Why seastar::shareded asserts instead of do its job?

[kelbon]

Why destructor seastar::sharded asserts about it deinitialized, instead of blocking deinitialize itself on all shards?

• send tasks to deinit on each shard
• wait all complete

Otherwise RAII breaks

[xemul]

Co-await-ing in destructors is not allowed by standard

[kelbon]

Co-await-ing in destructors is not allowed by standard

you dont need co await to blocking wait while all other threads are done

[xemul]

You wrote:

Why destructor seastar::sharded asserts about it deinitialized, instead of blocking deinitialize itself on all shards?

I read it as -- "why can't seastar::sharded::~sharded() block on de-initializing other cores' instances". Did you mean something else?

[kelbon]

You wrote:

Why destructor seastar::sharded asserts about it deinitialized, instead of blocking deinitialize itself on all shards?

I read it as -- "why can't seastar::sharded::~sharded() block on de-initializing other cores' instances". Did you mean something else?

1. i need to destroy N objects on N threads.
2. I destroy my object on my thread
3. I push N - 1 tasks to destroy to other threads
4. I wait while N - 1 task are not done.

[xemul]

In practice, the caller of sharded<>::stop() wants to know when exactly the instances on other cores are stopped and freed before proceeding. It's probably possible to do what you describe, but there's no use-case for it. You can achieve that with external RAII-sh wrapper that would call (void)sharded_service.stop() in its desctructor.