-
[NA] [BE] Fix silently-dropped demo experiment items in prod (#6454)
frostbyte_neo 发布于
2026-04-23 12:39:28 +00:00 - [NA] [BE] Use sync REST for experiment items to fix silent drops in prod
experiment.insert() goes through the SDK's async streamer queue — the actual
POST only happens on client.flush(). In production, flush() can return True
while silently dropping messages on timeout, leaving the experiment with 0
items in the UI even though the "Seeded experiment" log fires.Switch to client.rest_client.experiments.create_experiment_items(...) which
is a direct synchronous REST call that raises ApiError on non-2xx, guaranteeing
the items land before we log success. All other writes in this flow already
use the sync REST path.Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com
- [NA] [BE] Pin project_name on experiment items to avoid CH ingest race
The backend's create-experiment-items handler resolves project_id via two
paths (ExperimentItemService.java):- If project_name is set — resolve via projectService.retrieveByNamesOrCreate
- Otherwise — look up each trace by trace_id and use the trace's project_id
Path (2) races with ClickHouse ingestion of the traces we POSTed milliseconds
earlier — in prod the trace lookup can return empty, leaving experiment_items
with project_id=null and invisible in the project-filtered UI.Pin project_name explicitly so resolution is deterministic and independent of
CH replica lag.Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com
Co-authored-by: Claude Opus 4.7 (1M context) noreply@anthropic.com
下载附件