix RQ 2.x handle_job_failure signature
-
Revision 4: Address PR review - DRY env parsing and narrow exception handling
- Add get_env_int() utility to env_utils.py for defensive env var parsing
- Use get_env_int() in metrics_worker.py and optimizers.py (removes duplicate try/except)
- Define RQ_WORKERS_SET_KEY and RQ_WORKER_KEY_PREFIX constants in rq_worker_manager.py
- Narrow exception handling in _cleanup_stale_workers() to redis.exceptions.ResponseError
and redis.exceptions.RedisError instead of broad Exception catch
- Revision 5: Fix flaky test_executor_with_logging_module test
The test was failing intermittently because it expected logs in a specific
order, but stdout and stderr are separate streams that can interleave
unpredictably. Changed to order-independent validation by matching logs
on message content instead of strict positional ordering.Also fixed pre-existing lint issues (unused imports, undefined logger).
Co-authored-by: CometActions github-actions@comet.com
下载附件