Register the inflight_finalizes metric

[aarongable]

No description provided.

[aarongable]

We seriously need a way to guarantee that this bug never happens again; it has bitten us at least three times in the last year.

[pgporada]

There's some interesting discussion regarding ensuring all new metrics get registered over here (for a custom registry at least, but still some ideas I guess).

[aarongable]

That thread actually leads to a potential solution here: instead of having statsAndLogging return a prometheus.Registry, it could return a promauto.Factory, which exposes all of the usual NewCounterVec etc methods, except that they auto-register the metric when it's created.

[beautifulentropy]

That thread actually leads to a potential solution here: instead of having statsAndLogging return a prometheus.Registry, it could return a promauto.Factory, which exposes all of the usual NewCounterVec etc methods, except that they auto-register the metric when it's created.

I like this. Lexical confinement == best confinement.

[aarongable]

Unfortunately I did some digging and it would be pretty hard for us to use promauto.Factory, because we do lots of stuff with the registries that we pass around. Like, we don't only register individual metrics on them, we also register whole collectors (e.g. the whole grpc-go auto collector), and promauto.Factory doesn't expose the underlying registry at all. So we'd have to either completely overhaul when/where we register various third-party collectors with the registry, or we'd have to be passing around both the registerer and the factory, and that seems unfortunate.

Anyway, it's certainly worth digging into further, but I didn't see a nice clean improvement in about an hour of digging :(