Add Phoenix.PubSub.subscribe_once/3

[nathanl]

Given that the possibility of duplicate subscriptions was already discussed in the docs on subscribe/3, I assume there's a performance cost to checking for existing subscriptions that callers may not always want to pay. If so, I'd be happy to document that on subscribe_once/3. But it seems useful to have this function; I'm looking at a use case where the number of subscriptions to a topic won't be large but duplicate messages would be problematic.

Currently, any caller who wants to avoid duplicate subscriptions would need to use some kind of workaround, such as:

• Avoid calling subscribe/3 more than once by tracking subscriptions in their own process state, or by checking Registry.lookup/2 directly (relying on an implementation detail of PubSub)
• Call unsubscribe/2 each time before calling subscribe/3

[davydog187]

Unfortunately, I believe there is a race condition here. In betwen the Registry lookup and the else path, another process could have subscribed.

The only way this could work is by doing an insert_new operation to :ets

[josevalim]

Since you can only subscribe yourself, it cannot be a race condition. However, because this is a duplicate registry, there is basicaly no cheap way to check if you are already subscribed. If you have a topic with 100k entries, then this function will be very expensive because you are bringing a 100k elements list to memory, which is why we don't provide this functionality, it is basically an anti-pattern. So it is more scalable to guarantee code wise you won't subscribe multiple times. :)

[nathanl]

a topic with 100k entries

Meaning 100k subscribers?

[nathanl]

Given what you said, I'd like to do a followup docs PR to explain why the caller should guard against multiple subscriptions

[josevalim]

correct!

[josevalim]

That would be lovely!!! ❤️

[josevalim]

Closing per the above. :)