Home › AWS › AWS data services › AWS DMS

AWS data services · Ingestion and CDC · about 12 min read

AWS Database Migration Service (DMS)

Copies a database once, then keeps copying every insert, update and delete, here as change files into S3 landing.

RoleIngestion and CDC
StatusGenerally available. DMS Serverless and DMS Schema Conversion are generally available. DMS Fleet Advisor reached end of support on 20 May 2026; AWS points to Migration Evaluator instead.
Verdict on this sitePursue
AzureAzure Database Migration Service for moves; Data Factory or Fabric Mirroring for ongoing change capture
Google CloudDatastream (change capture) and Database Migration Service (moves)
DatabricksLakeflow Connect database connectors, which write straight into tables
Open sourceDebezium on Kafka Connect
Verdict: Pursue. The default mover for MES, quality and other database sources: a full load plus ongoing change capture written as Parquet files into S3 landing, so every source passes the same lander, tokenising and run log. Lakeflow Connect is the alternative to revisit per source later.

In plain words

What it is

AWS Database Migration Service, usually called DMS, reads rows out of one database and writes them somewhere else. It does this in two phases. First comes the full load, a one-time copy of every row in the tables you chose. Then comes change data capture, or CDC, which means DMS keeps reading the source database's own transaction log and replays every insert, update and delete that happens after the full load. The source keeps running the whole time, which is why DMS was built for migrations without downtime.

On a data platform the migration part matters less than the second phase. Take a tyre plant whose manufacturing execution system (MES) runs on SQL Server and whose quality lab system runs on PostgreSQL. Analysts want yesterday's batches, rejects and lab results in the lakehouse, and they want deleted rows to disappear there too. DMS connects to each database, copies the history once, and then writes a small Parquet file every few minutes holding the changed rows, each tagged with an operation flag (I for insert, U for update, D for delete) and a commit timestamp.

DMS comes in two shapes. DMS Standard runs on a replication instance, a virtual machine you size and pay for by the hour, which runs one or more tasks. DMS Serverless has no instance to size; you set a minimum and maximum capacity in DMS capacity units (DCUs) and it scales between them. Next to both sit DMS Schema Conversion, a free tool that translates schemas between engines, and homogeneous data migrations, which use the database's own native tools for like-to-like moves.

A clerk who photocopies the whole ledger once, then sits beside it and copies every new line onto a numbered slip, dropping a bundle of slips in the inbox every few minutes.

Big picture

Where it fits

SourcesMove inStreamStoreProcessServeUseAcross everything: catalog · security · orchestration · observability and cost · infrastructure as codegets data fromhands data toSQL Server, Oracle, PostgreSQL, MySQL (MES, LIMS)RDS PostgreSQLAurora PostgreSQLAWS DMSIngestion and CDCS3Databricks Auto Loader into bronzeKinesis Data Streams

The lit lane is where AWS DMS works. Boxes on the left feed it; boxes on the right use what it produces.

DMS sits in the move-in lane of the reference platform, on the path for relational databases. The MES and quality databases are the sources. DMS runs the full load and then ongoing CDC, and writes Parquet files into the S3 landing bucket for that catalog, one folder per table. It does not write into Delta tables and does not touch bronze.

From there the platform's own lander takes over. A Databricks job runs every 15 minutes, Auto Loader reads the new files since its checkpoint, tokenises personal fields, stamps an ingest run id and appends every change row into bronze, which keeps everything forever. Silver applies the changes in order, either with a MERGE or with AUTO CDC in Lakeflow Declarative Pipelines, which the streaming page rates Pursue after a spike for building SCD2 history from DMS changes.

Why files and not a stream? The site's rule is that Kinesis carries records and S3 events carry files. Database changes rarely need a seconds-level reader, so the decision recorded on the streaming page is DMS into S3 files by default, with DMS into Kinesis only if a live reader appears. Writing to S3 also means every source, whether DMS, AppFlow or Transfer Family, lands the same way and passes the same checks. CloudWatch watches DMS lag and failures and those metrics flow into the platform's mover-health table.

How it works

The ideas you need

IdeaWhat it means
Full loadThe one-time copy of every existing row in the selected tables, done before or alongside change capture.
Change data capture (CDC)Reading the database's transaction log to pick up every insert, update and delete after a point in time, without querying the tables.
Replication instanceThe virtual machine that runs DMS Standard tasks. You choose its class and storage and pay per hour while it exists.
Replication taskOne job that moves a set of tables from one source endpoint to one target endpoint, with its own settings and table mappings.
EndpointA saved connection to a source or target: host, port, credentials (ideally from Secrets Manager) and engine-specific settings such as the S3 file format.
Table mappingJSON rules that pick which schemas and tables to include or exclude and how to rename or add columns on the way.
DMS Serverless and DCUThe no-instance option. A DCU is a DMS capacity unit; you set a minimum and maximum and pay for what is used each hour.
LOB modeHow DMS handles large objects such as long text or binary columns. Full LOB mode is slow; limited LOB mode truncates anything over the size you set.
ValidationAn optional task feature that compares rows on source and target after the load and reports mismatches.

Choosing

When to use it, and when not

Use it when

  • You need ongoing change capture from MES, LIMS, ERP side databases or app databases into S3 landing, with deletes preserved.
  • The sources are common engines that DMS supports, such as SQL Server, Oracle, PostgreSQL, MySQL, Db2 or MongoDB.
  • You want every database source to land as files in S3 so it passes the same lander, tokenising and run log as every other source.
  • You are actually migrating a database to AWS and want little or no downtime during the cut-over.
  • A team wants a managed AWS service rather than running Debezium and Kafka Connect themselves.

Avoid it when

  • The source is SAP S/4HANA tables. The site rule is never to read SAP tables directly; use a certified route such as AppFlow with OData on released CDS views or a certified extractor.
  • A dashboard needs each change within seconds. Files every few minutes are the default; that case wants DMS into Kinesis or Debezium on MSK.
  • You want changes written straight into governed Delta tables with no landing zone. Lakeflow Connect does that, at the cost of skipping the shared lander.
  • The source database team will not enable log-based capture (supplemental logging, binlog in row format, logical replication). Without it DMS can only do full reloads.
  • The tables are huge and change their structure often. DMS handles some schema changes, but frequent DDL causes task errors and reloads.

Production grade

What goes wrong in production

ProblemWhat breaksWhat to do
Source log not kept long enoughIf the task stops for a few hours and the database purges its transaction log, DMS cannot resume. The only fix is a fresh full load.Set log retention on the source to outlast a long outage (for example 24 hours or more) and alarm when a task stops.
Replication slot fills the PostgreSQL diskOn PostgreSQL, a stopped task leaves its replication slot holding write-ahead log. The source disk fills and the production database can go down.Watch slot lag on the source, drop slots for retired tasks, and agree a maximum slot size with the database team.
Latency creeps upCDCLatencySource or CDCLatencyTarget climbs during busy shifts because the instance is too small or a large transaction is being applied.Alarm on both latency metrics; resize the instance or raise Serverless maximum DCUs; split very busy tables into their own task.
Large objects truncatedLimited LOB mode cuts long text or binary columns at the size you set, silently, so notes or images arrive incomplete.Measure the largest LOB per column first; set the limit above it or use inline LOB mode; add a length check in silver.
Out-of-order or duplicate change rowsAfter a restart DMS may resend some changes, and changes from a big transaction can span files. Naive appends double-count or apply an old update last.Keep the commit timestamp and change sequence in bronze; dedupe and order by them in silver; let AUTO CDC handle out-of-order SCD2.
Schema changes on the sourceA new column or a changed type arrives mid-stream. Parquet files change shape and downstream code breaks or the task fails.Enable Auto Loader schema evolution with a rescued data column, keep a data contract per source, and alert the owning team on DDL.
Instance version and maintenanceDMS Standard engine versions reach end of support and must be upgraded; maintenance windows restart tasks.Track engine versions in code, schedule upgrades, or move steady workloads to DMS Serverless, which AWS keeps patched.
Credentials in endpointsTeams paste database passwords into endpoint settings, where they are hard to rotate.Point endpoints at Secrets Manager with a role DMS can assume, and rotate the secret on a schedule.

Alternatives

What to compare it with

AlternativePick the alternative whenPick AWS DMS when
Lakeflow Connect (Databricks)You want changes written straight into Unity Catalog tables with no landing files and one less AWS service.You want every source to pass the same S3 landing, tokenising and run log, which is the v1 design here.
Debezium on MSK ConnectYou already run Kafka, want sub-second change events, or need open-source and portable change capture.Nobody wants to run connectors, Kafka and offsets; DMS is simpler inside AWS.
Zero-ETL integrationsThe source is Aurora, RDS or DynamoDB and the target is Redshift or the SageMaker lakehouse, with no transformation in between.The target is Databricks on S3 landing; zero-ETL does not write files for your own lander.
Azure Data Factory and Fabric MirroringThe platform is on Azure; Mirroring copies databases into OneLake as Delta with little setup.You are on AWS; cross-cloud change capture adds egress and another control plane.
Google DatastreamThe platform is on Google Cloud; Datastream writes changes into BigQuery or Cloud Storage.You are on AWS with Databricks.
Qlik Replicate or FivetranYou need sources DMS does not cover well, a strong SAP story, or a vendor that runs the connectors for you.Licence cost matters and DMS already covers your engines.

Cost

How you pay

DMS Standard bills each replication instance per hour by instance class while it exists, whether a task is moving data or not, plus storage beyond the included amount (used for logs and for caching changes when the target is slow). DMS Serverless bills per DCU-hour, scaling between the minimum and maximum you set, so the maximum caps the hourly spend. DMS Schema Conversion is free apart from the S3 storage it uses. Homogeneous migrations bill per hour while they run. Data coming into DMS is free; cross-AZ, cross-Region and outbound transfer are charged at normal rates. Check current list prices for your region.

What drives the bill

  • Instance class and hours: an instance running all month for continuous CDC is the main line item.
  • Multi-AZ, which roughly doubles instance cost for a standby.
  • Serverless maximum DCUs and how busy the source is during each hour.
  • Extra storage when the target falls behind and changes are cached on disk.
  • Downstream costs it creates: S3 PUT requests for many small CDC files and the Databricks job that reads them.

Cost traps

  • Leaving test replication instances running after a proof of concept; they bill every hour.
  • Writing very small CDC files every few seconds, which multiplies S3 requests and slows Auto Loader; tune the batch interval and minimum file size.
  • Choosing a burstable T-class instance for a busy source; it runs out of CPU credits and either throttles or bills extra.
  • Running the replication instance in another Availability Zone or Region from the source and paying for transfer on every change.

Prices change; check the AWS pricing page for your region before deciding.

Security and governance

Keeping it safe

  • Run replication instances in private subnets with no public IP, reaching sources over Direct Connect or VPN and S3 through a gateway VPC endpoint.
  • Store source credentials in Secrets Manager and grant DMS a role that can read only those secrets; never type passwords into endpoint settings.
  • Encrypt the replication instance storage and the S3 target with the catalog's KMS key, and use TLS for source connections.
  • Give the S3 target role write access only to its own landing prefix, so one task cannot overwrite another source's files.
  • Changes include personal data such as operator names or customer details; Macie scans the landing bucket and the lander tokenises those columns before bronze.
  • CloudTrail records who created, modified or started tasks and endpoints; keep that trail in the audit account.

Operating it

What to watch, and the limits

WatchWhy
CDCLatencySourceSeconds between a change committing on the source and DMS reading it. Rising means DMS cannot keep up with the log.
CDCLatencyTargetSeconds between a change committing on the source and being written to the target. This is the freshness the business sees.
CDCIncomingChangesChanges waiting to be applied. A growing number means the target or instance is the bottleneck.
FreeStorageSpaceDisk left on the replication instance. Low values mean changes are being cached because the target is slow.
CPUUtilization and FreeableMemoryInstance pressure; sustained high CPU or low memory means it is time to resize or split tasks.
Task status eventsEventBridge events or DMS event subscriptions for task stopped or failed, sent to on-call through SNS.
LimitValue and notes
Replication instances60 per account per Region (default quota)
Tasks600 per account and 200 per replication instance
Endpoints1,000 per account and 100 per replication instance
Instance storage30,000 GB total across replication instances in a Region
Serverless replications100 per account per Region
API rate200 requests per second burst, refilled at 8 per second; cannot be raised

Limits and quotas checked 2026-11-11; many can be raised through Service Quotas.

Decisions

Situations and verdicts

SituationChoiceVerdict
MES on SQL Server, changes needed every 15 minutesDMS full load plus CDC, Parquet files into S3 landingPursue: matches the platform's file path; Auto Loader reads them on the 15-minute run and deletes arrive as D rows.
A live screen wants order changes within secondsDMS into Kinesis, or Debezium on MSKPursue with care: only when a real live reader exists; it adds a stream to run and pay for.
Twenty small app databases, low change volumeDMS Serverless with a low maximum DCU per replicationPursue with care: no instance sizing, but check the capacity it picks and cap the maximum.
SAP S/4HANA order tablesDMS reading SAP database tables directlyDrop: unsupported by SAP licensing rules on the site; use AppFlow OData or a certified extractor.
A source that Lakeflow Connect supports nativelyLakeflow Connect straight into tablesKeep if you have it: fine where already running; for v1 the shared S3 landing wins.

FAQ

Questions people ask

Can AWS DMS write to Delta Lake?

Not directly. DMS writes CSV or Parquet files to S3, and Databricks Auto Loader reads them into a Delta bronze table. For a straight-into-tables route, Databricks Lakeflow Connect is the native option.

Is DMS only for one-off migrations?

No. A task can run full load only, CDC only, or full load followed by ongoing CDC for as long as you need. Many data platforms run DMS CDC tasks for years.

How fresh is DMS data in S3?

Usually a few minutes behind the source, set by the CDC batch interval and file size settings. The platform reads it on a 15-minute schedule, so end-to-end freshness is around 15 to 20 minutes.

Should I use DMS Standard or DMS Serverless?

Serverless suits variable or many small workloads and removes instance sizing and patching. Standard gives predictable hourly cost for steady busy sources and more tuning control. Check that your source and target are supported by Serverless first.

What happened to DMS Fleet Advisor?

It stopped taking new customers in May 2025 and reached end of support on 20 May 2026. AWS recommends Migration Evaluator for database inventory and sizing. The core DMS replication service is unaffected.

Sources

Checked against

Checked 2026-11-11.

Related

Where it connects

Common Ingestion Framework · Data platform on AWS and Databricks · SCD2: AUTO CDC vs PySpark · Streaming on AWS and Databricks · Zero-ETL integrations · Amazon S3 · All AWS data services

← CodePipeline and CodeBuild · AppFlow →