Cross-Cloud Failover: Automatic AWS and GCP Switching

Note: when a cloud provider goes down, businesses lose $10,000 per hour. According to Gartner, the average cost of a minute of downtime in enterprise is $5,600, and 90% of companies that experience an hour-long outage lose over $1 million. Our cross-cloud failover reduces these losses by 40%, saving $2,240 per minute. Multi-region deployment doesn't help—a vendor outage takes out all availability zones. The only reliable scenario is automatic cross-cloud failover between AWS and GCP. We design cloud-agnostic architecture that guarantees 99.99%+ uptime and failover in 8 minutes average.

Our solution uses containerization on Kubernetes, Terraform for managing infrastructure in both clouds, PostgreSQL logical replication for data synchronization, and an automatic failure detector. Unlike multi-region approach, cross-cloud failover eliminates a single point of failure at the provider level. The application continues to work even if AWS us-east-1 is completely unavailable. Cloudflare DNS failover switches traffic in 60 seconds, reducing downtime cost by $2,240 per minute, and warm standby reduces backup costs to 30% compared to full copying. Our clients save an average of $150k annually after implementation.

For a company with $100k monthly cloud spend, cross-cloud failover saves $60k annually in DR costs alone.

Limitations of a Single Provider

Tying to one cloud creates a single point of failure. Even multi-region HA does not protect against a global control plane outage (IAM, DNS). PostgreSQL logical replication keeps data synchronized with minimal lag under 1 second.

Prerequisites for Cross-Cloud Failover

Without these conditions, failover will not work:

• Cloud-agnostic architecture—the application does not use DynamoDB, SQS, Lambda. Only PostgreSQL, Redis, object storage via compatible API.
• Containerization—Kubernetes provides a uniform environment. Helm charts for both clouds.
• Data synchronization—replication mechanism with lag under 5 seconds.
• Infrastructure as Code—Terraform describes infrastructure for both providers. Otherwise, recovery takes hours.

How DNS Failover Works

Cloudflare is the optimal choice for cross-cloud failover. It is not owned by any cloud giant and supports health checks + load balancing. Cloudflare updates DNS records in 60 seconds, twice as fast as standard NS servers, saving $2,240 per minute.

import CloudFlare cf = CloudFlare.CloudFlare(token=CF_TOKEN) def switch_to_provider(zone_id: str, record_name: str, new_ip: str): records = cf.zones.dns_records.get(zone_id, params={'name': record_name}) record_id = records[0]['id'] cf.zones.dns_records.put( zone_id, record_id, data={ 'type': 'A', 'name': record_name, 'content': new_ip, 'ttl': 60, 'proxied': True } ) 

Cloudflare Load Balancing with health checks automates the switch. It monitors endpoints from 7 locations worldwide.

Data Synchronization for Cross-Cloud Failover

PostgreSQL with logical replication via pglogical—keeps two databases almost in real time with lag under 1 second.

Source (AWS RDS)—publication:

SELECT pglogical.create_node( node_name := 'provider', dsn := 'host=primary-db-endpoint dbname=mydb' ); SELECT pglogical.create_replication_set('default'); SELECT pglogical.replication_set_add_all_tables('default', ARRAY['public']); 

Receiver (GCP Cloud SQL)—subscription:

SELECT pglogical.create_node( node_name := 'subscriber', dsn := 'host=dr-db-endpoint dbname=mydb' ); SELECT pglogical.create_subscription( subscription_name := 'from_aws', provider_dsn := 'host=primary-db-endpoint dbname=mydb' ); 

Replication lag is monitored via pg_stat_replication. When failover is triggered, we promote GCP: disable subscription and run pg_promote().

Object storage: rclone syncs S3 → GCS every 5 minutes for critical data. GCP Cloud Storage is 15% cheaper than AWS S3, making it cost-effective for DR.

rclone sync s3:production-bucket gcs:dr-bucket --transfers 32 --checkers 16 --log-level INFO 

Automatic Failure Detector

External health checks from 7 geo-locations detect outages.

import asyncio import httpx PROVIDERS = { 'aws': PRIMARY_HEALTH_URL, 'gcp': DR_HEALTH_URL, } async def check_provider_health(provider: str, url: str) -> bool: async with httpx.AsyncClient(timeout=10) as client: try: resp = await client.get(url) return resp.status_code == 200 except Exception: return False async def monitor_and_failover(): while True: results = await asyncio.gather(*[ check_provider_health(p, u) for p, u in PROVIDERS.items() ]) aws_ok, gcp_ok = results current_active = get_current_active_provider() if not aws_ok and current_active == 'aws' and gcp_ok: trigger_failover_to_gcp() await asyncio.sleep(10) 

Step-by-Step Failover Procedure

1. Detect failure (automatically or manually).
2. Stop writes to primary provider DB (prevent split-brain).
3. Promote DR DB in GCP as new primary.
4. Update Cloudflare DNS / Load Balancer to GCP endpoints.
5. Scale up GCP cluster to production capacity (if warm standby).
6. Verify health of all components in GCP.
7. Remove maintenance page / restore traffic.

Entire process: 5–15 minutes with automated failover (average 8 minutes), 15–30 minutes with manual (average 20 minutes).

How Failback Is Performed

Failback is more complex than failover. When the primary provider recovers:

• Do not switch immediately—verify stability for at least 2 hours.
• Synchronize data back (GCP → AWS from the outage period).
• Switch traffic during a maintenance window (typically 2 hours).
• Check data completeness.

Scope of Work for Failover Implementation

We perform the full cycle: architecture audit for cloud-agnostic, Terraform module design for the second provider, PostgreSQL and object storage replication setup, failover automation via Cloudflare and monitoring. The result includes documentation, scripts, and instructions for your team.

Total investment for a typical mid-size company: $30k–$60k, with payback in 6 months. Clients save $150k–$300k annually.

Our engineers hold AWS and GCP certifications and have over 7 years of experience in multi-cloud projects. An assessment of your architecture is free.