Add db.cosmosdb.partition_key as a span attr on item ops

[kjg]

It would be useful to have the partition_key as an attribute on Item operations to make it easier to see if there are any performance based patterns in certain partitions, what are you thoughts on including it?

I know that db.cosmosdb.partition_key isn't in the OpenTelemetry Semantic Conventions (yet) as an attribute, but we could open a PR for that as well.

• The purpose of this PR is explained in this or a referenced issue.
• The PR does not update generated files.
  • These files are managed by the codegen framework at Azure/autorest.go.
• Tests are included and/or updated for code changes.
• Updates to module CHANGELOG.md are included.
• MIT license headers are included in each file.

[github-actions]

Thank you for your contribution @kjg! We will review the pull request and get back to you soon.

[Pilchie]

@FabianMeiswinkel @sourabh1007 @jcocchi - any thoughts here on the impact of this?

[analogrelay]

My main concern is PII leakage. The query text and parameters are in the Semantic Conventions but with explicit notes about sanitization. The partition key could be considered PII in some cases. I don't know for sure if we currently emit db.query.text or db.query.parameter.<key> but they're all in the same "class" of value to me, and should be opt-in.

I'm OK having it as an option, it does seem quite useful if you know your PKs are not PII.

[kjg]

azcosmos does not yet emit db.query.text but it actually is on my list of PRs I'd like to make. I'll see what I can do about making it optional. Any thoughts on the best place for consumers to specify the option and if it should be opt-in or opt-out?

For my team's use case, I could see us wanting to emit db.query.text and db.cosmosdb.partition_key but probably not db.query.parameter.<key>, so I'm not sure if we'd want to make them separately optional?

[analogrelay]

azcosmos does not yet emit db.query.text but it actually is on my list of PRs I'd like to make.

I absolutely think we need db.query.text to be opt-in, given the requirement in the OpenTelemetry guidance (which aligns with our own privacy/security guidance) to sanitize query text. I don't think it's feasible for us to add sanitization logic, so having it be opt-in is the appropriate alternative.

---

One thought is to have a list of "enabled tracing attributes" in azcosmos.ClientOptions. This would be a list of opt-in tracing attributes which the user is opting in to. Alas, my Go is rusty, but I don't think you can initialize a slice inside a struct initializer. What I'd like is to be able to write this:

options := azcosmos.ClientOptions {
    EnabledTracingAttributes: []string { "db.query.text", "db.cosmosdb.partition_key" },
}

But I don't know if that's feasible.

There's also the element of having to scan this list every time we build a span, to identify which attributes to specify.

I'd be curious what, if anything, other SDKs (both other Azure services' Go SDKs and Azure SDKs for other languages) do here 🤔 . Might poke around a bit and see what I can find.

[sourabh1007]

azcosmos does not yet emit db.query.text but it actually is on my list of PRs I'd like to make.

I absolutely think we need db.query.text to be opt-in, given the requirement in the OpenTelemetry guidance (which aligns with our own privacy/security guidance) to sanitize query text. I don't think it's feasible for us to add sanitization logic, so having it be opt-in is the appropriate alternative.

One thought is to have a list of "enabled tracing attributes" in azcosmos.ClientOptions. This would be a list of opt-in tracing attributes which the user is opting in to. Alas, my Go is rusty, but I don't think you can initialize a slice inside a struct initializer. What I'd like is to be able to write this:

options := azcosmos.ClientOptions {
    EnabledTracingAttributes: []string { "db.query.text", "db.cosmosdb.partition_key" },
}

But I don't know if that's feasible.

There's also the element of having to scan this list every time we build a span, to identify which attributes to specify.

I'd be curious what, if anything, other SDKs (both other Azure services' Go SDKs and Azure SDKs for other languages) do here 🤔 . Might poke around a bit and see what I can find.

@analogrelay We have added db.query.text as an opt-in attribute (ref. here) in cosmos DB SDK.

[sourabh1007]

@FabianMeiswinkel @sourabh1007 @jcocchi - any thoughts here on the impact of this?

As @analogrelay pointed out, the primary concern here is protecting PII, particularly if referencing logical partition keys tied to customer data. Instead, we could use physical identifiers such as partitionid (and replicaid) or pkrangeid, which would help in identifying potential hot partition issues without exposing sensitive information.

In the Cosmos DB SDK, we currently don't emit traces at the network level, as network interactions can involve multiple partition IDs per operation. Therefore, we haven't included this data in traces yet. However, we are planning to introduce it as an optional dimension in network-level OpenTelemetry (OTel) metrics (ref here). This would allow users to opt in as needed, helping to manage cardinality effectively.