Current state: Under Discussion
Discussion thread: Thread
Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).
Operators of Apache Kafka clusters have little information about the type of clients connected to their clusters besides the `clientId`. Having more information about the connected clients such as their software name and version could tremendously help them to (1) troubleshoot misbehaving clients; and (2) understand the impact of a broker upgrade to their clients and inform them proactively.
ApiVersionsRequest is bumped to version 3 with two new fields. ApiVersionsRequest version is a flexible version (KIP-482: The Kafka Protocol should Support Optional Tagged Fields).
ApiVersionsResponse is bumped to version 3 but does not have any changes in the schema. Note that ApiVersionsResponse is not a flexible version. This is necessary because the client must look at a fixed offset to find the error code, regardless of the response version, to remain backward compatible.
Errors.INVALID_REQUEST is added.
We will add few metrics in the broker to surface information about the connected clients.
|Metric||Type||Description||Can be plotted?|
|kafka.server:type=ClientMetrics,name=ConnectedClients||Gauge<Integer>||The total number of client connected.||Yes|
|kafka.server:type=ClientMetrics,name=ConnectedClients,softwarename=([\.\-_a-zA-Z0-9])+,softwareversion=([\.\-_a-zA-Z0-9])+||Gauge<Integer>||The number of client connected, broken down by softwarename and softwareversion. It gives an overview of the clients.|
The metric will be removed when it goes back to zero - when the all the clients with a given name and version are disconnected.
The clients connected to the broker where each Map represents a connection with the following metadata:
|No - Operator can get the active connections via JMX by using a tool such as jmxterm|
While the Request Log is not a public interface, it is worth mentioning that we will enrich it with the Client Name and the Client Version.
The idea is to re-use the existing ApiVersionsRequest to provide the name and the version of the client to the broker. Clients are responsible to provide their name and version.
ApiVersions Request/Response Handling
The client does not know which ApiVersions versions the broker supports as the ApiVersions is used for this purpose. Today, the client sends an ApiVersionsRequest with the latest schema it is aware of. The broker handles it with the correct version if it knows it or sends back an ApiVersionsResponse v0 with an UNSUPPORTED_VERSION error to the client if it doesn't. When the client receives such error, it retries the whole process with the ApiVersionsRequest v0. It means that the broker won't get any additional information about the client if the client uses a newer version that the broker doesn't know about. To circumvent this, we propose to provide the supported version of the ApiVersionsRequest in the response sent back to the client by populating the existing api_versions field when the version is not supported (ErrorCode equals to UNSUPPORTED_VERSION). This enables the client to use the latest version supported by the broker instead of defaulting to version 0.
At the moment, the ApiVersionsRequest is handled in two different places in the broker: 1) in the SaslServerAuthenticator (when used); and 2) in the KafkaApis. Both places will be updated to ensure that all clients work. We have decided to not refactor the handling of the ApiVersionsRequest for now and to leave it for further improvements.
We propose to attach the various metadata captured to the connection alongside existing metadata such as the principal or the listener. A registry will be created to store metadata about all the active connections. Connections will be removed when they are closed.
We propose to validate the client name and the client version with the following regular expression: ([\.\-a-zA-Z0-9])+. The INVALID_REQUEST error is returned to the client if the validation fails. When the client receives an INVALID_REQUEST, it must error out and close the connection.
Metrics & Log
The various metrics described above will be created based on the metadata available in the connection registry. Metrics will be removed when they are inactive (gauge equals to zero). The request log will be extended to include the metadata collected.
ApiVersions Request/Response Handling
As mentioned earlier, when the client receives an UNSUPPORTED_VERSION error, it will use the version provided in the api_versions field of the ApiVersionsResponse and fail back to the higher version know by the broker instead of defaulting to version 0. When the client receives an INVALID_REQUEST error, it will error out and close the connection.
When SASL is used, the (Java) client sends two ApiVersionsRequest to the broker. The first one is sent by the SaslClientAuthenticator and the second one is send by the NetworkClient when the KafkaChannel is established. The SaslClientAuthenticator always sends version 0 of the AVR. We have decided to not change this for now and to only update the second call which always happens. The reasoning behind this choice is to avoid multiplying the round trip when an unknown version is used by the client, version 0 always works.
ClientSoftwareName and ClientSoftwareVersion
The client uses the version provided in the `kafka/kafka-version.properties` file and the name `apache-kafka-java`.
Compatibility, Deprecation, and Migration Plan
What impact (if any) will there be on existing users?
Existing users extracting and parsing the Request Log may have to update their parsing logic to accommodate the new fields.
Put ClientSoftwareName and ClientSoftwareVersion in the RequestHeader
ClientSoftwareName and ClientSoftwareVersion could be sent in every request alongside to the clientId in the header. While this would be fairly simple to implement once KIP-482 is implemented, we believe it is not suitable if we want to collect more information in the future and would wast few bytes in every request for something which does not change within a session. It also makes the error handling weird as a request could be rejected due to its headers. Another issue is that we haven't found a way to evolve the header of the ApiVersionsResponse to support tagged fields.
Put ClientSoftwareName and ClientSoftwareVersion in the RequestHeader but provide it only once
ClientSoftwareVersion could be added to the RequestHeader but sent only in the first request to save bytes in the subsequent requests. The best would be to have it in the ApiVersionsRequest's header but it is impossible (see previous point). It would be weird to have the information in random requests and could make clients inconsistent.
Add a new request to communicate the client metadata to the broker
A new separate request/response could be used for the purpose. This option has been discarded because it would add another round trip to the broker in the establishment of the KafkaChannel.
ApiVersionsRequest combined with "prefix-based" compatibility
We have considered removing the extra round-trip to the broker when the version of the AVR is unknown by ensuring that new fields would be added to the end of the ApiVersions Request and Response. This way, we could parse newer version of the request or the response with any previous version. We have discovered this solution because it would have obliged us to freeze the RequestHeader forever which is not wise.