Debezium vs AWS DMS for change data capture
Both read a database's transaction log and replay the changes elsewhere. They differ in who runs them, where the changes can go and what happens when something breaks. Here is how to choose.
45 minutes with an engineer: we review your workload and sketch the design.
- ✓Docs read 18 Nov 2026
- ✓OSStend runs Debezium
Debezium is open-source CDC you operate, with Kafka and Debezium Server sinks. AWS DMS is an AWS service you configure as tasks, with a fixed list of sources and targets. Choose DMS for AWS-centred replication and migrations; choose Debezium for event streams and wider sinks. Docs read 18 Nov 2026.
- PostgreSQL
- MySQL
- MongoDB
- Redis
- Valkey
- OpenSearch
- Kafka
- RabbitMQ
- AWS
- Google Cloud
- Azure
- Kubernetes
- OpenShift
- Prometheus
- Grafana
- Datadog
- Dynatrace
- Splunk
Debezium vs AWS DMS side by side
Debezium and AWS DMS both do log-based CDC. The difference is the shape of the product: Debezium is a component you run in a streaming architecture, and DMS is a service that runs replication tasks between supported endpoints.
| Debezium | AWS DMS | |
|---|---|---|
| What it is | Open-source CDC platform, Apache License 2.0 | AWS service that migrates data and replicates ongoing changes |
| How you use it | Kafka Connect connectors, Debezium Server or the embedded engine | Source and target endpoints plus a replication task that runs on a replication instance |
| Sources | PostgreSQL, MySQL, MariaDB, SQL Server, Oracle, MongoDB, Db2, Cassandra and others (connectors) | Nearly all popular database engines as sources, per AWS (documentation) |
| Targets | Kafka topics through Kafka Connect; Debezium Server sinks include Kafka, Kinesis, Pub/Sub, RabbitMQ and others (sinks) | A fixed list that includes databases, Amazon S3, Amazon Kinesis, Apache Kafka, OpenSearch and others (targets) |
| Event shape | One event per change with before and after state, schema history and heartbeats; consumers handle the stream | Changes are applied to the target; to Kafka or Kinesis it writes records for consumers |
| Who runs it | You, or a provider; connectors, offsets and Kafka are yours | AWS runs the replication instance and manages its software; you own tasks and tuning |
| Where it runs | Anywhere: your cloud, a data center, Kubernetes | Within AWS, with sources and targets reachable from it |
| Cost basis | Software is free; you pay for infrastructure and operations | Pay as you go for DMS resources (AWS pricing) |
Sources: Debezium connector index, Debezium Server docs and license; AWS DMS overview and targets; all checked 18 Nov 2026. Neither tool was installed or benchmarked and we quote no performance figure or price.
Can DMS send changes to Kafka, and is Debezium AWS-only or open?
DMS can write to Apache Kafka and Amazon Kinesis as well as to databases and Amazon S3, so it is not limited to database targets. Debezium is not tied to AWS at all, and its sink list through Debezium Server is wider than a fixed endpoint list.
AWS documents Apache Kafka and Amazon Kinesis Data Streams as DMS targets (targets, checked 18 Nov 2026). The difference is in how the changes arrive: DMS carries them as task output, Debezium models them as change events with schema history, which suits teams building streams with several consumers. Debezium Server's documented sinks also include Pub/Sub, RabbitMQ, Redis streams, NATS, Pulsar and others (docs, checked 18 Nov 2026).
What happens when the source log is not there?
Both depend on the source keeping its change log until they have read it. AWS documents that a DMS task on SQL Server can lose events if the log is archived or truncated before DMS reads it, and Debezium has the same dependency on the source's retention.
- DMS and SQL Server.
- AWS's documentation says events may be lost if SQL Server archives or truncates them from the active transaction log before DMS reads them, and lists limits such as truncate commands not being captured and renamed tables or columns not being supported (SQL Server source docs, checked 18 Nov 2026).
- Debezium and SQL Server.
- The connector reads SQL Server's own CDC change tables. Its documentation covers snapshot modes (including re-snapshotting when a stored offset points at a position the server no longer has) and a schema history topic; the cleanup job on the source still decides how long changes exist (connector docs, checked 18 Nov 2026).
- The common part.
- In both, someone has to size log retention for the longest outage, alert when the reader falls behind, and have a re-snapshot plan. A tool does not remove that work.
When does each one fit?
Choose DMS for migrations and AWS-centred replication where you want AWS to run the infrastructure. Choose Debezium when changes are events many systems consume, when you need sinks or sources DMS does not offer, or when the pipeline must run outside AWS.
| Your situation | Usual fit | Why |
|---|---|---|
| One-time or phased database migration into AWS | AWS DMS | It is built for migration, with schema conversion tooling alongside it |
| Ongoing replication between AWS databases or into S3 | AWS DMS | AWS runs the replication instance and you configure tasks |
| Change events feed several services, a warehouse and a search index | Debezium | Events land in Kafka with schema history, and many consumers read them |
| Pipeline runs in your data center or another cloud | Debezium | It is software you can run anywhere |
| Source or sink is outside DMS's list | Debezium | Its connector and sink list is wider and extendable |
| Nobody wants to operate CDC software | DMS, or a provider running Debezium | Debezium leaves the operations to you; see managed Debezium |
OSStend runs Debezium and sells operations for it, so weigh this as a vendor's view. DMS is the better choice when its endpoints cover your case and a task-based service is what you want.
From our on-call engineers: why one customer chose Debezium over DMS
On one engagement the customer chose Debezium over DMS for a SQL Server pipeline that crossed clouds. We repeat the reasoning because it shows what to test, not because Debezium is better everywhere.
- ✓The worry with DMS was log access: the source's maintenance removed transaction logs, so a log-reading task could drift or lose access. Debezium would read the CDC change tables directly.
- ✓Whichever tool you pick, test recovery: while the capture tool is down the source has to keep its change logs, so source-side cleanup is paused until it has caught up.
- ✓Proofs of concept have missed delete events until the configuration was corrected, so delete capture is a named test in our reviews.
- ✓We have designed and hardened a standalone Debezium Server pipeline from SQL Server to RabbitMQ on Kubernetes, including node-failure and broker-outage tests with runbooks.
- ✓Schema evolution is the part teams leave untested; we run a schema change through the whole pipeline before go-live.
For the choice in the wider context, read change data capture explained. To have Debezium run for you, see managed Debezium.
FAQ
Is Debezium better than AWS DMS?
Neither is better everywhere. DMS fits migrations and AWS-centred replication. Debezium fits event streams, wider sinks and pipelines that run outside AWS, at the price of operating it.
Does AWS DMS use CDC?
Yes. AWS says DMS can replicate ongoing changes to keep sources and targets in sync, in addition to one-time migrations (overview, checked 18 Nov 2026).
Can DMS write to Kafka?
Yes. Apache Kafka is on the list of DMS targets (targets, checked 18 Nov 2026).
Is Debezium free?
The software is open source under the Apache License 2.0. You pay for the infrastructure and for whoever operates it.
Can I replace DMS with Debezium on AWS?
Yes: Debezium runs on Amazon MSK Connect or on your own Kafka Connect, and the move is done by running both until the targets agree. Managed Debezium covers who operates it.
Debezium is a trademark of its respective owner and Apache Kafka a trademark of the Apache Software Foundation. AWS and AWS DMS are trademarks of Amazon. OSStend is not affiliated with or endorsed by these projects or by Amazon.
Bring the decision to an engineer
45 minutes: your sources, targets and recovery needs are reviewed and an engineer says which tool fits.
- ✓A design sketch for your workload, not a plan menu
- ✓A zero-downtime cutover path for what you run today
- ✓A price for that design, after the review