Terraform, Kubernetes tradeoffs, and GitOps at prop-firm scale
Production infrastructure for trading systems: AWS instance selection, Terraform at multi-account scale, why matching engines do not belong in Kubernetes, blue-green deployments for stateful OMS, and cost engineering.
The infra decision that seems fine at $100K monthly volume creates a $200K incident at $10M volume. These eight posts document the tradeoffs - from c6i.metal vs c6i.4xlarge latency numbers to the ArgoCD DOKS gotchas - that only surface when you have run these systems in production.
How ZeroCopy chose c6i.metal with cluster placement groups and ENA Express to hit sub-100µs P99. A decision matrix for latency-critical AWS trading infrastructure.
How ZeroCopy manages 4 AWS accounts and DigitalOcean with isolated Terraform state, Atlantis PR workflows, and module patterns that prevent production drift.
A contrarian practitioner take on K8s for trading: why cgroup scheduling, iptables overhead, and container startup times disqualify it for the hot path - and what thrives on K8s instead.
How Upside's OMS canary release split position state across two versions for 20 minutes. The deployment patterns that actually work for stateful trading services.
Why active-active is almost never correct for trading. ZeroCopy's active-passive failover design, WAN latency budgets, and what 'multi-region' actually means for a real trading desk.
ZeroCopy runs 11 ArgoCD applications on DOKS. How we use app-of-apps, sync waves, resource hooks, and manual sync policies to deploy trading services safely.
At Akuna with 500+ test environments, manual provisioning was killing iteration speed. How we built PR-triggered ephemeral environments at $50 each with auto-destroy after 48 hours.
A line-by-line breakdown of crypto trading infrastructure costs: NAT gateway egress, cross-AZ transfers, RDS vs ClickHouse, reserved instances, and how to cut the bill by roughly two-thirds.