Kafka Configuration

General

Kafka is OSS message-oriented-middleware and is well documented. Brighter handles the details of sending to or receiving from Kafka. You may find it useful to understand the building blocks of the protocol. Brighter's Kafka support is implemented on top of the Confluent .NET client, and you might find the documentation for the .NET client helpful when debugging, but you should not have to interact with it directly to use Brighter (although we expose many of its configuration options).

Kafka has two main roles:

  • Producer: A producer sends events to a Topic on a Kafka broker.

  • Consumer: A consumer reads events from a Topic on a Kafka broker.

Topics are append-only streams of events. Multiple producers can write to a topic, and multiple consumers can read from one. A consumer uses an offset into the stream to indicate the event it wants to read. Kafka does not delete an event from the stream when it is ack'd by the consumer; instead a consumer increments its offset once an item has been read so that it can avoid processing the same event twice. See Offset Management for more on how Brighter manages consumer offsets. As a result the lifetime of events on a stream is instead a configuration setting for the stream.

As a consumer manages an offset to record events that is has read, you cannot scale an application that wishes to consume a topic by increasing the number of consumers--they don't share an offset--without partitioning the topic. If you supply a partition key, a partition uses consistent hashing to slice a topic into a number of streams; otherwise it will use round-robin. See this documentation for more. Each partition is only read by a single consumer within the application. All of the consumers for an application should share the same group id, called a consumer group in Kafka. As each consumer tracks the offset for the partitions it is reading, it is possible to have multiple consumers read and process the same topic.

A consumer may read from multiple partitions, but only one consumer may read from a partition at one time in a given consumer group. Kafka will assign partitions across the pool of consumers for the consumer group. When the pool changes, a rebalance occurs, which may mean that a consumer changes the partition that it is assigned within the consumer group. Brighter favors sticky assignment of partitions to avoid unnecessary churn of partitions.

In addition to the Producer API and Consumer API Kafka streams have features such as the Streams API and the Connect API. We do not use either of these from Brighter.

Connection

The Connection to Kafka is provided by an KafkaMessagingGatewayConnection which allows you to configure the following:

  • BootstrapServers: A bootstrap server is a well-known broker through which we discover the servers in the Kafka cluster that we can connect to. You should supply a comma-separated list of host and port pairs. These are the addresses of the Kafka brokers in the "bootstrap" Kafka cluster.

  • Debug: A comma-separated list of debug contexts to enable. Producer: broker, topic, msg. Consumer: consumer, cgrp, topic, fetch.

  • Name: An identifier to use for the client.

  • SaslMechanisms: If any, what is the protocol used for authenticated connection to the Kafka broker: plain, scram-sha-256, scram-sha-256, gssapi (kerberos), oauthbearer

  • SaslKerberosName: If using kerberos, what is the connection name.

  • SaslUsername: SASL username for use with PLAIN and SASL-SCRAM

  • SaslPassword: SASL password for use with PLAIN and SASL-SCRAM

  • SecurityProtocol: How are messages between client and server encrypted, if at all: plaintext, ssl, saslplaintext, saslssl

  • SslCaLocation: Where is the CA certificate located (see here for guidance).

  • SslKeystoreLocation: Path to the client's keystore

  • SslKeystorePassword: Password for the client's keystore

The following code connects to a local Kafka instance (for development):

	services.AddBrighter(...)
	.UseExternalBus(
	new KafkaProducerRegistryFactory(
		new KafkaMessagingGatewayConfiguration()
		{
			Name = "paramore.brighter.greetingsender",
			BootStrapServers = new[] {"localhost:9092"}
		},
		...//publication, see below
		)
	.Create())
	...

The following code connects to a remote Kafka instance. The settings here will depend on how your production broker is configured for access. We show getting secrets from environment variables for simplicity, again you will need to adjust this for your approach to secrets management:

	services.AddBrighter(...)
	.UseExternalBus(
	new KafkaProducerRegistryFactory(
		new KafkaMessagingGatewayConfiguration()
		{
			Name = "paramore.brighter.greetingsender",
			BootStrapServers = new[] { Environment.GetEnvironmentVariable("BOOSTRAP_SERVER")},
			SecurityProtocol = Paramore.Brighter.MessagingGateway.Kafka.SecurityProtocol.SaslSsl,
			SaslMechanisms = Paramore.Brighter.MessagingGateway.Kafka.SaslMechanism.Plain,
			SaslUsername = Environment.GetEnvironmentVariable("SASL_USERNAME"),
			SaslPassword = Environment.GetEnvironmentVariable("SASL_PASSWORD"),
			SslCaLocation = RuntimeInformation.IsOSPlatform(OSPlatform.OSX) ? "/usr/local/etc/openssl@1.1/cert.pem" : null;
		},
		...//publication, see below
		)
	.Create())
	...

Publication

For more on a Publication see the material on an External Bus in Basic Configuration.

We allow you to configure properties for both Brighter and the Confluent .NET client. Because there are many properties on the Confluent .NET Client we also configure a callback to let you inspect and modify the configuration that we will pass to the client if you so desire. This can be used to add properties we do not support or adjust how we set them.

  • Replication: how many ISR nodes must receive the record before the producer can consider the write successful. Default is Acks.All.

  • BatchNumberMessages: Maximum number of messages batched in one MessageSet. Default is 10.

  • EnableIdempotence: Messages are produced once only. Will adjust the following if not set: max.in.flight.requests.per.connection=5 (must be less than or equal to 5), retries=INT32_MAX (must be greater than 0), acks=all, queuing.strategy=fifo. Default is true.

  • LingerMs: Maximum time, in milliseconds, for buffering data on the producer queue. Default is 5.

  • MessageSendMaxRetries: How many times to retry sending a failing MessageSet. Note: retrying may cause reordering, set the max in flight to 1 if you need ordering by when sent. Default is 3.

  • MessageTimeoutMs: Local message timeout. This value is only enforced locally and limits the time a produced message waits for successful delivery. A time of 0 is infinite. Default is 5000.

  • MaxInFlightRequestsPerConnection: Maximum number of in-flight requests the client will send. We default this to 1, so as to allow retries to not de-order the stream.

  • NumPartitions: How many partitions for this topic. We default to 1.

  • Partitioner: How do we partition? Defaults to Partitioner.ConsistentRandom.

  • QueueBufferingMaxMessages: Maximum number of messages allowed on the producer queue. Defaults to 10.

  • QueueBufferingMaxKbytes: Maximum total message size sum allowed on the producer queue. Defaults to 1048576 bytes (so for 10 messages about 104Kb per message).

  • ReplicationFactor: What is the replication factor? How many nodes is the topic copied to on the broker? Defaults to 1.

  • RetryBackoff: The backoff time before retrying a message send. Defaults to 100.

  • RequestTimeoutMs: The ack timeout of the producer request. This value is only enforced by the broker and relies on Replication being != AcksEnum.None. Defaults to 500.

  • TopicFindTimeoutMs: How long to wait when asking for topic metadata. Defaults to 5000.

  • TransactionalId: The unique identifier for this producer, used with transactions

The following example shows how a Publication might be configured:

	services.AddBrighter(...)
	.UseExternalBus(
	new KafkaProducerRegistryFactory(
		...,//connection see above
		new KafkaPublication[] {new KafkaPublication()
                {
                    Topic = new RoutingKey("MyTopicName"),
                    NumPartitions = 3,
                    ReplicationFactor = 3,
                    MessageTimeoutMs = 1000,
                    RequestTimeoutMs = 1000,
                    MakeChannels = OnMissingChannel.Create 
                }
		)
	.Create())
	...

Configuration Callback

The Confluent .NET client has a range of configuration options. Some of those can be controlled through the publication. But, to allow you the full range of configuration options for the Confluent client, including new options that may appear, we provide a callback on the KafkaProducerRegistryFactory. The registry exposes a method, SetConfigHook(Action hook). The method takes a delegate (you can pass a lambda). Your delegate will be called with the proposed ProducerConfig (taking into account the Publication settings). You can adjust additional parameters at this point.

You can use it as follows:


	var publication = new KafkaPublication()
	{
		Topic = new RoutingKey("MyTopicName"),
		NumPartitions = 3,
		ReplicationFactor = 3,
		MessageTimeoutMs = 1000,
		RequestTimeoutMs = 1000,
		MakeChannels = OnMissingChannel.Create 
	};
	publication.SetConfigHook(config => config.EnableGaplessGuarantee = true)

	services.AddBrighter(...)
	.UseExternalBus(
	new KafkaProducerRegistryFactory(
		...,//connection see above
		new KafkaPublication[] {publication})
	.Create())
	...

Kafka Topic Auto Create

Brighter uses the Kafka AdminClient for topic creation. For this to work as expected you should set the server property of auto.create.topics.enable to false; otherwise the topic will be auto-created with the values defined by your server for new topics, such as the number of partitions. This error can be insidious because your code will still work against this topic, but without inspection you will not observe that its properties do not match those requested.

If you want to specify the topic through Brighter, or through your own IaaS code, we recommend always setting this setting to false; we recommend only setting it to true if you tell Brighter to assume that the infrastructure exists, as it will then be created on the first write.

Subscription

For more on a Subscription see the material on configuring Service Activator in Basic Configuration.

We support a number of Kafka specific Subscription options:

  • CommitBatchSize: We commit processed work (marked as acked or rejected) when a batch size worth of work has been completed (see below).

  • GroupId: Only one consumer in a group can read from a partition at any one time; this preserves ordering. We do not default this value, and expect you to set it.

  • IsolationLevel: Default to read only committed messages, change if you want to read uncommitted messages. May cause duplicates.

  • MaxPollIntervalMs: How often the consumer needs to poll for new messages to be considered alive, polling greater than this interval triggers a re-balance. Kafka default to 300000ms

  • NumPartitions: How many partitions does the topic have? Used for topic creation, if required.

  • OffsetDefault: What do we do if there is no offset stored in ZooKeeper for this consumer. Defaults to AutoOffsetReset.Earliest - Begin reading the stream from the start. Options include AutOffsetRest.Latest - Start from now i.e. only consume messages after we start and AutoOffsetReset.Error - which considers it an error if not reset is found

  • ReadCommittedOffsetsTimeOutMs: How long before attempting to read back committed offsets (mainly used in debugging) is an error. Defaults to 5000.

  • ReplicationFactor: What is the replication factor? How many nodes is the topic copied to on the broker? Defaults to 1. Used for topic creation if required.

  • SessionTimeoutMs: If Kafka does not receive a heartbeat from the consumer within this time window, trigger a re-balance. Default is Kafka default of 10s.

  • SweepUncommittedOffsetsIntervalMs: The interval at which we sweep, looking for offsets that have not been flushed (see below).

The following example shows how a subscription might be configured:

 private static void ConfigureBrighter(HostBuilderContext hostContext, IServiceCollection services)
{
	var subscriptions = new KafkaSubscription[]
	{
		new KafkaSubscription<GreetingEvent>(
			new SubscriptionName("paramore.example.greeting"),
			channelName: new ChannelName("greeting.event"),
			routingKey: new RoutingKey("greeting.event"),
			groupId: Environment.GetEnvironmentVariable("KAFKA_GROUPID"),
			timeoutInMilliseconds: 100,
			commitBatchSize: 5,
			sweepUncommittedOffsetsIntervalMs: 3000
		)
	};
		
	//create the gateway
	var consumerFactory = new KafkaMessageConsumerFactory(
	new KafkaMessagingGatewayConfiguration {...} // see connection information above
	);

	services.AddServiceActivator(options =>
	{
		options.Subscriptions = subscriptions;
		options.ChannelFactory = new ChannelFactory(consumerFactory);
	}).AutoFromAssemblies();


	services.AddHostedService<ServiceActivatorHostedService>();
}

Offset Management

It is important to understand how Brighter manages the offset of any partitions assigned to your consumer.

  • Brighter manages committing offsets to Kafka. This means we set the Confluent client's auto store and auto commit properties to false.

  • The CommitBatchSize setting on the Subscription determines the size of your buffer. A smaller buffer is less efficient, but if your consumer crashes any offsets pending commit in the buffer will be lost, and you will be represented with those records when you next read from the partition. We default this value to 10.

  • We do not add an offset commit to the buffer until you Ack the request. The message pump will Ack for you once you exit your handler (via return or throwing an exception).

  • Flushing the commit buffer happens on a separate thread. We only run one flush at a time, and we flush a CommitBatchSize number of items from the buffer.

    • A busy consumer may not flush on every increment of the CommitBatchSize, as it may need to wait for the last flush to finish.

    • We won't flush again until we cross the next multiple of the CommitBatchSize. For example if the CommitBatchSize is 10, and the handler is busy so that by the time the buffer flushes there are 13 pending commits in the buffer, the buffer would only flush 10, and 3 would remain in the buffer; we would not flush the next 10 until the buffer hit 20.

    • If your CommitBatchSize is too low for the throughput, you might find that you miss a flush interval, because you are already flushing.

    • If you miss a flush on a busy consumer, your buffer will begin to back up. If this continues, you will not catch up with subsequent flushes, which only flush the CommitBatchSize each time. This would lead to you continually being "backed up".

    • For this reason you must set a CommitBatchSize that keeps pace with the throughput of your consumer. Use a larger CommitBatchSize for higher throughput consumers, smaller for lower.

  • We sweep uncommitted offsets at an interval. This triggers a flush if no flush has run since the last flush plus the Subscription's SweepUncommittedOffsetsIntervalMs.

    • A sweep will not run if a flush is currently running (and will in turn block a flush).

    • A sweep flushes a CommitBatchSize worth of commits.

    • It is intended for low-throughput consumers where commits might otherwise languish waiting for a batch-size increment.

    • It is not intended to flush a buffer that backs up because the CommitBatchSize is too low, and won't function for that. Fix the CommitBatchSize instead.

  • On a re-balance where we stop processing a partition on an individual consumer, we flush the remaining offsets for the revoked partitions.

    • We configure the consumer to use sticky assignment strategy to avoid unnecessary re-assignments (see the Confluent documentation).

  • On a consumer shutdown we flush the buffer to commit all offsets.

Working with Schema Registry

If you want to use tools within the Kafka ecosystem such as Kafka Connect or KQSL you will almost certainly need to use Confluent Schema Registry to provide the schema of your message.

You will need to pull in the following package:

  • Confluent.SchemaRegistry

and a package for the serialization of your choice. Here we are using JSON, so we use

  • Confluent.SchemaRegistry.Serdes.Json

When working with Brighter, to use Confluent Schema Registry you will need to take a dependency on ISchemaRegistry in the constructor of your message mapper. To fulfill this constructor, in your application setup you will need to register an instance of schema registry. You should configure the schema registry config url to be the url of you schema registry. (Here we just use localhost for a development instance running in docker as an example).

var schemaRegistryConfig = new SchemaRegistryConfig { Url = "http://localhost:8081"};
var cachedSchemaRegistryClient = new CachedSchemaRegistryClient(schemaRegistryConfig);
services.AddSingleton<ISchemaRegistryClient>(cachedSchemaRegistryClient);

Once you can satisfy the dependency, you will want to use the serializer from the Serdes package to serialize the body of your message, instead of System.Text.Json. Note that 'under-the-hood' the Serdes serializer uses Json.NET and NJsonSchema, so you may need to mark up your code with attributes from these packages to create the schema you want and serialize a valid message to it. (Note that, at this time, the Serdes package does not support System.Text.Json so you will need to take a dependency on Json.NET if you want to use the schema registry).

It is worth noting the following aspects of the code sample below:

  • We need to set up a SerializationContext and tell Serdes that we are serializing the message body using their serializer

  • We provide two helpers, though you can pass your own settings if you prefer:

    • ConfluentJsonSerializationConfig.SerdesJsonSerializerConfig() offers default settings for JSON serialization (many of these are passed through to Json.NET).

    • ConfluentJsonSerializationConfig.NJsonSchemaGeneratorSettings() offers default settings for JSON Schema generation (such as using camelCase).

public class GreetingEventMessageMapper : IAmAMessageMapper<GreetingEvent>
{
private readonly ISchemaRegistryClient _schemaRegistryClient;
private readonly string _partitionKey = "KafkaTestQueueExample_Partition_One";
private SerializationContext _serializationContext;
private const string Topic = "greeting.event";

public GreetingEventMessageMapper(ISchemaRegistryClient schemaRegistryClient)
{
	_schemaRegistryClient = schemaRegistryClient;
	//We care about ensuring that we serialize the body using the Confluent tooling, as it registers and validates schema
	_serializationContext = new SerializationContext(MessageComponentType.Value, Topic);
}

public Message MapToMessage(GreetingEvent request)
{
	var header = new MessageHeader(messageId: request.Id, topic: Topic, messageType: MessageType.MT_EVENT);
	//This uses the Confluent JSON serializer, which wraps Newtonsoft but also performs schema registration and validation
        var serializer = new JsonSerializer<GreetingEvent>(_schemaRegistryClient, ConfluentJsonSerializationConfig.SerdesJsonSerializerConfig(), ConfluentJsonSerializationConfig.NJsonSchemaGeneratorSettings()).AsSyncOverAsync();
 	var s = serializer.Serialize(request, _serializationContext);
	var body = new MessageBody(s, "JSON");
	header.PartitionKey = _partitionKey;

	var message = new Message(header, body);
	return message;
}

public GreetingEvent MapToRequest(Message message)
{
	var deserializer = new JsonDeserializer<GreetingEvent>().AsSyncOverAsync();
	//This uses the Confluent JSON serializer, which wraps Newtonsoft but also performs schema registration and validation
	var greetingCommand = deserializer.Deserialize(message.Body.Bytes, message.Body.Bytes is null, _serializationContext);
	
	return greetingCommand;
}
}

Requeue with Delay (Non-Blocking Retry)

We don't currently support requeue with delay for Kafka. This is also known as non-blocking retry. With a stream if your app cannot process a record but it might be able to process the record after a delay (for example the DB is temporarily unavailable) then the options are:

  • Blocking Retry - keep retrying the processing of this record

  • Load Shedding - ack the record to commit the offset, skipping this record

  • Non-Blocking Retry - move the record to a new store or queue, skipping the original, append after a delay

Brighter supports the first two of these options.

  • Blocking Retry - use a Polly policy via the UsePolicy attribute

  • Load Shedding - allow the handler to complete, or throw an exception. This will cause the handler to commit the offset.

Note that Blocking Retry means you will apply backpressure as the blocking retry means you will pause consumption until the record can be processed.

A non-blocking retry typically creates a copy of the current record, and appends it to the stream so that it can be processed later:

  • Publish the message to be requeued to a new stream or store with a timestamp

  • Ack the existing message so at to commit the offset

  • Poll that stream or store and publish anything whose timestamp + delay means it is now due

(You may need multiple tables or streams to support different delay lengths)

Until Brighter supports this for you, implementation of non-blocking consumers is left to the user.

PreviousAWS SNS ConfigurationNextAzure Service Bus Configuration

Last updated