Should language SDK generate service.instance.id?

[jack-berg]

The resource semantic conventions say the following about service.instance.id:

If the service has no inherent unique ID that can be used as the value of this attribute it is recommended to generate a random Version 1 or Version 4 RFC 4122 UUID (services aiming for reproducible UUIDs may also use Version 5, see RFC 4122 for more recommendations).

The language is ambiguous as it's not clear "what" / "who" is supposed to generate the random Version 1 or Version 4 UUID. Is it the application owner? Is it the language SDK?

There's a request in opentelemetry-java that we generate service.instance.id. We discussed adding it in the Java SIG and there was debate about whether this was actually the intent of the spec, and if it is, whether its a good idea. The argument against including it can probably be summarized as:

• Adding a random service.instance.id that is not persistent across restarts increases cardinality. While this is true I disagree that it is a problem, since I think the number of users negatively impacted would be small and they have options available to solve the problem.
• Adding a random service.instance.id that doesn't reflect the identifier of the instance in its environment (i.e. is different than the pod uid) may be confusing to users. I disagree that this is a problem because users should set service.instance.id to an inherent unique ID if available.

Maybe I'm missing part of the debate but in any case it would be good if the spec could clear up this ambiguity.

This may be a duplicate of #1034, but that issue's description appears to be asking about alternates for service.instance.id where I'm asking about what is specifically responsible for generating service.instance.id in the current wording of the spec.

[svrnm]

Adding my 2 cents to that:

I would prefer a solution where the SDK is providing the service.instance.id if it is not already defined by the end user (via environment variable, config or code). It's not something most end-users care about until it comes to observability where they want to say "give me all data for that one particular service instance" and then they figure out they have no unique identifier...

I also don't think it should be persistent across restarts, here's my rational for that: Right now if I have a highly scaled service with lots of nodes I can not identify which service is the troublemaker, except through other IDs that might be available if running on K8s (pod.id), some container runtime (container.id*) or a physical host (host.id), etc., but it's not their purpose to identify a service, which brings us back to a restart: if I (soft) restart a service within a container, the pod&container&host ID stays the same: So in the worst case I have a service that restarted a few times, but I can not uniquely identify an instance.

[Oberon00]

This may be a duplicate of #1034, but that issue's description appears to be asking about alternates for service.instance.id where I'm asking about what is specifically responsible for generating service.instance.id in the current wording of the spec.

I think that is mostly what the discussion on #1034 is about. Repeating my comment from there: #1034 (comment):

IMHO this attribute is poorly defined right now as it may or may not be the same across service restarts, which IMHO can make quite a difference. It would be easiest if it MUST be the different for each restart, that way it could be used as primary key for all resources (not only service.*) sent by the same telemetry instance. On the other hand, maybe such an attribute would better be named telemetry.sdk.instance.id.

[tsloughter]

As a datapoint, in the Erlang SDK we use the node name as the instance id resource attribute. If the node is run without distributed erlang enabled (so no nodename) we simply use a random integer as the id -- same as using the default otel trace id generator.

nodename is somename@IP or somename@hostname, so stays the same between restarts if the node is on the same ip/hostname.

[sirzooro]

I also don't think it should be persistent across restarts, here's my rational for that: Right now if I have a highly scaled service with lots of nodes I can not identify which service is the troublemaker, except through other IDs that might be available if running on K8s (pod.id), some container runtime (container.id*) or a physical host (host.id), etc., but it's not their purpose to identify a service, which brings us back to a restart: if I (soft) restart a service within a container, the pod&container&host ID stays the same: So in the worst case I have a service that restarted a few times, but I can not uniquely identify an instance.

I have opposite use case: I have exactly 2 instances of service X and 10 of Y. These instances are numbered (1-2, 1-10) and their numbers are statically assigned to them during deployment. It makes sense to put them in service.instance.id as-is, and use resource attributes from Kubernetes semantic conventions to store other details.

[Oberon00]

I think both an ID that persists over restarts and one that doesn't can make sense to have, for different use cases. Also, the former is harder to impalement while the latter is trivial. There should probably be two different attributes here.

[svrnm]

I think both an ID that persists over restarts and one that doesn't can make sense to have, for different use cases.
+1

Agreed, there should be 2 different attributes, the questions remains which one is then "service.instance.id" and which one is something else?

[jsuereth]

I'm actually in favor of having the SDK generate this. The main concern, and others have raised this, is that we don't have a good specification for what to generate. I'm actually looking to tackle that and would be happy to brainstorm with you on what we could do.

[jsuereth]

Copying this here from #3202.

My requirements for a solution to this bug:

• Require service.instance.id to be generated by every SDK.
• Outline an algorithm/logic for service.instance.id synthesis in a majority of cases. Allow UUID for unknown scenarios.
• Allow users to override service.instance.id via configuration. Today that means OTEL_RESOURCE_ATTRIBUTES.
• Ensure folks using Prometheus scrape/pull-based metrics of any sort w/ OTEL have an opportunity to see consistent service.instance.id. This may actually not be technically feasible

[jsuereth]

Ok, my proposal - Looking for feedback:

https://docs.google.com/document/d/1BenPf9vsZHCf4JpHWGQBydKZAA4XdH38wuMD7JQnz9A/edit?usp=sharing