Scaling a Neo-Bank: Moving from AWS to Hybrid Bare Metal

The Challenge: "The noisy neighbor problem"

The client was hosted entirely on AWS (RDS for PostgreSQL and EKS for compute). As their user base grew, they began experiencing "micro-stalls" in their database performance.

Financial transactions require atomic consistency (ACID). You cannot afford a 500ms delay when a user is tapping their card at a terminal. The root cause was identified as I/O contention. Despite paying for "Provisioned IOPS," the shared nature of public cloud storage meant that during peak London trading hours, their disk latency fluctuated unpredictably.

The Solution: Hybrid Bare Metal

Migrating entirely away from the cloud is difficult due to the convenience of managed services. We proposed a Hybrid Architecture.

1. Frontend & Stateless Apps: Remained on Kubernetes (migrated to Forester Managed K8s) for easy auto-scaling.

2. The Core Ledger (Database): Moved to dedicated Bare Metal servers. We deployed a 3-node PostgreSQL cluster on AMD EPYC hardware with local NVMe drives (RAID 10).

Technical Implementation Details

Local NVMe vs Network Storage: Public clouds attach storage over the network (EBS). By moving to Bare Metal, we eliminated the network hop entirely. The database now talks directly to the PCIe bus of the NVMe drives.

• CPU Pinning: We isolated the database cores to prevent context switching.
• Private Networking: We established a 40 Gbps private VLAN between the Kubernetes nodes and the Database nodes, ensuring that financial data never traverses the public internet.

The Results

The migration took 6 weeks, with zero downtime during the final switchover (using logical replication).

Performance: The P99 latency dropped from 450ms to 45ms. The "micro-stalls" disappeared completely because the disk I/O was no longer shared with other tenants.

Cost: By eliminating the "Provisioned IOPS" fees and data egress charges, the monthly infrastructure spend dropped by 40%, despite having 2x the raw compute power available.