AWS Builder Center
Active-active vs active-passive: the multi-Region decision matrix: Resilient Cloud Networking on AWS — Patterns from the Field, Article 8

Active-active vs active-passive: the multi-Region decision matrix: Resilient Cloud Networking on AWS — Patterns from the Field, Article 8

Active-active vs active-passive across AWS Regions, in honest terms. Where the boundary actually sits, and how to choose without flattering the architecture.

A European payments firm I worked with had spent two years convincing itself it was running "multi-Region active-active." The architecture diagram showed two AWS Regions in EMEA, an Amazon Aurora Global Database stretched across them, Amazon DynamoDB global tables underneath the loyalty service, and Amazon Route 53 latency-based routing in front of it all. The team's internal documentation, their board reports, and their Digital Operational Resilience Act (DORA) self-attestation all used the words "active-active." Then a real Regional impairment arrived — precisely the kind of event a multi-Region architecture exists to absorb, and the scenario AWS's resilience guidance tells you to plan for as a design assumption rather than treat as a surprise. Reads from the secondary Region kept working. Writes did not. The Aurora cluster was configured with write forwarding from the secondary to the primary, which meant every write was still flowing back to the impaired Region. The team had to invoke a managed failover, promote the secondary cluster, repoint application configuration, and only then resume writes. The recovery time was forty-seven minutes. The board had been told the recovery time objective (RTO) was zero.
That afternoon is the kind of incident that exposes a vocabulary problem. The AWS Disaster Recovery whitepaper describes four broad strategies — Backup and Restore, Pilot Light, Warm Standby, and Multi-Site Active/Active[^dr] — and within active/active it draws a further distinction between hot standby (an active/passive deployment with two fully provisioned Regions where only one serves traffic) and a true multi-Region active/active deployment where users can be served from any participating Region. The boundary between those models is where most resilience strategies fail under examination. Choosing between them is the decision the regulator will actually want documented, and it is the one most teams have not made on paper.

What "active-active" actually requires

The strategies differ most concretely in the resilience outcomes they commit to. Pilot Light keeps data replicated and core infrastructure provisioned but switched off; AWS characterises its recovery point objective (RPO) as minutes and its RTO as tens of minutes. Warm Standby runs a scaled-down but functional copy in the recovery Region, which compresses the RPO into seconds and the RTO into minutes.[^rel13] Multi-Site Active/Active is the only model that targets an RPO near zero and an RTO potentially zero, and only because the workload is actively serving traffic from both Regions all of the time.
That last clause is the one that gets glossed over. Active/active in the AWS sense is not just two Regions provisioned. It is two Regions both processing user traffic, both holding state, and both designed for the operational reality that conflicting writes can happen and must be either avoided by design or reconciled deterministically. The whitepaper itself is explicit that active/active is the most complex and most costly approach, and that designing how concurrent writes are handled is the central problem the architect must solve.
The decision narrows down to three write postures. A write-global posture routes all writes to a single Region and treats other Regions as read replicas; Amazon Aurora Global Database, with up to 10 secondary Regions and typical replication latency under a second, is built for this.[^aurora] A write-local posture lets every Region accept reads and writes, with conflict handling delegated to the data store itself; Amazon DynamoDB global tables are the canonical example, and it is worth being current about what they actually offer. In their default mode — multi-Region eventual consistency (MREC) — writes replicate asynchronously (typically within a second) and simultaneous conflicting writes are reconciled by last-writer-wins on a per-item basis, which yields a non-zero RPO and is acceptable for some workloads and intolerable for others. Since 2024, global tables also offer multi-Region strong consistency (MRSC), which replicates synchronously across exactly three Regions, serves strongly consistent reads everywhere, and rejects a genuinely conflicting concurrent write with a ReplicatedWriteConflictException rather than silently resolving it — delivering an RPO of zero at the cost of higher write latency and a fixed three-Region footprint.[^ddb] The honest version of a write-local design names which mode it depends on: last-writer-wins is a property of the default, not an unavoidable law of the data store. A write-partitioned posture deterministically assigns each writer to a specific Region by a partition key — typically a user identifier — so that write conflicts cannot happen across Regions because the same key never gets written from two places at once. Each posture has its place. The honest version of an active/active design names which one it is and what failure modes it accepts.

Hot standby is more honest than most teams admit

The payments firm in the opening anecdote did not actually have active/active. It had hot standby. Two fully provisioned Regions, a single Region accepting writes, automatic read serving from both, and a deliberate decision — formalised through Aurora's write-forwarding feature[^wf] — that writes would be funnelled to one place to avoid the conflict problem. That is a legitimate architecture. It is, in fact, the right architecture for many regulated workloads where the data model cannot tolerate last-writer-wins reconciliation and partitioning is not workable. What it is not is RTO-zero. Aurora's managed planned switchover and unplanned failover both involve promoting a secondary Region, which AWS documents as completing in under a minute under normal conditions, with an RPO measured in seconds depending on replication lag at the moment of impairment.
Hot standby earns its keep when the workload's resilience outcome is "survive a Regional event with bounded data loss and bounded recovery time," when the team is willing to operate two production Regions full-time, and when the cost of building real multi-master semantics is greater than the cost of a one-minute promotion. Many of the deployments that fly the active/active flag would be more accurately and more defensibly described this way.
This is not a contrarian position, and it is worth being precise that it is now AWS's own. The AWS multi-Region fundamentals guidance published in September 2025 states plainly that, of the multi-Region continuity patterns, "the pattern with the lowest risk of meeting recovery objectives will involve running hot standby"[^mrf] — while acknowledging it will "double the cost for your workload." The older Disaster Recovery whitepaper still carries the softer instinct that "most customers find that if they are going to stand up a full environment in the second Region, it makes sense to use it active/active," and that instinct is exactly the one the payments firm followed into a forty-seven-minute recovery. Standing up the second Region is the expensive part. Being disciplined about whether it writes is the part that actually protects you.

Steering the traffic without becoming the failure

Whichever posture is chosen, the steering layer that decides which Region a user lands on is itself part of the failure model. Amazon Route 53 routing policies — weighted, latency, geolocation — let traffic flow to whichever Region is the right one under normal conditions. For an active/passive failover, the relevant tool is Amazon Application Recovery Controller (ARC). ARC routing controls are simple on/off switches whose state is propagated across a five-Region cluster and queried through a highly available data plane;[^arc] flipping a routing control is a data plane operation — and AWS's own guidance is to rely only on data plane operations during a failover, precisely because data planes carry higher availability design goals than control planes. AWS Global Accelerator offers traffic dials that perform an equivalent function for steering across endpoints, with one caveat the documentation calls out plainly: adjusting a traffic dial is a control plane operation, and relying on it as a data plane recovery primitive is not the resilient choice.[^ga]
Routing controls are the low-level primitive, but a real failover is rarely a single switch — it is a sequence: promote the database, scale the standby, flip the traffic, run post-cutover checks. Since 2025 AWS has offered a managed orchestration layer for exactly that, ARC Region switch, and it has matured quickly through late 2025 and into 2026 with post-recovery workflows, native Amazon RDS execution blocks, and Terraform support. It is built around a plan of workflows and steps,[^rs] it spans both postures in this article — failover and failback for active/passive, shift-away and return for active/active — and, most importantly for the argument here, it runs on "a data plane in each AWS Region, so that you can execute your Region switch plan without taking a dependency on the Region that you’re deactivating." That last property is the whole point: the recovery mechanism must not share a fate with the thing it is recovering from.
The discipline is to identify, for every active/passive design, the single mechanism that has to work during the worst hour and to make sure that mechanism lives in the data plane. ARC routing controls — and, for multi-step recoveries, ARC Region switch plans — were built for exactly that role. They are the recommended steering primitives for multi-Region failover precisely because their failure model is decoupled from the failure model they exist to handle.

The decision matrix

A decision matrix worth living with begins with a deliberately deflating premise, and AWS's 2025 multi-Region fundamentals guidance opens with exactly it: "most availability and continuity of operations goals can be met within a single Region." Multi-Region is for the small subset of workloads whose recovery time genuinely has to be measured in minutes rather than hours; for everything else, waiting for an in-Region remediation is the more honest — and far cheaper — commitment. Assuming the workload has cleared that bar, the matrix begins with the resilience outcome it is required to deliver. For workloads under DORA scope, that outcome is increasingly expressed in regulator-readable RTO and RPO targets and in the demand for evidence that the recovery can be exercised. The matrix then narrows on three orthogonal axes. The first is write semantics: can the data model tolerate last-writer-wins, can it be partitioned cleanly, or must writes funnel to a single primary at any given time? The second is operational maturity: can the team genuinely run two production Regions full-time, with on-call coverage, deployment pipelines, observability, and Game Day rehearsals reaching both, or is a scaled-down passive Region a more honest commitment? The third is cost: a hot standby roughly doubles steady-state cost, an active/active deployment doubles it and adds the engineering tax of conflict handling, and a warm standby costs meaningfully less while accepting a longer RTO.
What collapses out of those three axes is rarely the answer everyone wants. The payments firm's regulators did not need active/active. They needed evidence that the Region failover was exercised, that the steering control sat in the data plane, that the data loss boundary was understood, and that the recovery time was honestly documented. Hot standby, properly steered with ARC routing controls and properly rehearsed, would have given them all of that. What they had given themselves was a label that fell apart on first contact with reality.


[^dr]: AWS,  Disaster Recovery of Workloads on AWS: Recovery in the Cloud  — the four strategies, the hot-standby-vs-multi-site-active/active distinction, the "most complex and costly / near zero" characterisation, and the guidance to use only data plane operations during failover.
[^rel13]: AWS Well-Architected Framework, Reliability Pillar,  REL13-BP02 Use defined recovery strategies to meet the recovery objectives : "Pilot light (RPO in minutes, RTO in tens of minutes)… Warm standby (RPO in seconds, RTO in minutes)."
[^aurora]: AWS,  Using Amazon Aurora Global Database  and the  Aurora Global Database product page : up to 10 read-only secondary Regions, "typical latency of less than 1 second," and promotion of a secondary "in less than 1 minute" with an effective RPO of ~1 second.
[^ddb]: AWS,  How DynamoDB global tables work : multi-Region eventual consistency (MREC, the default) reconciles conflicts by "last writer wins" on a per-item basis with a non-zero RPO; multi-Region strong consistency (MRSC) replicates synchronously across exactly three Regions, supports an RPO of zero, and returns a  ReplicatedWriteConflictException  on a conflicting concurrent write.
[^wf]: AWS,  Using write forwarding in an Amazon Aurora global database : "secondary clusters… forward SQL statements that perform write operations to the primary cluster… the primary cluster is the source of truth."
[^mrf]: AWS Prescriptive Guidance,  AWS multi-Region fundamentals  (September 2025): "the pattern with the lowest risk of meeting recovery objectives will involve running hot standby, and will double the cost for your workload," and "most availability and continuity of operations goals can be met within a single Region."
[^arc]: AWS,  Amazon Application Recovery Controller (ARC) — routing controls : routing controls are "simple on/off switches" hosted on a cluster of "five redundant Regional endpoints." The whitepaper adds that toggling a Route 53 weighted policy is a control plane operation and "therefore not as resilient as the data plane approach using Amazon Application Recovery Controller (ARC)."
[^ga]: AWS Disaster Recovery whitepaper: "For manually initiated failover, you can adjust which endpoint receives traffic using traffic dials, but note this is a control plane operation."
[^rs]: AWS,  Region switch in ARC : plans of workflows and steps, support for "active/passive and active/active configurations" (failover/failback vs shift-away/return), and "a data plane in each AWS Region, so that you can execute your Region switch plan without taking a dependency on the Region that you're deactivating." Capability additions per the  February 2026 what's-new .
[^rc]: AWS,  ARC Developer Guide document history : the ARC readiness check feature closes to new customers on 30 April 2026; existing customers can continue to use it.


Cristian Critelli is the EMEA Lead for Networking and Resilience Specialist Partner Solutions Architect at AWS focused on networking and resilience across EMEA. His book, Cloud Networking and Resilience (Apress), is out 2026.
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article