profiles: use single Profile.sample_type and clarify use of timestamps

[florianl]

Implements #633 (comment).

FYI: @open-telemetry/profiling-approvers

[aalexand]

Should default sample type go to the scope message?

[christos68k]

Should default sample type go to the scope message?

That would be less flexible as it'd force us to have separate instrumentation scopes (ScopeProfiles) for different sample types. With the current proposal that's a possibility but not a requirement. The OTel documentation for instrumentation scope is sufficiently vague to allow both possibilities. For example, with the OTel eBPF profiling agent, we could have a single instrumentation scope (the eBPF profiler receiver/library) produce profiles with different sample types (cpu and memory).

Do you have a use-case in mind where associating different sample types with separate instrumentation scopes would be a benefit?

(I don't have a strong opinion on this, just that if we don't have a good reason to force a separate instrumentation scope per sample type, I'd go for more flexibility).

[aalexand]

Should default sample type go to the scope message?

That would be less flexible as it'd force us to have separate instrumentation scopes (ScopeProfiles) for different sample types. With the current proposal that's a possibility but not a requirement. The OTel documentation for instrumentation scope is sufficiently vague to allow both possibilities. For example, with the OTel eBPF profiling agent, we could have a single instrumentation scope (the eBPF profiler receiver/library) produce profiles with different sample types (cpu and memory).

Do you have a use-case in mind where associating different sample types with separate instrumentation scopes would be a benefit?

(I don't have a strong opinion on this, just that if we don't have a good reason to force a separate instrumentation scope per sample type, I'd go for more flexibility).

I mostly want to make sure there is some way to capture the default sample type. The current PR removes the field IIUC.

[florianl]

Should default sample type go to the scope message?
I mostly want to make sure there is some way to capture the default sample type. The current PR removes the field IIUC.

The main purpose of the proposal is to align values with timestamps as laid out in #633 (comment).

In general google/pprof it is possible to have multiple sample_types per Profile with special handling of sample_type.unit=count but does not have detailed timestamps but just a durations. With a slightly more focus on timestamps in OTel Profiling, I think, the concept of having multiple sample_types, like in google/pprof, conflicts. IIRC @felixge and I discussed the situation, where one wants to report profiles with different sample types and we concluded that it should be fine, to have multiple message Profile with their dedicated sample_type. I think, it also reduces complexity by not having a concept of a default sample_type and simplifies also the handling of message Profiles in the backend, if they just want to focus on a certain sample_type.

[christos68k]

Should default sample type go to the scope message?
I mostly want to make sure there is some way to capture the default sample type. The current PR removes the field IIUC.

The main purpose of the proposal is to align values with timestamps as laid out in #633 (comment).

In general google/pprof it is possible to have multiple sample_types per Profile with special handling of sample_type.unit=count but does not have detailed timestamps but just a durations. With a slightly more focus on timestamps in OTel Profiling, I think, the concept of having multiple sample_types, like in google/pprof, conflicts. IIRC @felixge and I discussed the situation, where one wants to report profiles with different sample types and we concluded that it should be fine, to have multiple message Profile with their dedicated sample_type. I think, it also reduces complexity by not having a concept of a default sample_type and simplifies also the handling of message Profiles in the backend, if they just want to focus on a certain sample_type.

@aalexand (I misread your previous comment and thought you wanted to lift sample_type to the scope message). Regarding default_sample_type, are you referring to introducing it in the scope message in order to make sample_type in each profile optional or do you have something else in mind? Lile @florianl mentioned the current proposal splits different sample_types across different profiles.

[aalexand]

Should default sample type go to the scope message?
I mostly want to make sure there is some way to capture the default sample type. The current PR removes the field IIUC.

@aalexand (I misread your previous comment and thought you wanted to lift sample_type to the scope message). Regarding default_sample_type, are you referring to introducing it in the scope message in order to make sample_type in each profile optional or do you have something else in mind? Lile @florianl mentioned the current proposal splits different sample_types across different profiles.

As one example, Go heap profiles have four sample types today: alloc_objects, alloc_space, inuse_objects, inuse_space. There are two profile types with this sample type set:

• The heap profile sets the default sample type to empty string (that is, unsets it) which means that tools visualizing the profile should use the last sample type as the default one (the "use the last one if the default is not set" is a non-intuitive convention but this is what the pprof format has historically).
• The alloc profile on the other hand sets the default sample type to alloc_space, so that tools like pprof focus the user on that metric by default.

The question is how such a profile with four sample types and one of them being the preferred default as chosen by the profile producer should be encoded in OpenTelemetry profiles. Before this change it would be the array of sample types plus the default sample type. What would it be after this change and how would the default be captured?

[christos68k]

The question is how such a profile with four sample types and one of them being the preferred default as chosen by the profile producer should be encoded in OpenTelemetry profiles. Before this change it would be the array of sample types plus the default sample type. What would it be after this change and how would the default be captured?

The current proposal doesn't have an explicit default sample type (and we probably should steer away from introducing implicit conventions) so assuming four different sample types in a single OTel messsage, samples will be grouped according to sample type to separate profiles. One possible encoding would have four different profiles, each with a different sample type (of course one could further split the data according to other criteria).

Some options to accommodate the notion of a default sample type (e.g. for a consumer visualization tool) include:

1. Use an InstrumentationScope attribute
2. Introduce default_sample_type in ScopeProfiles

1 is more restricted than 2 as it could be seen as purely advisory metadata that a visualization tool can use. Adding default_sample_type to ScopeProfiles could also act as a fallback sample_type for profiles that do not declare it.

[felixge]

LGTM in general. Sorry for the delay. PTAL at my comments. (I still need to find time to reply to some of the comments above, will do soon)

[florianl]

I have added default_sample_type as suggested by @christos68k and discussed in the OTel Profiling SIG (2025-06-29):

2. Introduce default_sample_type in ScopeProfiles

I didn't mark Profile.sample_type as optional, as this would conflict with the effort to avoid optional in the proto (see #659).

[felixge]

This LGTM, including the proposed default_sample_type field. Sorry for the delay 😞.