DynamoDB Global Tables: Multi-Region Replication Explained
A global table is a single DynamoDB table replicated across multiple AWS Regions, where every replica is writable.
DynamoDB keeps the replicas in sync, so you get low-latency local reads and writes in each Region plus cross-Region disaster recovery without running your own replication.
The examples use an append-only audit log for a multi-tenant SaaS: one item per
event, partitioned by tenant (PK = TENANT#acme, SK = EVENT#<time>#<id>), as
set up in the Operations & Cost overview.
Most tenants run in us-east-1, a group of EU customers runs in eu-west-1,
and as a compliance record the log has to survive a full Regional outage.
How do DynamoDB Global Tables work?
- Multi-Region, active-active. Each replica is readable and writable; a write in any Region propagates to every other replica.
- Every replica holds every item. A global table copies all of its data to all of its Regions. It is a resilience and latency feature, not a way to keep some tenants' data in one Region.
- In the default mode (MREC), replication is asynchronous and eventually consistent across Regions, typically within a second. Concurrent writes to the same item resolve last-writer-wins.
- It backs the 99.999% availability SLA, against 99.99% for a single-Region table.
The problem: one Region isn't enough
If us-east-1 has an outage, a single-Region audit log is unreadable and
unwritable for the duration, which is exactly when you most need the record of
what happened.
Building cross-Region replication, failover and conflict handling yourself is a large, error-prone project. Global tables make it a configuration choice.
Data residency is a different requirement. If the EU customers' events must
be stored only in the EU, a global table spanning us-east-1 and eu-west-1
breaks it, because their events replicate to the US too.
Keep those tenants in a separate table whose Regions are all in the EU (for
example eu-west-1 plus eu-central-1 for DR), and route each tenant's
requests to its table.
Keep that table's backup copies, S3 exports and stream consumers in EU Regions too: an AWS Backup cross-Region copy or a US-hosted consumer moves the data out just as a replica would.
Replication mechanics
You add a replica Region to an existing table, and DynamoDB copies the data there and keeps the replicas in sync. On a provisioned table, turn on write auto scaling first (or switch to on-demand), or DynamoDB refuses the replica:
aws dynamodb update-table \
--table-name audit-log \
--replica-updates '[{"Create": {"RegionName": "eu-west-1"}}]'Watch the ReplicationLatency CloudWatch metric (per Region pair) to see how far
behind a replica runs.
In the default MREC mode:
- A read in the other Region may not see a write yet. A write in
us-east-1is acknowledged locally, then propagated toeu-west-1. Strongly consistent reads return the latest version only for items last written in the Region you read from. - Conflicts are last-writer-wins, per item. If the same item is written in two Regions at nearly the same time, DynamoDB keeps the write with the latest timestamp and discards the other. AWS doesn't record these conflicts in CloudWatch or CloudTrail.
- Transactions are atomic only in the Region that ran them. Another Region
can briefly see part of a
TransactWriteItemswhile it replicates. - Recovery point = replication delay. AWS gives MREC an RPO of "usually a few seconds"; MRSC's RPO is zero.
A worked example: an EU replica for latency and DR
You add eu-west-1 as a replica of the audit-log table. Now:
| write region | item | visible in | |
|---|---|---|---|
| us-east-1 | TENANT#acme | EVENT#…#a1 | both Regions (~1s lag to EU) |
| eu-west-1 | TENANT#initech | EVENT#…#e7 | both Regions (~1s lag to US) |
EU-based tenants write to and read from the local eu-west-1 replica at low
latency. The same replica is your disaster recovery: if us-east-1 goes down,
eu-west-1 holds everything replicated before the outage (typically all but the
last few seconds of writes) and serves traffic once you fail over.
The audit log is append-only and partitioned per tenant, and each tenant's events are written from one Region with unique keys.
Two Regions rarely race on the same item, so last-writer-wins costs this table nothing. That is why an append-only log fits global tables well.
Do it in DynoTable
After adding a replica you want to confirm it's healthy and that the data landed:
that the EU replica holds acme's events with the right attributes.
In DynoTable, the Replicas section of the Table settings dialog (the Settings button in a table's toolbar) lists each replica Region with its live status, and adds or removes replicas on an MREC table.
A connection profile targets one Region, so to read the eu-west-1 copy you
open the table from a profile set to eu-west-1 and query the same tenant
there.

To keep the replica list in code rather than in a console click, the
infrastructure code generator
turns aws dynamodb describe-table output into Terraform, AWS CDK or
CloudFormation, replicas included.
Pitfalls and next steps
- Don't read-your-own-write across Regions. Route each tenant's reads to the Region that takes its writes, and fail a tenant over to another Region as a whole. AWS's guidance is that an application homed in one Region calls only that Region's endpoint.
- Last-writer-wins silently drops data. For mutable items, give each item (or tenant) a home Region and send all its writes there, so concurrent writes never come from two Regions. If the table can start empty and needs no TTL, LSIs or transactions, MRSC rejects the conflicting write instead.
- Every replica costs. Each Region stores a full copy and pays for every replicated write, so one extra Region roughly doubles the write and storage bill (AWS doesn't charge for the replication data transfer). Add Regions for a DR or latency need, not by default.
- Backups are per replica. PITR settings are never synchronized between replicas, so turn PITR on in every Region you might restore from. A restore comes back as an independent single-Region table; add replicas to it again before you repoint the app. See backup & point-in-time recovery.
Global tables protect against losing a Region. Protecting against losing data, a bad deploy or accidental delete, is the job of backup & point-in-time recovery.
Download DynoTable to check your replicas' status and query each Region's copy of the table.


