Kafka: support templated topic name

[axw]

Component(s)

exporter/kafka

Is your feature request related to a problem? Please describe.

The kafkaexporter component can be configured in various ways to determine the topic to which messages are published:

• by default, messages are sent to a signal-specific topic: otlp_spans, otlp_metrics, etc.
• you can configure topic to explicitly specify the topic name
• you can configure topic_from_attribute to extract the topic name from a resource attribute
• you can programmatically inject the topic name into the context with https://github.com/open-telemetry/opentelemetry-collector-contrib/tree/main/pkg/kafka/topic

This is complex, and there are problems with each of these options:

• You cannot explicitly configure topic in a way that matches the default, i.e. you cannot configure signal-specific topic names: [exporter/kafka] Replace "topic" setting by "traces_topic", "logs_topic" and "metrics_topic" #35432
• If multiple resources are combined into one batch, then topic_from_attribute may behave incorrectly: topic_from_attribute does not work as expected #37470
• Programmatic topic name injection is flexible but will only work for custom components. Nothing in-tree uses it.

Describe the solution you'd like

Replace all of these with a templated topic name, e.g. otlp_{signal_record_type}s where "signal_type" expands to "log", "span", or "metric"; variable name is just for illustration. The {...} bit would be treated as an expression, e.g. using OTTL. So you could for example include:

• client metadata (for multi-tenancy): topic: {request["x-tenant-id"]}.whatever
• resource attributes: topic: {resource.attributes["service.name"]}-logs

See also #35432 (comment)

Assuming we use OTTL for the template expressions, then we may be able to rely on OTTL context inference to decide when to determine the topic: per request, resource, scope, record, etc. Otherwise, we could limit determining the topic to the resource level.

We should also support a templated topic for kafkareceiver to resolve #32735, but in that case the only expression that would make sense is the signal type. We'll need to consider how the templating syntax interacts with support for topic regular expressions, which the receiver should also ideally support.

Describe alternatives you've considered

No response

Additional context

No response

[github-actions]

Pinging code owners:

• exporter/kafka: @pavolloffay @MovieStoreGuy @axw

See Adding Labels via Comments if you do not have permissions to add labels yourself.

[aklemp]

Not sure, if this single change would also resolve the similar receiver issue (see #32735) or if the same changes would be necessary there.

[axw]

@aklemp yes, we should sort out the receiver at the same time. I updated the description.

[crobert-1]

Issue filed by code owner, removing needs triage.

[MovieStoreGuy]

I like the idea but I am worried about the potential impact, say I have templated out the topic name based on a tenant context, what happens if the client doesn't have permission to create topics, or if I have exceeded the amount of topics possible? (I don't remember what is the upper limit on topics).

Then on the flip side, say the exporter is creating all of these topics and sending data fine, how will the consumer know what topics to consume from?

[axw]

I like the idea but I am worried about the potential impact, say I have templated out the topic name based on a tenant context, what happens if the client doesn't have permission to create topics or if I have exceeded the amount of topics possible? (I don't remember what is the upper limit on topics).

Then on the flip side, say the exporter is creating all of these topics and sending data fine, how will the consumer know what topics to consume from?

Clarified on Slack, but for posterity: these are already issues that we would need to deal with, since it's already possible to dynamically choose the topic. Regarding permissions and limits, I wouldn't propose changing the defaults, just making it possible to include additional client metadata in the topic name. We should make it clear in logs and metrics when errors are related to lack of permissions.

Regarding the receiver side, there are a couple of options: there's https://github.com/open-telemetry/opentelemetry-collector-contrib/tree/main/extension/observer/kafkatopicsobserver, which you can use with receivercreator to dynamically create receivers for Kafka topics. I think this may be problematic at scale, as every receiver would have its own client, and make its own network connections -- so the network connections (and other resources) will be proportional to the number of topics.

The other option is to add support for topic regex on the receiver side, which is what I suggest since it avoids the issue described above. Sarama does not support this (see IBM/sarama#1379 and IBM/sarama#1382). On the other hand, it is supported by franz-go (https://pkg.go.dev/github.com/twmb/franz-go/pkg/kgo#ConsumeRegex) and confluent-kafka-go (https://github.com/confluentinc/confluent-kafka-go?tab=readme-ov-file#examples).

[aklemp]

My 5 cents on this: dynamically created topics create more problems than they help in my experience.

• the mentioned limit problems
• reduced security and increased attack vector (compromised client can do a lot to a Kafka cluster)
• difficult to estimate load on the system
• impossible to tune the topics according to their content (partitions, replication, retention, ...)

In the regulated context i'm currently working, it is actually forbidden to create topics on demand.

[axw]

@aklemp I agree with your points, but that is not applicable in my use case. The scenario I am trying to enable is multi-tenancy (https://kafka.apache.org/documentation/#multitenancy-topic-naming), where the tenant ID is associated with a client, but which the client cannot spoof. The number of tenants, and thus topics, are controlled -- so we can estimate load, and there is no additional security risk. We also do not rely on auto-topic creation, we orchestrate this separately -- so we can (and do) tune partitions, retention, etc.

Anyway, after thinking about this a bit more, I'm coming around to having per-signal type topic and encoding config. So my new proposal is:

1. in both the kafka exporter and receiver, introduce logs.topic, metrics.topic, traces.topic and deprecate topic config
2. in both the kafka exporter and receiver, introduce logs.encoding, metrics.encoding, traces.encoding, and deprecate encoding config
3. in the kafka exporter only, add support for templated topic name
4. in the kafka receiver only, add support for regex topic names

(1) will fix #32735 and #35432
(2) will enable specifying different encodings for different signals, e.g. "raw" for logs (only supported for logs), "jaeger_proto" for traces (only supported for traces), and "otlp_proto" for metrics
(3) and (4) will enable our use case of per-tenant topics