Most enterprise data migrations run into trouble in the assessment, not the transfer. The tooling AWS provides is mature and well documented. What costs teams time is picking the wrong service for the workload, underestimating schema conversion, or planning a cutover with no rollback path.
This guide covers the AWS data migration services on offer, which service fits which job, how a migration actually runs, and what changes when the data is regulated.
What AWS data migration services cover
AWS does not sell a single migration product. It sells a set of services, each built for a different combination of source, target, data volume and downtime tolerance.
They fall into three groups:
- Database migration. Moving relational and non-relational databases between engines or into AWS. AWS DMS and the Schema Conversion Tool.
- Bulk and file transfer. Moving file systems, object storage and large datasets. DataSync, the Snow Family and Transfer Family.
- Server and application migration. Lifting whole workloads. AWS Application Migration Service.
Choosing between them starts with two questions: how much data, and how much downtime the business can absorb.
AWS data migration services compared
| Service | What it moves | Use it when |
|---|---|---|
| AWS DMS | Relational and NoSQL databases, live | The source must stay running during the move, or you need ongoing replication after it |
| AWS Schema Conversion Tool | Schemas, stored procedures, database code | Source and target run different database engines |
| AWS DataSync | File systems and object storage, over the network | You have bandwidth and want scheduled, validated, automated transfers |
| AWS Snow Family | Terabyte to petabyte volumes, on physical devices | Network transfer would take weeks, or the site has limited connectivity |
| AWS Transfer Family | Files over SFTP, FTPS and AS2 | Partners or external systems send you data over file protocols and it needs to land in Amazon S3 |
| AWS Application Migration Service | Whole servers, lift and shift | You are moving an application as it stands before modernising it |
| AWS Glue | Data in transit, with transformation | The data needs reshaping, cleaning or cataloguing on the way in |
| AWS Migration Hub | Nothing. It tracks the others | Several migrations are running and you need one view of progress |
Most enterprise migrations use three or four of these together. A live database moves through AWS DMS, its schema through the Schema Conversion Tool, its historical archive through Snowball, and the whole programme is tracked in Migration Hub.
Homogeneous and heterogeneous migration
Homogeneous migration moves data between identical or compatible database engines. Oracle to Amazon RDS for Oracle, or MySQL to Amazon Aurora MySQL. Schema, data types and database code carry across unchanged. Create the DMS task, connect source and target, start it.
Heterogeneous migration moves data between different engines. Oracle to Amazon Aurora PostgreSQL, or SQL Server to Amazon Redshift. Here the schema and any stored code have to be converted before data moves. The Schema Conversion Tool handles most of it and produces an assessment report listing what it cannot convert automatically.
That report is the real output of the assessment phase. It tells you how much manual work sits between you and cutover. Projects that skip it are the ones whose timelines slip.
Choosing a strategy: the 7 Rs
AWS groups migration approaches into seven strategies, usually called the 7 Rs. The choice drives tooling, cost and duration more than any other decision on the project.
- Rehost. Move as it stands. Fastest route, no optimisation.
- Replatform. Move with targeted changes, such as swapping a self-managed database for Amazon RDS.
- Repurchase. Replace the workload with a different product, usually SaaS.
- Refactor. Rebuild for cloud-native architecture. Highest cost, highest long-term return.
- Retire. Switch it off. Most portfolios carry workloads nobody uses.
- Retain. Leave it in place for now.
- Relocate. Move infrastructure wholesale without changing the workload itself.
For data workloads specifically, replatform is the common landing point. Moving a self-managed Oracle instance to Amazon RDS or Aurora removes the maintenance burden without a rewrite.
How an AWS data migration runs
Six phases. The order matters more than the pace.
- Assessment. Inventory sources, map dependencies, size the data, run the Schema Conversion Tool against every heterogeneous pair. Output is a conversion report and a target architecture.
- Schema conversion. Apply the tool’s output, resolve what it flagged for manual work, test the converted schema against sample data.
- Full load. Move historical data. At large volumes this is where Snowball earns its place over network transfer.
- Change data capture. DMS replicates changes made at the source while the full load runs, so the target stays current and the source stays live.
- Validation. Row counts, checksums, referential integrity, then application-level testing against the target.
- Cutover. Point applications at the target, keep reverse replication running for a defined rollback window, then decommission the source.
Phases two and five are where projects overrun, and both are predictable from a proper assessment. Our data engineering services team runs this sequence as standard, and the wider pattern is covered in our guide to the seven steps of cloud migration.
Migrating a data lake or analytics workload
Analytics workloads follow different rules from transactional databases. The target is usually Amazon S3 with the AWS Glue Data Catalog on top, and Amazon Redshift or Amazon Athena for query.
The sequence:
- Land raw data in Amazon S3, using DataSync over the network or Snowball in bulk.
- Catalogue it with Glue crawlers so the schema is discoverable.
- Apply governance and access control through AWS Lake Formation.
- Transform with Glue jobs into the shape analytics needs.
- Connect Redshift, Athena or the reporting layer.
The mistake to avoid is treating the lake as a dumping ground. Data landed without a catalogue and without partitioning is data nobody can query economically. Partitioning strategy belongs in the data warehouse architecture design, before anything moves, not after.
A catalogued lake is also the precondition for anything built on top of it. A model retrieving against uncatalogued data inherits every gap in it, which is why AWS Bedrock development work starts at the data layer rather than the model.
What AWS data migration services cost
AWS charges for DMS only through the compute and storage its replication instances use, and Snow devices are priced per device per job. The costs that catch people out sit elsewhere.
- Egress from the source. Moving data out of another cloud or a colocation provider carries transfer charges that can exceed the migration tooling cost.
- Network capacity. A dedicated AWS Direct Connect link is often cheaper across a long migration than saturating an existing internet connection.
- Replication duration. DMS runs until cutover. A migration that stalls in validation keeps paying for replication instances.
- Parallel running. Source and target both run until the rollback window closes.
Estimate egress and replication duration first. They move the total more than instance sizing does.
Migrating regulated workloads
Healthcare, energy, financial services and public sector workloads add constraints that change the design, not just the paperwork.
- Encryption in transit and at rest, with keys held under your control through AWS KMS.
- Data residency. Some data cannot leave a jurisdiction, which constrains region choice and rules out certain transfer methods.
- Audit trail. Every access and every transformation during migration has to be logged and retained.
- Access scope. Migration engineers should not hold standing access to production data. Time-bound, least-privilege roles instead.
Liquid Technologies designs regulated migrations to meet HIPAA-aligned architecture, SOC 2 and GDPR requirements, and operates under ISO 27001. Where a workload cannot use shared inference or public cloud AI at all, the target environment is built around private AI deployment inside your own perimeter.
Consolidation, testing and continuous replication
Three uses of the same tooling that sit outside a one-time migration.
Consolidation. DMS can merge several source databases into a single target. The sources do not need to sit in one location, and the work can be homogeneous or heterogeneous.
Development and testing. Copy production data into AWS for development work, then replicate changes back to the source environment. Teams can also pull a current cloud database down into an on-premises system for local testing.
Continuous replication. DMS can run indefinitely rather than stopping at cutover. This covers disaster recovery targets, keeping a reporting replica current, and holding two systems in step. It works across different engines, so the DR target does not have to match the primary.
Choosing an AWS data migration partner
Migration is rarely a standalone project. Most of this work sits inside a wider AWS consulting services engagement covering architecture review, cost optimisation and the target design. Whoever runs it, check, in order:
- Migrations completed on your specific source and target pair, not AWS experience in general. Oracle to Aurora PostgreSQL is a different project from SQL Server to Redshift.
- A named validation method. Ask how they prove the target matches the source, and what the acceptance criteria are.
- A documented rollback plan with a defined window.
- A compliance position that matches your regulatory obligations, evidenced rather than asserted.
- Who owns the environment after cutover, and what that costs.
A partner who quotes a timeline before running an assessment has not run the assessment.
Scope your migration before you price it
The Liquid Technologies Cloud Assessment Workshop maps your data estate, identifies the right AWS services for each workload, and produces a migration plan with a validated scope.