Multiple issues with unsubscribe

[gbrew]

Version
All crates are v2.0.10 as of commit e0e79df

Platform
Macos Darwin Kernel Version 22.6.0, Ubuntu with kernel version 5.15.0-86-generic

Description

There are multiple issues preventing ethers unsubscription from working. The end result is that the rpc server does not get proper eth_unsubscribe calls, so the subscriptions are not cancelled on the server. This leads to problems if the server enforces a maximum number of live subscriptions per connection.

For example, this code will never unsubscribe and eventually run out of subscriptions:

   loop {
        let stream = provider.subscribe_blocks().await.unwrap();
        stream.unsubscribe().await.unwrap()
    }

This also happens when relying on the drop() implementation for SubscriptionStream to clean up when a SubscriptionStream goes out of scope.

There are several distinct issues here:

1. SubscriptionStream::unsubscribe(id) calls Provider::unsubscribe(id). This makes a properly formatted eth_unsubcribe json-rpc call to the provider, but with the ws PubsubClient, the id in the SubscriptionStream is a client-side one which does not match the server-side ID, so the wrong ID is passed and the call fails.
2. SubscriptionStream::drop() calls PubsubClient::unsubscribe(id), which maps the IDs properly to the server-side one, but incorrectly formats the eth_unsubscribe json-rpc request by passing params as a string rather than the required list, so that request fails as well (and prints a warning if using reth as the RPC server).
3. The IPC PubsubClient doesn't send an eth_unsubscribe call when PubsubClient::unsubscribe() is called, so the subscription is removed on the client side but not the server side.
4. The ws-legacy PubsubClient has the same issue as in point 3.

I've observed these issues for the ws client by watching the connection in wireshark.

I am putting together a pull request to deal with all of the above by making all unsubscribe calls (both explicit via SubscriptionStream::unsubscribe() and implicit via SubscriptionStream::drop()) call through PubsubClient::unsubscribe(), fixing the ws implementation of that to properly format the rpc request, and adding eth_unsubscribe calls in the IPC and ws-legacy clients. Let me know if this approach makes sense and I can put that up for review soon.

[mattsse]

@gbrew thanks for digging

looks like you fixed that on your branch?
mind opening a PR for that?

[gbrew]

@gbrew thanks for digging

looks like you fixed that on your branch? mind opening a PR for that?

No problem -- I actually opened a PR for this right after the bug, but it looks like it didn't get linked here. Please have a look at #2670 . I'm noticing that several of the CI tasks failed, but from a quick look they seem unrelated to my changes.

[gbrew]

@gbrew thanks for digging
looks like you fixed that on your branch? mind opening a PR for that?

No problem -- I actually opened a PR for this right after the bug, but it looks like it didn't get linked here. Please have a look at #2670 . I'm noticing that several of the CI tasks failed, but from a quick look they seem unrelated to my changes.

@mattsse Who's the right person to ask to take a look at this? Would like to get it merged if appropriate, as I'm still seeing this issue and it looks like the alloy replacement isn't quite ready.