CSE Roadmap

Specialist tier · about 2 hours · 8 parts

Part 3 of 8 · 1 min read

Cutting cost

Common practice, not from the books.

The SLO of module 1 sets the floor. Cost can be cut as long as the budget is not spent. Common ways:

  • Right-size. Lower instance sizes and counts to what load tests (Level 8) show is needed, keeping headroom.
  • Scale with load. Remove instances at night if traffic is low, and add them before peaks. Orchestrators (module 6) can do this.
  • Use cheaper classes. Move old data to slower, cheaper storage. Use discounted machines for work that can be restarted.
  • Keep traffic local. Place chatty services in the same zone or region. Use a CDN (Level 1) so fewer bytes leave the origin.
  • Cut observability waste. Apply the sampling, levels and rate limits of module 3, and keep fewer metric labels.
  • Delete what is not needed. Which also helps privacy (section 4).

Two cautions. First, some savings cut resilience: fewer zones lower the bill and also the availability (Level 8). The error budget tells you whether the saving is allowed. Second, an engineer's time is a cost too. A change that saves 50 a month but costs a week is a bad trade.

A made-up case. Raising instance use from 30 percent to 60 percent lets a service run on 10 instances, not 20: saving 730 a month at the earlier price. A load test shows p99 still within the SLO.

Check yourself: 1. What limits how far cost can be cut? 2. Name one saving that lowers resilience. 3. Why is engineer time part of cost?