The Real Cost Model of Matrix Builds — How a 4-Axis Matrix Cost $6,400/Month

Matrix builds feel almost free. One line, N environments validated. Past four axes, cost explodes multiplicatively.

This post is our cost ledger from a six-month 3 → 4 axes progression.

Start — 3 axes, 9 jobs

matrix:
  os: [ubuntu-22.04, ubuntu-24.04, macos-14]
  node: [18, 20, 22]

3 × 3 = 9 jobs. Expected CI time: 1 job × 4 min × 9 = 36 min total, parallel-8 → ~8 min wall clock.

Pleasing. Every OS × every Node validated in 8 minutes.

Six months later — 4 axes, avg 5, 60 jobs

The matrix grew. Each new requirement added an axis:

matrix:
  os: [ubuntu-22.04, ubuntu-24.04, macos-14, windows-2022]   # +1
  node: [18, 20, 22]
  database: [postgres-15, postgres-16, mysql-8]              # new axis
  region: [us-east-1, eu-west-1]                             # new axis

Now 4 × 3 × 3 × 2 = 72 jobs. Some combos are invalid (Windows + macOS-only DB), so avg 60 actual.

CI time — 5 min per job (DB setup adds time). 60 × 5 = 300 min total, parallel-8 → ~38 min wall clock.

Cost ledger

Initial (3 axes, 9 jobs):
  CI time per PR:          8 min
  PRs per month:           280
  CI minutes per month:    2,240 = 37 hr
  GitHub Actions cost:    ~$30 / mo

Six months later (4 axes, avg 60 jobs):
  CI time per PR:          38 min
  PRs per month:           280  (unchanged)
  CI minutes per month:    10,640 = 177 hr
  GitHub Actions cost:    ~$1,800 / mo

Net increase:              $1,770 / mo = $21,240 / yr

That's direct cost. Indirect is larger.

Indirect

Extra wait time per PR:    30 min
PRs per month:             280
Engineer wait per month:   8,400 min = 140 hr
Annualized:                ~$84,000 (engineer rate)

Four times the direct cost. Each axis added means the whole team waits longer.

Decision matrix for adding an axis

Before adding, check:

What we did — axis tier separation

Six months in, we split the matrix into three tiers:

1) PR matrix (every PR):
   os × node = 9 jobs (8 min)
   Purpose: fast feedback

2) main-merge matrix (right after merge):
   os × node × database = 27 jobs (15 min)
   Purpose: integration validation

3) Nightly matrix (once daily):
   os × node × database × region = 60 jobs (38 min)
   Purpose: edge-combination validation

After splitting, PR CI time recovered to 8 min. Monthly cost $1,800 → $400. Indirect cost $84,000 → $14,000.

What matrix builds actually mean

A matrix isn't a tool for validating every combo on every PR. It's a distribution tool for which combos to validate when. Without explicitly designing this distribution upfront, you default to every combo on every PR — which is the cost-explosion path.

So should you add that matrix axis?

Adding a new axis to a CI build matrix is justified only when it passes the four-question decision matrix above. Any team running matrices on 2+ axes that runs the check on every new axis prevents the cost explosion six months out. We skipped it, and only turned back at $1,800/month direct and $84,000/year indirect.