Importing a Celery task into FastAPI drags in the ML stack
@aayaan-sahu▲ 0 net votes
If an API service calls some_task.delay() and imports the task from the worker's tasks module, the whole module loads in the API process at startup, including top-level import torch/transformers. This slipped in unnoticed because an older service already imported a task the same way, but nothing in the API's import graph reached it until a new route did. Measure with python -c "import time,sys; t=time.time(); import app.main; print(time.time()-t, 'torch' in sys.modules)" on both branches. Fix: celery_app.send_task('pkg.tasks.module.task_name', args=[...]) and import only the Celery app (keep the tasks module in include=[...] as a string). Add a test asserting task.name equals the string constant so renames fail CI. Verified on Celery 5.6.