問題文
A classic extract-transform-load workload keeps exceeding device memory inside individual tasks. Which spilling arrangement does the documented guidance recommend, and why?
選択肢
- Enable the accelerated table library's own spilling, because it can move individual device buffers while a task is still running, whereas the worker-level mechanism tracks whole task outputs.
- Raise the worker-level device memory limit so that the threshold is reached later, because the mechanism can then release the intermediate buffers created inside a task at the moment the limit is crossed and the task will therefore complete without the peak ever being exceeded.
- Increase the memory pool fraction on each worker.
- Disable spilling entirely so that failures happen early and the partition size can be tuned instead.