How severe does this issue affect your experience of using Ray?
- Low: It annoys or frustrates me for a moment.
I am performing a ray pipeline task. After loading the data, it is allocated to two actors on the GPU for model inference using the
map_batches method. Subsequently, the results of the model inference are allocated to an actor on the CPU for further post-processing, again using the
map_batches method. The number of CPU actors used for post-processing is controlled by the
ActorPoolStrategy , with
However, I have observed in the Ray timeline that the actors used for post-processing are frequently killed, and Ray initializes new actors for post-processing. Strangely, if I specify
min_size=max_size=8 while keeping everything else the same, no actors are killed at all. After examining the dashboard and logs, I don’t believe this is caused by an Out Of Memory (OOM) issue. This has been bothering me for several days, and I wonder why the actors will be killed.