Skip to main content
Cost control is one of the sections in the Insights dashboard. It helps platform teams identify topics consuming unnecessary storage and make informed cleanup and retention decisions. Cost control dashboard
Free ebook: Where Kafka costs hide: a field guide Where Kafka spend actually accumulates, and how to find and cut it.

​
Overview

The Cost Control section displays three categories of topics that may be wasting storage:
  • Empty topics - topics with no messages that consume metadata overhead
  • Stale topics - topics with no writes in a configured time period (default: 7+ days)
  • Tiny topics - topics with minimal data volume relative to their partition overhead
Each category shows the number of affected topics. Individual topics are listed with idle time in days and partition count. You can switch the view of data on the page between a graphs view and a table view using the “Graphs” toggle found at the top right of the page. Before taking action on any topic, verify it is truly unused by checking consumer groups, producer configurations, ownership and documentation.
Deleting a topic is permanent and cannot be undone. All partition metadata and configuration will be lost.

​
What the dashboard shows

​
Empty topics

Topics containing zero messages. These topics have been created but never received data, or have had all messages deleted through retention policies. For each topic: name, idle time in days, and partition count. Empty topics consume cluster resources even without data: metadata storage in ZooKeeper or KRaft, memory for buffers, and often per-partition charges on cloud Kafka services.
Empty topics are common in development and staging environments where topics are created for testing but not cleaned up afterward.
Good candidates for deletion:
  • topics created for abandoned projects or experiments
  • temporary topics used for testing or debugging
  • topics created by mistake with typos in the name
  • legacy topics from deprecated applications
  • topics created automatically by frameworks but never used
Reasons to keep empty topics:
  • topics pre-created for upcoming features or applications
  • topics serving as placeholders for future data pipelines
  • topics referenced in consumer group configurations awaiting data
  • topics used as dead letter queues that only receive messages in error scenarios
1

Verify the topic is unused

Click the topic name to go to the Topics page. Check:
  • Consumer Groups tab for active consumers or unconsumed data (lag > 0)
  • Monitoring tab graphs to confirm zero produce activity
  • Application references in codebases and deployment configs
  • Dependencies - verify no streaming jobs or external systems depend on the topic
2

Delete from the list

In the Cost Control section, click the … menu next to the topic and select Delete topic.
3

Confirm deletion

Type DELETE to confirm and click the delete button.
Topic deletion can take up to 5 minutes to be reflected in the dashboard. Consumer groups that were subscribed will need to be updated to remove the topic from their subscriptions.

​
Stale topics

Topics with no write activity for 7+ days and no active consumer groups. For each topic: name, idle time in days, and partition count. Stale topics often indicate abandoned data pipelines from shut down applications, failed producers or cancelled projects. They continue consuming storage through retention policies and add operational overhead.
A topic with no writes may still have active consumers reading historical data. Stale write activity does not necessarily mean the topic is unused or safe to delete.
Investigation checklist:
  1. Check consumer groups - Are any consumers actively reading from the topic?
  2. Review consumer lag - Is there unconsumed data that applications still need?
  3. Verify business requirements - Is the historical data needed for compliance or analytics?
  4. Contact topic owners - Do the application teams still need this topic?
  5. Check data lineage - Do downstream systems depend on this data?
Topics that are safe to delete:
  • No active consumer groups subscribed
  • All consumers show zero lag (all data has been consumed)
  • Application owners confirm the topic is no longer needed
  • No compliance requirements for data retention
  • Historical data has been archived to long-term storage
Topics to keep:
  • Active consumers are still reading historical data
  • Data is required for compliance or audit purposes
  • Topic serves as a source of truth for event sourcing
  • Downstream analytics or reporting systems depend on the data
  • Occasional writes expected (seasonal or infrequent events)
Deleting a topic with unconsumed messages will result in permanent data loss for consumers that have not yet processed those messages.
For topics containing valuable historical data, export and archive the data before deletion.
1

Export the topic data

Navigate to the topic, click the Consume tab, configure the consumer to read from Beginning offset, and export data to a file or external storage.
For large topics, use a dedicated archival tool or Kafka Connect to export data to object storage (S3, Azure Blob Storage) or a data lake.
2

Verify the archive

Confirm all messages were exported successfully, validate the data format and completeness, and store in long-term storage with documentation of location and format.
Maintain a record of archived topics including the archive location, date, original message count and retention period for the archived data.
3

Delete the topic

Follow the deletion steps above.
If a topic might receive future writes but historical data is no longer needed, reduce the retention period instead of deleting.
1

Update retention settings

Navigate to Topics, select the topic, click Configuration, and edit:
  • retention.ms - Time-based retention (e.g., 86400000 for 1 day)
  • retention.bytes - Size-based retention per partition (e.g., 1073741824 for 1 GB)
Kafka automatically deletes segments older than the retention period during its next cleanup cycle (typically runs every few minutes based on broker configuration).
2

Monitor cleanup

Monitor the topic size in the Overview tab to verify that old data is being cleaned up as expected.

​
Tiny topics

Topics meeting all of the following: more than 1 partition, less than 10 MB of data OR fewer than 1000 messages, and name does not end with _repartition. For each topic: name, idle time in days, and partition count. Topics with many partitions but little data represent inefficient resource allocation. Cloud Kafka services often charge per partition, and high partition counts increase controller overhead.
A common cause is using default partition counts (often 6-12) for topics that only receive occasional messages, such as configuration topics or low-volume event streams.
How to evaluate:Keep the topic as-is if it expects significant growth, requires high parallelism, is part of a standardized policy or would require complex migration. Consider optimization if the topic has been tiny for months without growth, has arbitrary partition counts or handles low-frequency data without parallelism requirements.Since Kafka does not support reducing partition count, you have three options:
If you have multiple tiny topics serving similar purposes, consolidate them into a single topic with categorization through message headers or keys. This reduces total partition count while maintaining logical separation.Considerations:
  • Identify candidates from the same application with similar schemas
  • Design consolidated topic structure with category or type fields
  • Update producers and consumers accordingly
Topic consolidation changes the partition key distribution and may affect message ordering. Ensure consumers can handle messages from multiple original topics in a single stream.

​
Our recommendations

Prevent cost control issues through governance and automated ownership tracking rather than manual remediation. Use Self-service for ownership and governance - it brings a GitOps approach to topic lifecycle management where applications represent streaming apps or data pipelines and dictate ownership of Kafka resources. Benefits for cost control:
  • Automatic ownership tracking - topics created through Self-service have clear ownership from creation
  • Policy enforcement - application instance policies can prevent creation of tiny topics by enforcing minimum data volume or maximum partition count requirements
  • Naming conventions - policies enforce consistent naming that includes environment and ownership information
  • Lifecycle management - clear application ownership reduces abandoned topics through better tracking and accountability
Learn about Self-service Configure RBAC - Use role-based access control to enforce naming conventions, limit partition counts and require labeling at topic creation. Learn more about RBAC Implement cleanup policies - Document organization-wide policies for retention periods, deletion criteria and approval processes. Schedule regular reviews of cost control reports (weekly for new topics, monthly for trends).

​
Troubleshoot

Producer health checks, monitoring systems or automated tests may send periodic messages that are immediately consumed and deleted. Kafka’s log cleaner activity can also appear as writes in some monitoring systems.Check the Messages in per second graph over 7+ days to distinguish sustained activity from sporadic checks. Review the Consume tab to see actual message content.
No, topic deletion is permanent. All partition data and topic metadata are removed. Consumer group offsets are retained but the topic cannot be recreated with the same history.Always archive important data before deletion. Implement a grace period policy where topics are first marked for deletion (using labels) before actual removal. To recreate a deleted topic, create a new topic with the same name in Console and restore data from archives if available.
Kafka creates internal topics automatically: __consumer_offsets (consumer group offsets), __transaction_state (transactional state), _schemas (Schema Registry metadata) and topics matching .*-changelog or .*-repartition (Kafka Streams internals).Never delete __consumer_offsets or __transaction_state as this causes cluster-wide failures. Only delete Kafka Streams topics when the application is permanently decommissioned and all instances are shut down.
Kafka’s log cleanup process runs periodically (default: every 5 minutes). Data is deleted at segment boundaries, not individual messages, and the current active segment is never deleted even if older than retention.Wait 15-30 minutes after reducing retention, then check topic size in the Overview tab.
Check topic naming patterns, review consumer groups in the Consumer Groups tab, search code repositories and deployment configs for references, and consult internal documentation.If ownership remains unclear, add a scheduled-for-deletion label with a future date (for example, 30 days) and send announcements asking owners to identify themselves. After the grace period expires with no claims, proceed with deletion.Implement a policy requiring application labels at topic creation time. Use RBAC permissions to enforce labeling as a prerequisite for topic creation.
This is normal for seasonal data (holiday sales, tax season), emergency systems (disaster recovery), compliance events (audit logs) or legacy integrations (backup systems).Verify the producer is authorized, add a usage-pattern: seasonal or usage-pattern: emergency label, and document the activity pattern. Do not delete topics showing this pattern without understanding the business requirement.