Skip to main content
Hello and welcome back. In this lesson we will learn about offset management in Apache Kafka — what offsets are, how consumers use them, common commit strategies, and why correct offset handling matters for correctness and reliability. Overview
  • An offset in Kafka is a partition-scoped, monotonically increasing integer that identifies a message’s position within a partition.
  • Consumers use offsets as a “bookmark” to remember where to resume reading after a restart or after a rebalance.
  • Proper offset management prevents duplicate processing and data loss in downstream systems.
Analogy: imagine you are reading a book. After an hour you place a bookmark to remember where to continue. That bookmark tells you which page you have read up to and where to resume next. In Kafka, the “book” is the stream of events, the “bookmark” is the offset, and the reader is the consumer. Each message in Kafka within a partition has a unique integer called an offset. Consumers read messages and use offsets to track how far they’ve progressed.
The image shows a diagram for Kafka offset management, featuring a broker with Topic A, Partition 1, and a sequence of offsets numbered from 0 to 10.
How offsets are scoped and incremented
  • Offsets start at 0 for each partition and increase monotonically for that partition.
  • Offsets are scoped to a partition — the same numeric offset can exist in multiple partitions but point to different messages.
  • The offset identifies the position of a message; consumers use the offset of a processed message to determine the next message to read.
Consumer groups and committed offsets When multiple consumers work together in a consumer group, Kafka assigns partitions to consumers. Each consumer reads from its assigned partitions and should record progress by committing offsets. The committed offset indicates the position from which the consumer group should resume next time (usually the next message to read). Consider a consumer reading from Topic A as part of a consumer group:
The image illustrates Kafka offset management, showing a broker with a topic partition and a consumer group with a single consumer.
A typical processing flow
  1. Consumer polls messages from an assigned partition.
  2. Application processes each message (for example, writes a record to a commissions database).
  3. After successful processing, the consumer commits the offset to indicate progress.
  4. If the consumer crashes, a new consumer that takes over will read the last committed offset and resume from there.
When the consumer has processed the message at offset 3, it records a committed offset indicating progress — that is, the position from which the consumer group should resume next time (generally the next message to read). This committed offset prevents reprocessing already-handled messages when a new consumer takes over.
The image illustrates Kafka offset management, showing a broker with a topic partition, current and committed offsets, a consumer group, and a commission database.
Why committed offsets matter Offsets are per-partition. If a consumer crashes and a different consumer joins the group, the new consumer consults the committed offset for each partition to decide where to start reading. Without committed offsets, the new consumer might start from earliest or latest depending on configuration, which can lead to duplicates or missing data in downstream stores. Imagine the commissions database is used to pay restaurant owners. If reprocessing occurs due to incorrect offset handling, you could accidentally pay vendors twice — a real financial risk. Proper offset management prevents this by letting the replacement consumer resume exactly where the previous one left off (for example, starting at offset 4 if 0–3 were already processed and committed). Offset management is therefore critical: Kafka assigns offsets and stores messages, but it is the consumer’s responsibility to record how much it has processed.
The image illustrates Kafka offset management tracking, showing how consumer offsets track the position of each consumer within a partition to resume from the correct position after failures or rebalancing.
Where offsets are stored
  • By default, consumers commit offsets to Kafka’s internal offsets storage, the __consumer_offsets topic.
  • Some applications prefer to persist offsets externally (for example, in a relational database or a durable key-value store) when they need tight coupling between processing and offset commits or custom recovery semantics.
Commit strategies There are two common ways to commit offsets. Choosing the right strategy affects delivery semantics (at-most-once, at-least-once, or exactly-once when used with transactional APIs).
The image illustrates Kafka's offset management and persistence, showing a consumer committing an offset to Kafka, ensuring fault tolerance and preventing message loss or duplication.
Table: Commit strategies comparison Configuration examples
  • Enable automatic commits (Java properties):
  • Manual commit (Java example):
  • Manual async commit:
Best practice: if your processing is not idempotent (i.e., you cannot safely reprocess the same message twice), prefer manual commits after successful processing or use exactly-once processing patterns (transactions or idempotent sinks) to avoid duplicates.
Automatic commits simplify development but can hide problematic edge cases; manual commits add complexity but provide stronger guarantees when combined with proper error handling and transactional techniques.
The image compares manual and automatic commit strategies in Kafka offset management, highlighting pros like simplified management and cons like data loss risk for each approach.
Rebalancing and offset recovery Rebalancing occurs when partition assignments change — for example, when consumers join or leave the group or when the number of partitions changes. When a rebalance happens, the new owners of partitions will check the last committed offsets for those partitions and resume processing from that point. This ensures continuity and helps avoid unnecessary reprocessing or message loss. Suppose a topic has four partitions and a consumer group has three consumers. Partitions are assigned across consumers; if a consumer fails or leaves, the group coordinator triggers a rebalance so the remaining consumers take over the orphaned partitions. Each new owner consults the committed offsets to continue where the previous consumer left off.
The image illustrates Kafka offset management and rebalancing, showing how partitions from "Topic A" are distributed between two consumers within a consumer group.
Practical tips
  • For non-idempotent sinks (billing, transfers), always commit offsets only after the downstream write has succeeded.
  • Consider using transactions (exactly-once semantics) for end-to-end guarantees when producing to Kafka and consuming from it in the same transactional unit.
  • If using external offset storage, ensure atomicity between processing and offset persistence, or use two-phase commit patterns where appropriate.
  • Monitor __consumer_offsets lag and consumer group rebalances — frequent rebalances can increase duplicate processing and complicate offset management.
Summary
  • Offsets are unique integers per partition that identify the position of messages.
  • Consumers track and commit offsets to record how far they have processed.
  • Committed offsets (by default stored in Kafka’s __consumer_offsets topic) allow new consumers to resume from the correct position after failures or rebalancing.
  • Automatic commits are convenient but risk duplicates or data loss; manual commits offer more control and should be used when precise guarantees are required.
References and further reading I hope this lesson clarifies the importance of offsets and how offset management works in Apache Kafka. That is it for this lesson.

Watch Video