I like the end result of OpenTelemetry tracing when using Axiom and the like, but the SDKs have been a nightmare. Too much emphasis on automatic instrumentation, Java-isms, everything is stateful and abstracted away.
It can do distributed tracing of otherwise traditional long running microservices, but breaks down when your functions are distributed like in durable execution engines, Cloudflare Workflows, “functions” that span hours/days/weeks and steps that retry many times.
I had to reverse engineer how SDKs work and how tracing UIs display data so I could make simpler functions that fit wider variety of runtimes and more freely parent spans, start spans and end them from different function instances.
I think most of the API and terminology complexity is self inflicted. Would love to see a rebooted developer experience that is less Kubernates-brained.
What always puzzles me about OpenTelemetry is that tracing, metrics and logs are all designed independently. I wish there was a way I could just annotate my code base once, and let the ultimate decision to expose something as a metric/log/trace be dynamic at runtime.
For example, if I look at a graph in monitoring dashboard and see something suspicious, I’d like to say: “The next time something like this occurs again, please save me a trace.” I should be able to just do that with a single mouse click.
I remember them releasing the tracing spec/SDKs and saying “now let’s move on to metrics/logs.” That never sat right with me.
I think it is almost a inevitability where otel came as a standardised aggregate of OpenTracing (which was the same but only for tracing over multiple tracing implementations), logging, and metrics into a single observability standard without alienating all the individual supporting vendors.
Historically, logging and metrics have been different problem domains with different implementations for ages.
Now to your point:
Note that tracing does get the most of love, and that it does include constructs to add logging and metrics into these traces (spans actually). So you could argue that they are trying to develop a single interface.
> “The next time something like this occurs again, please save me a trace.”
Well, if you want this you either need to propagate this predicate to all points that might be involved, or always emit all traces and have the predicate included in the filter. And then you need to be able to dynamically propagate this predicate from the system/ui where you click to where you filter.
This is one of the reasons why we always propagate and emit traces and just post filter it in processing before it lands in the persistence layer.
I don't think OTEL is necessarily "at fault" here. It's a split that's carried all throughout the observability ecosystem. e.g. in the Grafana suite of solutions you have Loki (logs), Tempo (tracing) and Mimir (metrics) to cover storage & querying for all three axis, as all of them have very distinct processing & performance characteristics.
While it may intuitively may look like there is a large overlap in the three areas there is suprisingly little, and for the few parts there are (e.g. trace <-> log correlation), OTEL does offer a standard.
I just don’t get this sentiment. How would you represent metrics as traces? You cannot. Even reconstructing traces from logs would be challenging at best. How would you get, say, Garbage Collector metrics from logs or traces? You cannot.
There is no magic bullet. Observability isn’t something you can just slap on and call it a day. While traces and logs might share superficial similarities, they are not the same. And metrics are something else altogether. Trying to somehow unify them would be a prime example of "wrong abstraction".
> “The next time something like this occurs again, please save me a trace.”
The building blocks for this exist. The observability platform must simply (haha) implement the pattern detectors and use them for sampling decisions.
I am not sure if this is what they mean, but e.g. with Micrometer in Java you can instrument your code once with observations that produces observation events, then you can register handlers that can turn them into metrics, or logs, or traces without having to instrument your code three times.
Technically, you can use the same places in the code where you stop/start/fork traces to also be the places where you increment the counters/gauges, etc. Which I think the GP was alluding to when describing the micrometer solution. Similarly, you can derive metrics for log lines without having to emit the actual log lines.
Then separately you can have log levels or verbosity levels that control to which level you actually emit traces/logs and/or roll up metrics.
At that point you almost might as well just log everything. The decision logic is likely about as complex as just doing it. Then I suppose you have a watchdog task that fires off every, say, 15 minutes or an hour or something, looks at the collected data, and either decides to keep it or trash it while recording a tiny "nothing interesting" datapoint.
In the code define everything as a span with a name, scope (start-end), description and tags... and then you can easily dynamically produce traces, spans, logs or metrics based on what you need.
Tracing is the most general of them, and the most expensive unless you're careful with the implementation.
Trace spans are time-delimited units of "stuff that happened", with a tree relationship among the spans, and each span can have arbitrary tags (key/value pairs) and events (time/value).
From that, if you chose, you could derive metrics and logs. The trick is to start with tracing and to actually put it in your program, rather than trying to mostly-automatically tack it on later.
If I understand that correctly, it means your app always creates traces, and Grafana Cloud is responsible for sampling/aggregating. That may be prohibitively expensive in terms of CPU/network load.
What I’m suggesting is that your apps by default only send metrics to your monitoring system, but that the monitoring system can specifically ask to “upgrade” metrics to traces. Or to log entries.
The same thing with metric cardinality: by default, only report metrics in a fully aggregated manner. But do tell the monitoring system how they can potentially be broken up if needed (i.e., which labels to add).
You're pitching a solution that's incredible brittle and unnecessarily complicated if you think about it in technical terms.
For your feature to work you need bi-directional communication between the otel receiver and your application - that's still doable in general, but now you want a synchronous "upgrade" to traces.
Now we're talking about a massive performance impact - and you need to somehow cache all otel data locally so they're available for the upgrade and only then submit then.
It is a architecture that's not very smart, honestly. And precisely the reason why you'd simply submit everything and let the receiver figure out which samples it wants to keep - as thorian pointed out earlier.
> The same thing with metric cardinality: by default, only report metrics in a fully aggregated manner. But do tell the monitoring system how they can potentially be broken up if needed (i.e., which labels to add).
How does the monitoring system have any of the context to add labels? That would only exist in application memory.
> That may be prohibitively expensive in terms of CPU/network load.
In practice I've not experienced this even on quite high request rates. While it isn't free, exporting everything has been cheap enough that the real cost in dollars spent is basically marginal (it's _storing_ the data that's expensive)
> How does the monitoring system have any of the context to add labels? That would only exist in application memory.
Indeed. If you have a protocol that doesn’t allow exposing that kind of information, then that only lives in application memory. But my suggestion is that it’s exposed.
> If I understand that correctly, it means your app always creates traces
Yes, because otherwise what you propose requires modifying the binary in-place and that's too big of a security hole for lots of (production) environments. Some variants of that could work with an out-of-process method like Dtrace or eBPF, but that means mutating the kernel, even more of a no-no.
Its a shame that the various implementations are pretty horrible. Global state, static methods etc etc.
If you get rid of that, and just pass dependencies around, create some appropriate local abstraction around them.. the tooling, be it datadog or honeycomb does a great job making it useful. Can't really say the same for grafana, but ymmv - depending on budget
OTel is so frustrating. If it wasn't shaping to be the clear winner in the space, I wouldn't complain about it as much. But today:
1. Every major vendor is still in some weird alpha/beta support for OTel even after all this time.
2. The performance hit is substantial and makes you question what the point of performance instrumentation is if you need twice as much compute/RAM to run the same workload now.
3. Serverless runtimes pay a heavy penalty for cold starts with OTel.
4. You're basically forced to run both gateway collectors and edge collectors for any realistic usage.
5. You still need to configure destination exporters in unique ways. This leaves you questioning what the value of OTel was.
6. Vendors that go beyond the scope of what OTel covers still need their own bespoke instrumentation. What was the point of any of this then?
> 4. You're basically forced to run both gateway collectors and edge collectors for any realistic usage.
You most certainly don't. You can run your app (especially if it's "serverless") without the collector agent.
App-to-agent and agent-to-sink use the same protocol, so all you need to do is set up the tracing/logging/metrics exporters to directly speak with the sink. These days, it typically means specifying the URL and the DSN header.
Perhaps there's a gap in my understanding. Can you clarify on this a bit more? I run a mix of serverless and non-serverless workloads.
Gateway collectors are unavoidable because various SaaS platforms require you to be running publicly reachable endpoints to send telemetry to.
In a runtime like Lambda, how would you avoid the need to run an edge collector? The only thing that comes to mind is to write to logs and then have a log stream processor that then writes to your gateway collector. Other than that, it seems unavoidable, no? Sure, in something like Fargate you could go app to sink. But even that has its own tradeoffs.
We use Node.js, so all we need to do is run a script initializing Otel before running the app. We set this up following the docs a few years ago, and haven’t had to change it much since then.
A typical setup is to run a separate OpenTelemetry collector process on the same host as the app. The app connects to it via localhost on a standard port (although you can override it using env vars).
The collector process then sends the metrics/traces/logs to the observability sink. But there's nothing at all preventing you from sending telemetry directly to the observability sink.
It's just outbound HTTP or GRPC, and it doesn't have to go over public Internet.
> In a runtime like Lambda, how would you avoid the need to run an edge collector?
Here's my setup (in Go, very simplified):
> // Instantiate a new slog logger
> logger := otelslog.NewLogger("root", otelslog.WithLoggerProvider(otelLogger))
> // Use the logger as needed
My code uses proper Go loggers exclusively. I also redirected the stdout and stderr to a goroutine (via the usual close(2)+open() trick) to serve as a catch-all sink for anything that slips the net.
There isn't really a great alternative without vendor lock-in. If you go all-in on AWS Cloudwatch/X-Ray, it's a really easy setup with low effort. If you go all-in on Datadog, it's pretty easy. But if you want to mix Sentry, Langfuse, Datadog, etc, OTel is still probably the best option. It's just a letdown that this is the best there is.
I don't mean to disparage anyone working on OTel. I can appreciate that it has ambitious goals and it's not an easy problem to get alignment and interop here. Especially with all the stakeholders involved. But as a user, it feels simultaeneously over-engineered and under-engineered.
OpenTelemtry is the perfect example of an overengineered mess.
While I usually think that at least having some standard that people agree on I think OpenTelemtry should be dropped.
A lot of the less popular alternatives (just going with Prometheus, Victoriametrics, etc) are de-facto competing smaller standards and a lot better both in terms of less added complexity and the results you get.
I think OpenTelemetry turned metrics into a farce. In many situations even self-rolled telemetry works better even with the added stuff. The annoying thing is that OpenTelemtry is that big standard now one kind of has to to add compatibility. So please, if you write software, make sure you don't lock yourself into OTel.
I disagree. I'm an observability geek, and OTel is... fine.
It's missing a few things that I'd like, but I was able to implement them myself. I guess the major design issue is that the sampling decision is made at the _start_ of the segment. So I hacked up a few improvements:
1. Ability to mark segments as "boring", so they are dropped before the export. For things like healthchecks, empty "get the pending jobs" queries, etc.
2. Ability to downgrade errors for segments that are expected to return an error (e.g. HEAD on a non-existing object in S3 to check if there's a cached blob).
I understand the author's perspective in the linked article, but none of that data shows a project in trouble? Some languages have more resources than others, but those all look like healthy open source projects
Every time I share your blog (and I share it a lot) I tell people:
"This guy started a blog in 2024. Wrote three posts and all three of them would still make my top ten list of 'greatest posts on observability' today".
'A practitioner's guide to wide events' especially is still my number 1.
I'd make a wager that things would go better smoother faster if folks tried more stuff, ventures forth more on their own. It's obviously not great that there's no semantic convention that's perfect and just works for everything, and yeah it takes a while. I feel like the real data I'd want is who else, how many people show up to say they've tried something. Is that happening? Whether specs are really good enough advance or not, to me, is often whether enough people have tried it to find out.
The net of this is, otel is a very flexible system you can use and adapt in all kinds of ways and while the spec is important, using the toolkit to FAFO yourself, ahead of any beaten path, should really be encouraged. That's the message I'd want to see being radiated out about otel.
The alternative is vendor lockin, $$$, and spotty support for complex environments with zero chance of ever getting 100% coverage.
At least with Open Telemetry, anyone can write an OTLP "source" using free, open specifications, and it'll "just work" with dozens of third-party "sinks". That's huge!
Sure, there's a lot of experimental tags on semantic conventions, but at the end of the day, that's not that critical. It's just data: most sinks don't "interpret" these tags, they just display them as-is, so changes aren't breaking changes.
The alternative is Prometheus (which is freaking great) and Jaegar (which is freaking great), each alone. This is better, because Otel is trying to put two distinct things (monitoring and metrics, distributed tracing) into one package, because they know how to use neither.
Neither Prometheus metrics nor Jaeger traces are magic bullets. Neither of them are complicated, either, and in fact the fact that they're not complicated is their greatest strength. You can and should understand every facet of what they entail. You should build the (very small) shims that they need for your company's framework every time. It's not hard. It's not hard because it's not complicated. The fact that it's not complicated seems to break people's brains. They are accurate because they're simple and they're easy to work with because they're simple, and OTel is neither.
Prometheus is so easy to add and if you need more scale, there is mimir and a few other options with similar client semantics. I really can't imagine reaching for a framework APK that tries to anticipate every possible thing I would want telemtered, and is inevitably missing all the domain specific derived channels I need. Even prepackaged Prometheus exporters are usually overkill.
Hard Agree on Prometheus. And esp on the complexity - OTel is dizzyingly complex. You can get started ASAP on Prometheus whereas you get lost in analysis-paralysis when dealing with OTel.
It is crazy to me how often people don't grok how to design software well.
1. The worst thing you can do is try to stuff too many things into one specification. So you want an API? That's great. What's that? You want a rigid set of types so that any tiny changes over time aren't compatible? You want to try to define every conceivable use case as a new call? You want to combine multiple elements from different domains into one flat set of functions? You don't have any hierarchy or inheritance? You don't support extensions?
2. The second-worst thing you can do is to force a whole lot of different people to go through a single standards body. So you want to support a thousand different 3rd party components. What's that? You want to require everyone get their adapter approved by one group? And there's only one supported adapter per 3rd party component?
If you're trying to feed an entire city, it's logistically incredibly difficult to try to do it all yourself. If instead you just define where food can be dropped off or picked up, and ask volunteers to bring their own food there whenever they can/want, now you don't have a logistical nightmare on your hands anymore. The tech alternative? Add support for "plugins", make the plugin interface incredibly loose/backwards-compatible/layered, and invite people to publish their own plugins. If you under-engineer it, it actually works better.
It can do distributed tracing of otherwise traditional long running microservices, but breaks down when your functions are distributed like in durable execution engines, Cloudflare Workflows, “functions” that span hours/days/weeks and steps that retry many times.
I had to reverse engineer how SDKs work and how tracing UIs display data so I could make simpler functions that fit wider variety of runtimes and more freely parent spans, start spans and end them from different function instances.
I think most of the API and terminology complexity is self inflicted. Would love to see a rebooted developer experience that is less Kubernates-brained.
reply