Cut Compute Spend by 30% Before Q3

Compute costs $14.2M a year, 61% of the cloud bill. It grew 48% last year while traffic grew 22%, and finance has asked engineering to cut it by 30% before the end of Q3.

Move every Kubernetes node pool to spot instances, including the pools that run Kafka, Elasticsearch, and the Postgres replicas. Spot capacity costs 60 to 70% less than on-demand.

Each service's CPU and memory requests are set to its average usage over the last 30 days plus 10% headroom.

Production services and databases consolidate from three availability zones into one, which removes $1.1M a year in cross-zone data transfer.

Once usage settles, we buy three year, all upfront reserved instances for the remaining baseline.

Spot saves $5.0M, rightsizing saves $2.4M, and the zone change saves $1.1M, for $8.5M a year, well above the $4.3M target.

The infrastructure team applies every change centrally, so product teams don't need to spend time on the effort.

Progress is tracked against the monthly cloud invoice and reported to finance at the end of each quarter.