Collator protocol followup

[montekki]

Fixes #1694

[ordian]

Metrics-related comments.

[ordian]

I think &Registry is not always available, real_overseer accepts Option<&Registry>. we could do the same and call register function instead of try_register.

Wouldn't it be better to make ProtocolSide public and pass it here instead?

ProtocolSide could have two constructors:

impl ProtocolSide {
    pub fn validator(registry: Option<&Registry>) -> Result<Self> {
        ...
    }
    pub fn collator(id: CollatorId, registry: Option<&Registry>) -> Result<Self> {
        ...
    }
}

[montekki]

In the original PR there were grumbles agains it

[ordian]

But how was it resolved? We still are making a choice at startup depending on whether id is None or not?

[montekki]

Yep, I see no other options

[ordian]

Looks good, just a couple of questions.

[coriolinus]

I think that the recently_removed_heads thing is nice, though it necessarily limits us to precisely one update-duration of leeway for late blocks. As has already been discussed, we're implicitly assuming that we get precisely one of these updates per block, and I'm not certain that's a formal requirement for the overseer.

That said, I don't see a better way of dealing with this until we change the ActiveLeavesUpdate to explicitly name parent-child relationships, which would allow us to specify a particular duration to keep the leaves.

[coriolinus]

I don't really understand why we expose the Metrics type in the trait if it's ok to just fake it and use a different type internally. I assume this is fine as @ordian has already approved this PR, but it may be worth considering whether we could remove type Metrics from trait Subsystem.

If on the other hand there's a reason to publish this type, then we really should be distributing it via an enum.

[montekki]

bot merge

[ghost]

Trying merge.