Table of Contents |
---|
Status
Current state: Accepted
Discussion thread: here
...
JIRA:
Jira | ||||||
---|---|---|---|---|---|---|
|
PRPRs: https://github.com/apache/kafka/pull/10851 https://github.com/apache/kafka/pull/1078510802 https://github.com/apache/kafka/pull/11837
Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).
Motivation
Similar to many distributed systems, Kafka Streams instances can also be grouped in different racks. When Kafka Stream's standby task is properly distributed in different rack compared to the corresponding active task, it provides fault tolerance and faster recovery time if the rack of the active task goes down.
Below we will explore how other distributed systems implement rack awareness and what kind of guarantees they aim to provide.
Elasticsearch
Rack awareness in Elasticsearch works by defining a list of tags/attributes, called awareness attributes to each node in the cluster. When Elasticsearch knows the nodes' rack specification, it distributes the primary shard and its replica shards to minimize the risk of losing all shard copies in the event of a failure. Besides defining an arbitrary list of tags/attributes for each node, Elasticsearch provides a means of setting which tags/attributes it must consider when balancing the shards across the racks.
...
In the example above, if we start two nodes with node.attr.zone
set to zone1
and create an index with five shards and one replica, Elasticsearch creates the index and allocates the five primary shards but no replicas. Replicas are only allocated once nodes with node.attr.zone
set to zone2
is available.
Hadoop
In the case of Hadoop, rack is a physical collection of nodes in the cluster, and it's the mean of fault tolerance, as well as optimization. The idea in Hadoop is that read/write operation in the same rack is cheaper compared to when the process spans across multiple racks. With the rack information, Namenode chooses the closest Datanode while performing the read/write operation, which reduces network traffic.
...
- There should not be more than 1 replica on the same Datanode.
- More than 2 replica's of a single block is not allowed on the same rack.
- The number of racks used inside a Hadoop cluster must be smaller than the number of replicas.
Redis
Rack "awareness" in Redis is called "Rack-zone awareness" and it's very similar to Kafka Broker's rack awareness. Rack-zone awareness only works in a clustered Redis deployment, and it's an enterprise feature.
...
In the event of a rack failure, the remaining racks' replicas and endpoints will be promoted. This approach ensures high availability when a rack or zone fails.
Proposed Changes
This KIP proposes to implement similar semantics in Kafka Streams as in Elasticsearch. Rack awareness semantics in Elasticsearch seems the most flexible and can cover more complex use-cases, such as multi-dimensional rack awareness. To achieve this, KIP proposes to introduce a new config prefix in StreamsConfig that will be used to retrieve user-defined instance tags of the Kafka Streams
...
Info |
---|
Standby task distribution algorithm is not specified in this KIP, but is left as an implementation detail. However, every distribution algorithm must handle gracefully when ideal standby task distribution is not possible; In that case, Kafka Streams must not fail the assignment but try to distribute the standby tasks on best-effort bases. With an ideal task distribution, each client of the set of clients that host a given active task and the corresponding standby replicas has a unique value for each tag with regard to the other clients in the set. |
Example of the ideal task distribution
Suppose we have the following infrastructure setup: Three Kubernetes Clusters, let us call them K8s_Cluster1, K8s_Cluster2, and K8s_Cluster3. Each Kubernetes cluster is spanned across three availability zones: eu-central-1a, eu-central-1b, eu-central-1c.
...
Algorithm will chose either 1st or 2nd option, but not both.
Compatibility, Deprecation, and Migration Plan
The changes proposed by this KIP shouldn't affect previously setup applications. Since we introduce new configuration options, existing ones shouldn't be affected by this change.
Changes in Task Assignment logic
Implementation of this KIP will not affect task assignor behaviour specified in the KIP-441. Proposal mentioned in this KIP will merely extend the behaviour of the distribution of standby tasks. Behaviour will be extended only if required configurations mentioned in this KIP specified by the Kafka Stream's user.
Rejected Alternatives
The initial idea was to introduce two configurations in StreamsConfig,
rack.id
, which defines the rack of the Kafka Streams instance andstandby.task.assignor
- class that implementsRackAwareStandbyTaskAssignor
interface.The signature of RackAwareStandbyTaskAssignor was the following:
Code Block language java public interface RackAwareStandbyTaskAssignor { /** * Computes desired standby task distribution for a different {@link StreamsConfig#RACK_ID_CONFIG}s. * @param sourceTasks - Source {@link TaskId}s with a corresponding rack IDs that are eligible for standby task creation. * @param clientRackIds - Client rack IDs that were received during assignment. * @return - Map of the rack IDs to set of {@link TaskId}s. The return value can be used by {@link TaskAssignor} * implementation to decide if the {@link TaskId} can be assigned to a client that is located in a given rack. */ Map<String, Set<TaskId>> computeStandbyTaskDistribution(final Map<TaskId, String> sourceTasks, final Set<String> clientRackIds); }
By injecting custom implementation of RackAwareStandbyTaskAssignor interface, users could hint Kafka Streams where to allocate certain standby tasks when more complex processing logic was required — for example, parsing rack.id, which can be a combination of multiple identifiers (as seen in the previous examples where we have cluster and zone tags).
The above mentioned idea was abandoned because it's easier and more user-friendly to let users control standby task allocation with just configuration options instead of forcing them to implement a custom interface.
Defining multiple
client.tag
with combination ofrack.aware.assignment.tags
gives more flexibility, which, as already mentioned above, could have been only possible with pluggable custom logic Kafka Streams's user must provide.For instance, if we append multiple tags to form a single rack, it may not give desired distribution to the user if the infrastructure topology is more complex. Let us consider the following example with appending multiple tags to form the single rack.
Code Block Node-1: rack.id: K8s_Cluster1-eu-central-1a num.standby.replicas: 1 Node-2: rack.id: K8s_Cluster1-eu-central-1b num.standby.replicas: 1 Node-3: rack.id: K8s_Cluster1-eu-central-1c num.standby.replicas: 1 Node-4: rack.id: K8s_Cluster2-eu-central-1a num.standby.replicas: 1 Node-5: rack.id: K8s_Cluster2-eu-central-1b num.standby.replicas: 1 Node-6: rack.id: K8s_Cluster2-eu-central-1c num.standby.replicas: 1
In the example mentioned above, we have three AZs and two Kubernetes clusters. Our use-case is to distribute standby task in the different Kubernetes cluster and different availability zone. For instance, if the active task is in Node-1 (K8s_Cluster1-eu-central-1a), the corresponding standby task should be in either on Node-5 (K8s_Cluster2-eu-central-1b) or on Node-6 (K8s_Cluster2-eu-central-1c).
Unfortunately, without custom logic provided by the user, this would be very hard to achieve with a single
rack.id
configuration. Because without any input from the user, Kafka Streams might as well allocate standby task for the active task either:- In the same Kubernetes cluster and different AZ (Node-2, Node-3)
- In the different Kubernetes cluster but the same AZ (Node-4)
On the other hand, with the combination of the new "client.tag.*" and "rack.aware.assignment.tags" configurations, standby task distribution algorithm will be able to figure out what will be the most optimal distribution by balancing the standby tasks over each client.tag dimension individually. And it can be achieved by simply providing necessary configurations to Kafka Streams.
- The second approach was to refactor
TaskAssignor
interface to be more user-friendly and expose it as a public interface. Users then could implement customTaskAssignor
logic and set it viaStreamsConfig
. With this, Kafka Streams users would effectively be in control of Active and Standby task allocation.
Similarly to the point above, this approach also was rejected because it's more complex.
Even though it's more-or-less agreed on the pluggable TaskAssignor interface's usefulness, it was decided to cut it out of this KIP's scope and prepare a separate one for that feature.
...