Vorner's random stuff

Extracting tracing metrics to dipstick

While I enjoy programming, there are certain aspects that are less fulfilling than others. When writing a „business“ service, there’s quite a lot of distracting boilerplate besides the actual useful code. I’ve tried to cut down on that previously with the spirit family of libraries, because I didn’t want to handle reading configuration, logging setup, etc every time a new service was started.

But this one is not about spirit. Some time ago, I’ve noticed the appearance of the tracing crate and was wondering since how much traditional functionality could be derived from that single instrumentation of a program. Logging, certainly. But I was more annoyed by the need to gather metrics through the program. And I was specifically annoyed by the fact that I’ve had a log message near each metric collection point ‒ if I want to count something, it’s probably important enough for it to appear in logs. So, shouldn’t it be possible to write that just once?

Annotating the tracing statements

The tracing crate offers similar macros as the log crate, like info!. It also adds macros to create so called spans ‒ the „normal“ macros are one-off ‒ they just happen (they are called events in the tracing terminology). But spans have a start and an end. They also nest into each other (well, they can do some more crazy things for advanced tracking of what’s happening in a program and why, but that’s not needed here).

Events can be used quite naturally to collect counters, they can contain values for gauges, etc.

The spans are ideal for timing things, for nesting metric naming scopes and for adjusting levels (eg. how many requests are being handled right now in parallel).

The only problem is, how does the bridge (let’s call the thing that takes tracing information and turns that into traditional metrics that can be put somewhere into graphana a bridge) know which events and spans to consume?

What I’ve come up with is placing an attribute onto the span or event macro call. The bridge can look at the attribute and decide what to do about it. It has a small downside, though ‒ these are more like instructions for the bridge than human-useful information directly, but they still end up in logs or other places these things travel to.

let _span = info_span!("Handling yak", metrics.timer="yak").entered();
info!(metrics.counter="shave", "Shaving yak")

These would provide the timing information of the whole handling of each yak and count how many yaks the program shaved.

I think this can be used even in combination with the instrument procedural macro to lower the amount of typing even more (but I didn’t have opportunity/time to play with that one yet).

The dipstick bridge

Eventually, I’ve decided to roll up my sleeves and actually try writing a bridge. I’ve used the dipstick metrics library previously, so I’m using that one.

The crate is published as tracing-dipstick, with an example how to set it up and how to use it (with the list of supported attributes). It’s quite early, so there’s a chance there could be some bugs. There are places where I believe the overhead could be lowered. Some features might be missing (for example, one might want to also measure the time an asynchronous task spends doing something vs. waiting).

There’s also one problem with filtering. The crate sits on top of tracing-subscriber, which is the thing that allows composing multiple tracing “backends” together. It also allows for filtering of the tracing information by severity. The issue is, it does so globally ‒ for all the backends. While logging backend wants them to be filtered, this bridge likely wants to consume all the events and spans ‒ even if one doesn’t want to log each request the service handles, it wants to count them.

This may be solved in a future version of tracing-subscriber or maybe worked around in a future version of the bridge. For now, it is possible to combine logging and metrics by using the logging-full feature of tracing ‒ so logging doesn’t go through the subscriber and is routed directly to the traditional log crate.

Anyway, that looks as a good enough first step for trying the experiment out and figuring if it makes sense to extend or not.

Compatibility with other bridges

I’ve used the dipstick mostly because I’ve already worked with it before (and because spirit knows how to configure it).

But the names of the attributes are quite general. It is conceivable that a different bridge for different metrics backend will be created (by me or others, as the need arises). In such case a whole application could be migrated to a different backend by replacing just the initialization part, not by going through all the instrumentation of the whole code. I guess it’s another instance of „There’s no problem that couldn’t be solved by another layer of abstraction“ thing.

Desirable improvements

I’ve identified few areas that are annoying at best.

The first one is the problem with filtering, described above. That one is already known and tracing-subscriber really wants to have an answer to that, long-term.

The other one is, it would be nice to somehow hide the metrics.* attributes from the logs. But it’s currently not possible to filter individual attributes, only whole events and spans, as they travel through the subscribers. I’m not sure if such feature is generally desirable, though, or if there’s a way to implement it in an effective way.

The last thing that comes to my mind is ‒ if one makes a typo in the attribute name and writes something like metrics.coutn or metric.count, there won’t be a compilation error. The attribute (and therefore the event/span) will simply be ignored by the bridge and nothing happens. It could be solved in part by checking the attributes for metrics.* and reporting the ones that are not understood at startup, but that’s still not great.

The future

So, what happens now? I want to experiment with it a bit more and see if it is usable. I have a good first feeling from it, because it cuts down on the part of programming I don’t like ‒ all that instrumentation boilerplate thing.

It’s another crate that I have to maintain now, but this one seems to be quite small so I hope its maintenance overhead would be low.

But I’d also like to have feedback from others. If you have a use case, or just some time to play with something, giving it a try and coming up with ideas for better interface, finding some of the bugs, etc, is welcome :-).