Eventing-Kafka Distributed Channel
This is a Kafka Channel implementation, originally contributed by
SAP's Kyma project.
See https://github.com/knative/eventing-contrib/issues/1070 for discussion
of the donation process.
This repo falls under the
Knative Code of Conduct
This project is a Knative Eventing implementation of a Kafka backed channel
which provides advanced functionality and production grade qualities as an
alternative to the "consolidated" implementation offers.
Rationale
The Knative consolidated implementation (originally from the 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.
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.
Background / Status
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 implementation initially utilized the
Confluent Go Client, and
the librdkafka C library on which it
is based. Initially the Sarama Go Client
had limitations which have since been resolved so we have refactored to that
library instead.
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 further in the hopes that this project can merge with the
default "kafka" implementation. The project which previously existed in the open
source kyma-incubator, but was moved into
it's current home here in the
knative-sandbox The focus is now on
bringing the implementation into alignment with the other knative channel
implementations.
Architecture
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.
Project Structure
Eventing-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 the shared KafkaChannel
CRD, apis/, and
client implementations in this repository.
-
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 the KafkaChannel
.
-
config - Eventing-kafka
ko installable YAML files for installation.
Control Plane
The control plane for the Kafka Channels is managed by the
eventing-kafka-controller which is installed in the
knative-eventing namespace. KafkaChannel
Custom Resource instances can be
created in any user namespace. The eventing-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. Eventing-kafka supports several different Kafka (and
Kafka-like)
infrastructures.
Data Plane
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.
Messaging Guarantees
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.
Installation
For installation and configuration instructions please see the config files
README.