Skip to content

Apache Kafka primer

ksqlDB is a database built specifically for stream processing on Apache Kafka®. Although it's designed to give you a higher-level set of primitives than Kafka has, it's inevitable that all of Kafka's concepts can't be, and shouldn't be, abstracted away entirely. This section describes the minimum number of Kafka concepts that you need to use ksqlDB effectively. For more information, consult the official Apache Kafka documentation.


The primary unit of data in Kafka is the event. An event models something that happened in the world at a point in time. In Kafka, you represent each event using a data construct known as a record. A record carries a few different kinds of data in it: key, value, timestamp, topic, partition, offset, and headers.

The key of a record is an arbitrary piece of data that denotes the identity of the event. If the events are clicks on a web page, a suitable key might be the ID of the user who did the clicking.

The value is also an arbitrary piece of data that represents the primary data of interest. The value of a click event probably contains the page that it happened on, the DOM element that was clicked, and other interesting tidbits of information.

The timestamp denotes when the event happened. There are a few different "kinds" of time that can be tracked. These aren’t discussed here, but they’re useful to learn about nonetheless.

The topic and partition describe which larger collection and subset of events this particular event belongs to, and the offset describes its exact position within that larger collection (more on that below).

Finally, the headers carry arbitrary, user-supplied metadata about the record.

ksqlDB abstracts over some of these pieces of information so you don’t need to think about them. Others are exposed directly and are an integral part of the programming model. For example, the fundamental unit of data in ksqlDB is the row. A row is a helpful abstraction over a Kafka record. Rows have columns of two kinds: key columns and value columns. They also carry pseudocolumns for metadata, like a timestamp.

In general, ksqlDB avoids raising up Kafka-level implementation details that don’t contribute to a high-level programming model.


Topics are named collections of records. Their purpose is to let you hold events of mutual interest together. A series of click records might get stored in a "clicks" topic so that you can access them all in one place. Topics are append-only. Once you add a record to a topic, you can’t change or delete it individually.

There are no rules for what kinds of records can be placed into topics. They don't need to conform to the same structure, relate to the same situation, or anything like that. The way you manage publication to topics is entirely a matter of user convention and enforcement.

ksqlDB provides higher-level abstractions over a topic through streams and tables. A stream or table associates a schema with a Kafka topic. The schema controls the shape of records that are allowed to be stored in the topic. This kind of static typing makes it easier to understand what sort of rows are in your topic and generally helps you make fewer mistakes in your programs that process them.


When a record is placed into a topic, it is placed into a particular partition. A partition is a totally ordered sequence of records by offset. Topics may have multiple partitions to make storage and processing more scalable. When you create a topic, you choose how many partitions it has.

When you append a record to a topic, a partitioning strategy chooses which partition it is stored in. There are many partitioning strategies. The most common one is to hash the contents of the record's key against the total number of partitions. This has the effect of placing all records with the same identity into the same partition, which is useful because of the strong ordering guarantees.

The order of the records is tracked by a piece of data known as an offset, which is set when the record is appended. A record with offset of 10 happened earlier than a record in the same partition with offset of 20.

Much of the mechanics here are handled automatically by ksqlDB on your behalf. When you create a stream or table, you choose the number of partitions for the underlying topic so that you can have control over its scalability. When you declare a schema, you choose which columns are part of the key and which are part of the value. Beyond this, you don't need to think about individual partitions or offsets. Here are some examples of that.

When a record is processed, its key content is hashed so that its new downstream partition will be consistent with all other records with the same key. When records are appended, they follow the correct offset order, even in the presence of failures or faults. When a stream's key content changes because of how a query wants to process the rows (via GROUP BY or PARTITION BY), the underlying records keys are recalculated, and the records are sent to a new partition in the new topic set to perform the computation.

Producers and consumers

Producers and consumers facilitate the movement of records to and from topics. When an application wants to either publish records or subscribe to them, it invokes the APIs (generally called the client) to do so. Clients communicate with the brokers (see below) over a structured network protocol.

When consumers read records from a topic, they never delete them or mutate them in any way. This pattern of being able to repeatedly read the same information is helpful for building multiple applications over the same data set in a non-conflicting way. It's also the primary building block for supporting "replay", where an application can rewind its event stream and read old information again.

Producers and consumers expose a fairly low-level API. You need to construct your own records, manage their schemas, configure their serialization, and handle what you send where.

ksqlDB behaves as a high-level, continuous producer and consumer. You simply declare the shape of your records, then issue high-level SQL commands that describe how to populate, alter, and query the data. These SQL programs are translated into low-level client API invocations that take care of the details for you.


The brokers are servers that store and manage access to topics. Multiple brokers can cluster together to replicate topics in a highly-available, fault-tolerant manner. Clients communicate with the brokers to read and write records.

When you run a ksqlDB server or cluster, each of its nodes communicates with the Kafka brokers to do its processing. From the Kafka brokers' point of view, each ksqlDB server is like a client. No processing takes place on the broker. ksqlDB's servers do all of their computation on their own nodes.


Because no data format is a perfect fit for all problems, Kafka was designed to be agnostic to the data contents in the key and value portions of its records. When records move from client to broker, the user payload (key and value) must be transformed to byte arrays. This enables Kafka to work with an opaque series of bytes without needing to know anything about what they are. When records are delivered to a consumer, those byte arrays need to be transformed back into their original topics to be meaningful to the application. The processes that convert to and from byte representations are called serialization and deserialization, respectively.

When a producer sends a record to a topic, it must decide which serializers to use to convert the key and value to byte arrays. The key and value serializers are chosen independently. When a consumer receives a record, it must decide which deserializer to use to convert the byte arrays back to their original values. Serializers and deserializers come in pairs. If you use a different deserializer, you won't be able to make sense of the byte contents.

ksqlDB raises the abstraction of serialization substantially. Instead of configuring serializers manually, you declare formats using configuration options at stream/table creation time. Instead of having to keep track of which topics are serialized which way, ksqlDB maintains metadata about the byte representations of each stream and table. Consumers are configured automatically to use the correct deserializers.


Although the records serialized to Kafka are opaque bytes, they must have some rules about their structure to make it possible to process them. One aspect of this structure is the schema of the data, which defines its shape and fields. Is it an integer? Is it a map with keys foo, bar, and baz? Something else?

Without any mechanism for enforcement, schemas are implicit. A consumer, somehow, needs to know the form of the produced data. Frequently this happens by getting a group of people to agree verbally on the schema. This approach, however, is error prone. It's often better if the schema can be managed centrally, audited, and enforced programmatically.

Confluent Schema Registry, a project outside of Kafka, helps with schema management. Schema Registry enables producers to register a topic with a schema so that when any further data is produced, it is rejected if it doesn't conform to the schema. Consumers can consult Schema Registry to find the schema for topics they don't know about.

Rather than having you glue together producers, consumers, and schema configuration, ksqlDB integrates transparently with Schema Registry. By enabling a configuration option so that the two systems can talk to each other, ksqlDB stores all stream and table schemas in Schema Registry. These schemas can then be downloaded and used by any application working with ksqlDB data. Moreover, ksqlDB can infer the schemas of existing topics automatically, so that you don't need to declare their structure when you define the stream or table over it.

Consumer groups

When a consumer program boots up, it registers itself into a consumer group, which multiple consumers can enter. Each time a record is eligible to be consumed, exactly one consumer in the group reads it. This effectively provides a way for a set of processes to coordinate and load balance the consumption of records.

Because the records in a single topic are meant to be consumed by one process in the group, each partition in the subscription is read by only one consumer at a time. The number of partitions that each consumer is responsible for is defined by the total number of source partitions divided by the number of consumers. If a consumer dynamically joins the group, the ownership is recomputed and the partitions reassigned. If a consumer leaves the group, the same computation takes place.

ksqlDB builds on this powerful load balancing primitive. When you deploy a persistent query to a cluster of ksqlDB servers, the workload is distributed across the cluster according to the number of source partitions. You don't need to manage group membership explicitly, because all of this happens automatically.

For example, if you deploy a persistent query with ten source partitions to a ksqlDB cluster with two nodes, each node processes five partitions. If you lose a server, the sole remaining server will rebalance automatically and process all ten. If you now add four more servers, each rebalances to process two partitions.

Retention and compaction

It is often desirable to clean up older records after some period of time. Retention and compaction are two different options for doing this. They are both optional and can be used in conjunction.

Retention defines how long a record is stored before it's deleted. Retention is one of the only ways to delete a record in a topic. This parameter is particularly important in stream processing because it defines the time horizon that you can replay a stream of events. Replay is useful if you're fixing a bug, building a new application, or backtesting some existing piece of logic.

ksqlDB enables you to control the retention of the underlying topics of base streams and tables directly, so it's important to understand the concept. For more information see Topics and Logs in the Kafka docs.

Compaction, by contrast, is a process that runs in the background on each Kafka broker that periodically deletes all but the latest record per key. It is an optional, opt-in process. Compaction is particularly useful when your records represent some kind of updates to a piece of a state, and the latest update is the only one that matters in the end.

ksqlDB directly leverages compaction to support the underlying changelogs that back its materialized tables. They allow ksqlDB to store the minimum amount of information needed to rebuild a table in the event of a failover. For more information see Log Compaction in the Kafka docs.

Last update: 2021-04-05