Railway and Render deploy to their own managed regions, so running one app on both AWS and Google Cloud means adding infrastructure outside them or moving the app. Northflank can deploy into AWS and GCP accounts you own and can geo-route one subdomain to services in different projects, but each project deploys to one cluster, and its documentation does not describe built-in cross-cloud failover or data replication. Platforms such as Control Plane run one workload across AWS, GCP, Azure, and Oracle Cloud locations behind a single endpoint.
Disclosure: Control Plane publishes this guide.
Platform details were checked against each vendor’s documentation in November 2026.
TL;DR
Railway, Render, and Northflank each run a full application stack, but none of their documentation describes built-in automatic failover for one app running across AWS and Google Cloud.
- Railway supports multi-region replicas for stateless services across its four Railway Metal regions, with public traffic routed to the nearest region. Railway’s documentation states it does not offer a bring-your-own-cloud option.
- Render runs web services, workers, and datastores in five Render-managed regions (Oregon, Ohio, Virginia, Frankfurt, and Singapore); static sites are served from a global CDN.
- Northflank supports bring-your-own-cloud on AWS, GCP, Azure, and other providers. A project’s region cannot be changed after creation, and all resources in that project deploy in that region; a geo-routed subdomain can send users to the closest of several projects.
- Building it yourself means assembling compute in each cloud, container registries, health-checked traffic routing, database replication, object storage, secrets management, and unified observability across both clouds.
- Control Plane runs one workload across AWS, GCP, Azure, and Oracle Cloud locations behind a single canonical endpoint that DNS geo-routes to the nearest healthy location; you can run the same app on AWS, GCP, and Azure without building the routing layer yourself.
Can Railway, Render, or Northflank run one app on both AWS and Google Cloud?
Each platform stops at a different layer.
Railway runs on its own Railway Metal infrastructure across four regions. You can create multi-region replicas of a stateless service, and Railway will route public traffic to the nearest region; volumes follow the region of the service they are attached to. Services in the same Railway environment connect over Railway’s private network even when they run in different regions. Railway’s own comparison page states it does not offer a bring-your-own-cloud option, so you cannot use Railway to run the same workload inside your own AWS or GCP account.
Render runs web services, workers, and datastores in five Render-managed regions. Render does not currently support changing the region of an existing service or database, and its private network is limited to one region and workspace. Render’s PostgreSQL high availability standby runs in the same region as the primary, in a separate zone. Render’s documentation does not describe an option to deploy into your own AWS or GCP account or across multiple cloud providers; its private links connect Render services outbound to resources in an AWS VPC in the same region.
Northflank provides bring-your-own-cloud deployment into AWS, GCP, Azure, and other providers, so your application runs inside your cloud accounts. Each Northflank project is pinned to the region chosen at creation, and every resource in it deploys there. Multi-project networking allows ingress only between projects in the same team and on the same cluster. Database addons deploy to the project’s region, and cross-region restoration from global backups is an Enterprise feature. To run across both AWS and GCP, you would create a project on a cluster in each cloud. A geo-routed subdomain, which requires Northflank CDN, can send users to the closest backend, but Northflank’s documentation does not describe health-based failover or data replication across the clouds, and private networking between projects on different clusters runs through a Tailscale network you connect, so you build and operate those yourself.
What does each platform document today?
Each platform’s documentation sets these limits on running one app on AWS and Google Cloud.
| Capability | Railway | Render | Northflank | Control Plane |
|---|---|---|---|---|
| Deploy targets | Four Railway Metal regions on Railway-owned infrastructure | Five Render-managed regions | Northflank’s managed cloud, or your own AWS, GCP, Azure, or other cloud accounts | AWS, GCP, Azure, and Oracle Cloud regions, plus your own Kubernetes clusters as BYOK locations |
| Run in your own AWS or GCP account | Not available, per Railway’s docs | Not described in Render’s docs | Yes, via BYOC or an imported cluster (BYOK) | Yes, by attaching your own Kubernetes clusters as locations (BYOK), or with Control Plane’s Managed Kubernetes (MK8s), which runs worker nodes in your AWS account or GCP project |
| One app in two clouds at once | Multi-region replicas across Railway regions only | No; separate services per Render region only | Separate projects per cloud; a geo-routed subdomain can route users across them | Yes, one workload runs in every location of its Control Plane Global Virtual Cloud (GVC) |
| One public endpoint routing to the nearest healthy copy | Nearest-region routing for multi-region replicas; health checks run at deploy time only | Not applicable; each service runs in one region | Geo-routed subdomain to the closest or per-continent backend; requires Northflank CDN | Yes, canonical endpoint with DNS geo-routing to the nearest location passing readiness probes |
| Private networking between two copies | Railway’s private network per environment, including across Railway regions | Private network limited to one region and workspace | Ingress between projects in the same team and on the same cluster; across clusters, through a Tailscale network you connect | Internal endpoint (workload.gvc.cpln.local) over mutual TLS on Control Plane’s service mesh, once the internal firewall allows the caller; calls are served in the caller’s location and, by default on new workloads, fail over to other locations when no local replica is ready |
| Database replicated across clouds | No cross-region replication documented; a database with a volume is locked to one region; PostgreSQL can run as a high-availability cluster with automatic failover | No cross-region replica documented; HA standby in the same region, separate zone; logical replication to a self-managed external database on Pro plans and higher | Addons deploy to the project’s region; no cross-region replication documented; cross-region restore through global backups is an Enterprise feature | Template Catalog includes replicated databases such as Postgres Multi-Location and pgEdge |
| Automatic failover between clouds | Not applicable; Railway runs only on Railway-owned infrastructure. Failed replicas restart automatically | No; HA failover is within one region | Not described in Northflank’s docs; geo-routing selects a backend by location | When a workload runs in two or more locations and one fails, users are routed to healthy locations |
Sources: platform documentation linked in the table and in the platform sections. Cells without a link summarize the sources linked elsewhere in the same row or in that platform’s section; “not described” means the capability does not appear in that platform’s documentation as of November 2026.
What do you have to build yourself to run across two clouds?
Running one app on AWS and Google Cloud yourself means building and maintaining seven layers and managing one new cost line.
| Layer | What it takes | Common tools | Watch out for |
|---|---|---|---|
| Compute in each cloud | A container runtime or Kubernetes cluster in both AWS and GCP | Amazon EKS, Google GKE, or managed container services | Keeping cluster versions, node pools, and autoscaling policies in sync |
| Container images and registries | A single image pushed to registries accessible from both clouds | Amazon ECR, Google Artifact Registry, or a third-party registry | Image tag consistency, pull latency across regions, and pull credentials: pulling from Amazon ECR on GKE needs an auth token that expires after 12 hours |
| Traffic routing and health checks | Latency-based or geolocation DNS routing with health checks, or a global load balancer, that routes users to the nearest healthy instance | Amazon Route 53, Cloudflare Load Balancing, Google Cloud DNS | Resolvers and clients that cache records past the TTL; detection takes roughly the check interval times the failure threshold, plus probe timeouts, and users move only after DNS propagates and cached records expire |
| Database replication | PostgreSQL logical or physical replication across clouds, or a distributed SQL database; if the primary is Amazon RDS, replicate out with logical replication, since physical streaming replication needs a self-managed primary | PostgreSQL streaming or logical replication over VPN or TLS, CockroachDB, pgEdge | Replication lag from cross-cloud latency; split-brain if both sides accept writes after a network partition |
| Object storage | Uploads and assets stored in cloud-native object storage, replicated or dual-written | Amazon S3, Google Cloud Storage, cross-cloud sync tools | Consistency windows during replication; egress costs for cross-cloud sync |
| Secrets and cloud credentials per provider | Separate IAM roles, service accounts, and secret stores for each cloud | AWS IAM, GCP IAM, HashiCorp Vault, cloud-native secret managers | Credential rotation schedules; avoiding long-lived keys in application code |
| Observability across both clouds | Unified logging, metrics, and tracing from two different cloud environments | OpenTelemetry for instrumentation, sending to a backend such as Datadog or Grafana Cloud | Correlating traces across cloud boundaries; alert fatigue from duplicated monitors |
| Cost of cross-cloud data transfer | Egress fees for traffic between AWS and GCP, including replication traffic | Cloud provider billing dashboards, cost allocation tags | Over the internet or a VPN, replication and sync traffic is billed at each cloud’s internet data transfer rates and grows with write volume; a dedicated link such as Google Cross-Cloud Interconnect changes the pricing |
How do you add Google Cloud next to AWS yourself, step by step?
These steps add Google Cloud as a second location for an app already running on AWS.
| Step | What you do | Done when |
|---|---|---|
| 1. Make the app stateless | Move uploads off local disk to object storage such as Amazon S3 or Google Cloud Storage. Move session state to a shared cache or database. | The app starts cleanly with no local state and serves requests identically from any instance. |
| 2. Containerize consistently | Ensure the app builds as a single container image, built for the CPU architecture both clouds run or as a multi-arch image, with no cloud-specific build steps baked in. | One image runs successfully on both AWS and GCP container runtimes. |
| 3. Provision compute on Google Cloud | Create a GKE cluster or Cloud Run service in a GCP region close to your users. Deploy the same container image. | The app starts and passes health checks on GCP. |
| 4. Set up traffic routing | Configure a global load balancer, or latency-based or geolocation DNS records with health checks, that routes users to the nearest healthy endpoint. Until the database replica in step 5 exists, requests served from GCP reach the AWS database across clouds. | Requests from different geographies resolve to the closer cloud. |
| 5. Replicate the database | Set up PostgreSQL streaming replication to a same-version replica over a site-to-site VPN or TLS (if your primary is Amazon RDS, use logical replication instead), or deploy a distributed SQL database with nodes in both clouds. Decide on single-primary or multi-primary. | Writes on the primary are visible on the replica within your acceptable lag window. |
| 6. Unify secrets and credentials | On GKE, use Workload Identity Federation for GKE; on Cloud Run, attach a service account as the service identity. For calls back into AWS services such as S3, create an AWS IAM role that trusts Google as an OIDC identity provider so the GCP copy gets temporary AWS credentials. Replicate application secrets to Google Secret Manager. | The app on GCP accesses all required services and secrets without hardcoded credentials. |
| 7. Connect observability | Send logs, metrics, and traces from both clouds to a single observability platform. Set up alerts for replication lag, error rates, and health check failures. | A single dashboard shows the health of both deployments side by side. |
| 8. Run a failover drill | Simulate a failure in the AWS deployment, fence the AWS primary, promote the GCP replica, confirm traffic shifts to GCP, and measure the error spike and recovery time. | Traffic shifts to GCP within your target recovery time, and the error rate returns to baseline. |
How do you keep the database consistent across two clouds?
Database consistency is usually the hardest part of a multi-cloud setup. Two main approaches exist, each with different trade-offs.
A single-primary architecture keeps one PostgreSQL instance as the writable primary, typically in AWS, with a read replica in Google Cloud using streaming replication. Writes always go to the primary. If the primary fails, you promote the replica. This model is simpler to reason about because there is one source of truth at any moment. The trade-off is that each write query from the app running in GCP crosses clouds, so a request that runs several queries pays several round trips, and writes are unavailable from the moment the primary fails until the replica is promoted and the app reconnects.
A multi-primary architecture runs writable database nodes in both clouds. Distributed SQL databases like CockroachDB, or PostgreSQL-based solutions like pgEdge, can handle this. pgEdge accepts writes on each node and replicates them to the others, which keeps write latency low near either cloud but requires resolving conflicting writes. Two users updating the same row at the same instant from different clouds can produce conflicts that the database must resolve automatically or surface to the application. CockroachDB needs at least three regions to survive a full region failure, so a two-cloud setup needs a third location.
Recovery point objectives matter. With asynchronous streaming replication, you lose any committed writes the replica had not received when the primary failed, which can be several seconds of writes, and more when replication lags. Synchronous replication avoids losing committed writes but adds a cross-cloud round trip to every commit, and writes stall if the replica is unreachable. Multi-primary systems differ: CockroachDB commits a write only after a majority of its voting replicas acknowledge it, while pgEdge resolves conflicting rows when replicated changes arrive.
Failover order should be documented and drilled. Fence the old primary, promote the replica, then shift traffic. If traffic shifts before the database is promoted, write requests fail, so make the GCP health check report unhealthy until its database accepts writes. Automate the promotion only if you have tested it and your recovery time objective requires it.
When should you stay on your current platform?
If your customers do not require cross-cloud resilience in their contracts, and your application can tolerate the recovery time of a single-region restore from backups, same-region high availability may be sufficient. Each platform covers part of this today. Railway offers multi-region replicas for stateless services and a PostgreSQL high-availability cluster with automatic failover, Render offers a Postgres high-availability standby in a separate zone of the same region, with automatic failover to it, and Northflank runs projects in its managed cloud or in your own AWS, GCP, or Azure account, with geo-routing across them.
Signals that it is time to add a second cloud include a region outage that would take your whole app down, a customer contract or SLA that requires multi-cloud, credits from a second provider, or a data residency rule tied to a specific cloud.
Moving to multi-cloud increases operational cost. You will operate two sets of infrastructure, two sets of credentials, and a replication pipeline. If the business case does not justify that cost today, invest in solid backups, health checks, and a documented recovery plan on your current platform.
How Control Plane fits
Control Plane handles Day 2 operations for you, including autoscaling, built-in observability, and security. A Global Virtual Cloud (GVC) is Control Plane’s named set of locations, which can include AWS, GCP, Azure, and Oracle Cloud regions and your own Kubernetes clusters. A workload runs in every location of its GVC, and by default every location serves traffic. Each workload that serves traffic gets a canonical endpoint that DNS geo-routes to the nearest location where the workload passes its readiness probes, so when a workload runs in two or more locations and one fails, users are routed to the healthy ones.
During the October 20, 2025 AWS us-east-1 outage, 20% of Control Plane customers had AWS us-east-1 in their GVC. Their workloads were active and running in healthy locations within 10 seconds, with no customer downtime, according to Control Plane’s internal incident telemetry.
Control Plane’s Template Catalog includes replicated databases such as Postgres Multi-Location and pgEdge. Its Universal Cloud Identity lets workloads call AWS, GCP, and Azure services with no long-lived cloud keys in your code. Control Plane is PCI DSS Level 1, SOC 2 Type II, HIPAA, and GDPR compliant.
Control Plane runs container images, so an app that already builds as a container image for Railway, Render, or Northflank can be deployed as a Control Plane workload. Control Plane runs and routes the workloads; your application must be stateless or use a replicated data store, and you choose the database replication model. Learn more about adding a second cloud provider for failover without doubling ops or running the same app on AWS, GCP, and Azure.
FAQ
These answers cover running one app across AWS and Google Cloud on each platform.
Can Railway run on AWS and Google Cloud?
Railway runs on its own Railway Metal infrastructure across four regions, and its documentation states it does not offer a bring-your-own-cloud option, so all replicas run on Railway-owned infrastructure. To run on both AWS and Google Cloud, you would deploy and manage infrastructure outside Railway.
Does Render support multi-cloud deployment?
No. Render web services, workers, and datastores run in five Render-managed regions, and Render does not currently support changing the region of an existing service or database. Its documentation does not describe an option to deploy into your own AWS or GCP account. Render’s Postgres standby stays in the primary’s region.
Can Northflank run one app across AWS and GCP?
Northflank supports bring-your-own-cloud deployment into AWS, GCP, Azure, and other providers, and each project deploys to one cluster and region, so running across both clouds takes a project in each. A geo-routed subdomain can route users to the closest one, but health-based failover and cross-cloud data replication are not described as built-in features.
How do I fail over between AWS and Google Cloud automatically?
Automatic failover between AWS and Google Cloud requires a health-checked traffic layer in front of both clouds, such as Amazon Route 53 records with health checks or Cloudflare Load Balancing. The health check must detect when your application in one cloud is unhealthy and send traffic to the other cloud. Database failover must be coordinated separately, either through automated fencing and replica promotion or a distributed SQL database that handles node failure internally.
Do I need Kubernetes to run on two clouds?
Kubernetes is one way to standardize your container runtime across AWS and GCP, using Amazon EKS and Google GKE. Managed container services such as Amazon ECS with AWS Fargate and Google Cloud Run also work, as does a platform that abstracts the infrastructure. The key requirement is that both clouds run the same container image and that you have unified traffic routing and health checks in front of them.
How do I replicate Postgres between AWS and Google Cloud?
If you run PostgreSQL yourself, streaming replication can replicate data from a primary on AWS to a same-version replica on Google Cloud over an encrypted connection. If the primary is Amazon RDS, use logical replication instead. You configure the replica to connect to the primary’s VPN-connected endpoint or an IP-allowlisted TLS endpoint. Logical replication copies selected tables and works across major Postgres versions, but it does not replicate schema changes or keep sequence values continuously in sync, so sync both before promoting a logical replica. For multi-primary writes, consider a distributed SQL database like CockroachDB, which keeps replicas consistent through consensus, or a PostgreSQL-based solution like pgEdge, which resolves write conflicts between locations.
How does Control Plane run one app on AWS and Google Cloud?
Control Plane uses a Global Virtual Cloud (GVC) that spans locations across AWS, GCP, Azure, Oracle Cloud, and your own Kubernetes clusters. A workload runs in every GVC location behind one canonical endpoint that routes users to the nearest healthy location, so when it runs in two or more locations and one fails, users reach the healthy ones. Read about how to route users to the nearest healthy region and test failover without taking production down.
What is BYOC and which platforms support it?
BYOC stands for bring your own cloud, a deployment model where the platform manages your application but runs it inside your own cloud account. Among the three platforms covered here, Northflank supports BYOC on AWS, GCP, Azure, and other providers. Railway’s documentation states it does not offer BYOC, and Render’s documentation does not describe a BYOC option. Explore BYOC PaaS options for running apps in your own AWS or GCP account.
Is multi-cloud worth the complexity for a small SaaS?
Usually not, unless a contract, regulator, or region-outage risk justifies running two sets of infrastructure. Otherwise single-cloud high availability with tested backups is often enough.
Talk to the Control Plane team to see how a Global Virtual Cloud works for your app, or sign up free to deploy your first workload across clouds.
Originally published on the Control Plane blog.


![Mobile Manipulation Software and Operations — Ch.13: OTA Updates [EN]](https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fb0t5n8yyl10kkhac8kc0.png)
![Mobile Manipulation Software and Operations — Ch.12: Safety Validation [EN]](https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fmync7cwwizyn31zfxaud.png)