The knative-kafka project is no longer being maintained
in the kyma-incubator repository. The v0.14.0
release is the final
release from this repository. The project has been moved to a new knative-sandbox
repository called eventing-kafka.
This project is a Knative Eventing implementation of a Kafka backed channel which provides advanced functionality and production grade qualities beyond what the eventing-contrib/kafka implementation offers.
The Knative eventing-contrib repository already contains a Kafka implementation, so why invest the time in building another one? At the time this project was begun, and still today, the reference Kafka implementation does not provide the scaling characteristics required by a large and varied use case with many different Topics and Consumers. That implementation is based on a single choke point that could easily allow one Topic's traffic to impact the throughput of another Topic. It is also not horizontally scalable as it only supports a single instance of the dispatcher/consumer. Further, no ordering guarantees on the consumption side are provided which is required in certain use cases. The reference implementation was also based on the Sarama Go Client which has limitations compared to the Confluent Go Client, and the librdkafka C library on which it is based. We also had the need to support a variety of Kafka providers, including Azure EventHubs in Kafka compatibility mode. Finally, the ability to expose Kafka configuration was very limited, and we needed the ability to customize certain aspects of the Kafka Topics / Producers / Consumers.
The Knative-Kafka project originated as an internal SAP implementation and was based on very early knative-eventing implementations. At the time this meant kube-builder and the controller-runtime library were used for the foundation of the controller. This also predated any of the more recent duck-typing and higher level abstractions (brokers, triggers, etc) which have since been added to knative-eventing.
The internal SAP project for which this was intended also underwent several variations in it's requirements and approach, which sometimes meant the development of this project languished behind the fast moving and ever changing knative-eventing implementation. Further, internal corporate CI/CD constraints imposed some of the structure of the current project.
Recently, however, the commitment to this effort was renewed, and the
implementation is now current with knative-eventing /master. Work is
in progress to align with implementation structure in eventing-contrib in
the hopes that this project can replace the default "kafka" implementation.
The project has moved to the open source kyma-incubator
repository and the focus is on bringing it into alignment with the other
eventing-contrib implementations.
As mentioned in the "Rationale" section above, the desire was to implement
different levels of granularity to achieve improved segregation and scaling
characteristics. Our original implementation was extremely granular in that
there was a separate Channel/Producer Deployment for every KafkaChannel
(Kafka Topic), and a separate Dispatcher/Consumer Deployment for every Knative
Subscription. This allowed the highest level of segregation and the ability to
tweak K8S resources at the finest level.
The downside of this approach, however, is the large resource consumption
related to the sheer number of Deployments in the K8S cluster, as well as the
inherent inefficiencies of low traffic rate Channels / Subscriptions being
underutilized. Adding in a service-mesh (such as Istio) further exacerbates the
problem by adding side-cars to every Deployment. Therefore, we've taken a step
back and aggregated the Channels/Producers together into a single Deployment per
Kafka authorization, and the Dispatchers/Consumers into a single Deployment per
KafkaChannel
(Topic). The implementations of each are horizontally scalable
which provides a reasonable compromise between resource consumption and
segregation / scaling.
Knative-Kafka is comprised of three distinct runtime K8S deployments as follows...
-
channel - The event receiver of the Channel to which inbound messages are sent. An http server which accepts messages that conform to the CloudEvent specification, and then writes those messages to the corresponding Kafka Topic. This is the "Producer" from the Kafka perspective. A separate Channel Deployment is created for each Kafka Secret detected in the knative-eventing namespace.
-
controller - This component implements the
KafkaChannel
Controller. It is using the current knative-eventing "Shared Main" approach based directly on K8S informers / listers. The controller is using theKafkaChannel
CRD, api, and client from the current Knative eventing-contrib/kafka implementation in order to align as closely as possible with the existing implementation. -
dispatcher - This component runs the Kafka ConsumerGroups responsible for processing messages from the corresponding Kafka Topic. This is the "Consumer" from the Kafka perspective. A separate dispatcher Deployment will be created for each unique
KafkaChannel
(Kafka Topic), and will contain a distinct Kafka Consumer Group for each Subscription to theKafkaChannel
.
...and the controller can be deployed via the Helm Chart found in the resources/ directory...
- resources - Knative-Kafka Helm Chart used to install
the
KafkaChannel
CRD and controller. The use of ko has been explored but is currently not feasible due to limitations with its ability to handle our build / packaging of the external librdkafka C library.
The control plane for the Kafka Channels is managed by the
knative-kafka-controller which is installed
in the knative-eventing namespace. KafkaChannel
Custom Resource instances can
be created in any user namespace. The knative-kafka-controller will guarantee
that the Data Plane is configured to support the flow of events as defined by
Subscriptions
to a KafkaChannel. The underlying Kafka infrastructure to be used is defined in
a specially labeled K8S Secret in the
knative-eventing namespace. Knative-Kafka supports several different Kafka
(and Kafka-like) infrastructures.
The data plane for all KafkaChannels
runs in the knative-eventing namespace.
There is a single deployment for the receiver side of all channels which accepts
CloudEvents and sends them to Kafka. Each KafkaChannel
uses one Kafka topic.
This deployment supports horizontal scaling with linearly increasing performance
characteristics through specifying the number of replicas.
Each KafkaChannel
has one deployment for the dispatcher side which reads from
the Kafka topic and sends to subscribers. Each subscriber has its own Kafka
consumer group. This deployment can be scaled up to a replica count equalling the
number of partitions in the Kafka topic.
An event sent to a KafkaChannel
is guaranteed to be persisted and processed
if a 202 response is received by the sender.
The CloudEvent is partitioned based on the CloudEvent partitioning extension
field called partitionkey
. If the partitionkey
is not present, then the
subject
field will be used. Finally if neither is available, it will fall-back
to random partitioning.
Events in each partition are processed in order, with an at-least-once guarantee. If a full cycle of retries for a given subscription fails, the event is ignored and processing continues with the next event.
For installation and configuration instructions please see the Helm Chart README.
The project uses the Confluent Go Client which relies on the following...
brew install pkg-config
brew install librdkafka
...alternatively, if the current/latest version of the librdkafka C library is not compatible with knative-kafka, then you can manually install an older version via the following....
curl -LO https://github.com/edenhill/librdkafka/archive/v1.0.1.tar.gz \
&& tar -xvf v1.0.1.tar.gz \
&& cd librdkafka-1.0.1 \
&& ./configure \
&& make \
&& sudo make install