72% Cloud Cost Reduction: A FinOps Playbook for SaaS

A monolithic ECS service was running on six on-demand Fargate tasks per region, 24/7, across three regions. The bill hit $41.6K/mo. Traffic was spiky but predictable — yet the autoscaler was set to on-demand only because nobody trusted spot for stateful workloads.

The diagnosis

The first thing I did was pull 90 days of CloudWatch metrics and plot actual CPU/memory utilization against provisioned capacity. The result: average utilization was 14%. The other 86% was paying for idle.

The second thing: I looked at the traffic pattern. It was spiky but predictable — peaks at 9am and 2pm local time, troughs overnight. The autoscaler was treating every spike as an emergency and never scaling down fast enough.

The fix: split, shift, and schedule

**Split** the monolith into twelve stateless services behind Caddy. Stateful bits (Postgres, Redis) stay on-demand; everything else can run on spot.

**Shift** non-critical workloads to whichever region has surplus spot capacity that hour. A Terraform-managed scheduler checks spot pricing across regions every 15 minutes and rebalances.

**Schedule** the autoscaler to pre-warm at 8:45am (before the 9am spike) and scale down aggressively at 6pm. The pre-warm eliminates the cold-start latency that justified the 24/7 on-demand baseline.

# Terraform: spot fleet with drain grace
resource "aws_spot_fleet_request" "stateless" {
  iam_fleet_role                      = aws_iam_role.spot_fleet.arn
  spot_price                          = "0.10"
  target_capacity                     = var.target_capacity
  allocation_strategy                 = "diversified"
  instance_interruption_behavior      = "terminate"
  terminate_instances_with_expiration = true

  launch_specification {
    instance_type = "t3.medium"
    ami           = data.aws_ami.latest.id
    spot_instance_metadata_options {
      http_put_response_hop_limit = 1
    }
  }
}

The CDN tuning nobody talks about

CDN tuning cut egress by another 18%. The trick: set `Cache-Control: public, max-age=31536000, immutable` on every hashed asset (JS/CSS bundles, images with content hashes in the filename). This lets the CDN cache them for a year — and more importantly, lets the browser cache them for a year.

Before: every page load fetched 2.3MB of JS from the origin. After: 180KB (just the changed chunks).

The payoff

• Bill dropped from $41.6K/mo to $11.6K/mo — **72% reduction**
• p99 latency fell 33% because the services could now scale independently — the old monolith's connection pool was the actual bottleneck, not CPU
• Zero customer-facing downtime during the migration

The lesson: most cloud bills aren't a pricing problem — they're an architecture problem. Fix the architecture, and the bill follows.