Persistent Task Group

[achimnol]

Let's add a persistent task group in addition to the current aiotools.TaskGroup.

While existing TaskGroup targets to group and ensure termination of a set of tasks with a short period of lifetime (e.g., a single operation that is composed of many async tasks), PersistentTaskGroup targets to group and ensure termination of a continuously changing set of tasks that are created and terminate through a very long time (e.g., the entire process lifecycle), with proper cancellation and exception handling upon shutdown.

One key difference to TaskGroup is that it has its own fallback exception handler because gathering the results (and exceptions) of all tasks at once at the point of "exit" as MultiError is meaningless due to sparsity across time. Custom fallback exception handlers could be implemented using decorators (with functools.wraps() to keep the function signature consistent with the original coroutine).

To automatically detach terminated tasks from the tracking set, we could use weakref.WeakSet.

In fact, the above pattern is widely used in most Backend.AI components, and adding PersistentTaskGroup will reduce boilerplates for Backend.AI.

[asvetlov]

/cc @1st1
Yuri, you might consider this mode also when designing the asyncio TaskGroup class.

[achimnol]

To aggregate the results of belonging tasks, we could use a separate async-generator channel to subscribe the results/exceptions of tasks finished through time, though this feature would be completely optional as the main use case for PersistentTaskgroup would be server applications where most tasks would return just None while serving their results to some I/O channels such as client connections.