Skip to content

cost-model.md

Latest commit

 

History

History
124 lines (96 loc) · 5.61 KB

cost-model.md

File metadata and controls

124 lines (96 loc) · 5.61 KB
Raw

Cost model — what each piece actually costs

An explicit price list for every moving part of the Iris stack. Updated when a new component lands or when GCP/providers change their pricing. The ADR-0022 target is ≤€2/month idle; this document is how we keep it honest.

How to use this table

  • Before adding a new component, look up its row below and add its projected cost to your mental total. If it pushes the monthly bill over €10 (the alert threshold in docs/ops/cost-control.md), you should either ask why, cut something else, or raise the budget explicitly with an ADR.
  • When the bill drifts, cross-reference the component section — whatever crept in since the last clean audit shows up here.
  • When a cost row changes (GCP price list update), bump it here together with the live bin/budget/budget.sh set <amount> call. The two should never disagree.

Prices in EUR, europe-west1, standard tier, as of 2026-04 (spot-check at https://cloud.google.com/pricing). Prices are rounded to 2 decimals or to 1 significant figure for small amounts.

Idle baseline (cluster DOWN, nothing running)

Component Unit Monthly cost Notes
GCS bucket *-tf-state 182 B <€0.01 tf state + lock, single-region, 1 object
Artifact Registry images ~500 MB ~€0.05 PD-backed repo storage in europe-west1
Secret Manager 5 active secrets €0 free tier = 6 active secrets/month
Pub/Sub topic iris-budget-kill 1 topic, 0 msg €0 billed per-message; no message = no cost
Cloud Function budget-kill 0 invocations €0 gen2 scale-to-zero, billed per invocation + vCPU-s
IAM bindings — €0 always free
DNS zones (DuckDNS external) — €0 external free service
TOTAL idle — ~€0.10 genuine idle, ephemeral pattern at rest

Per-demo transient (cluster UP for 2h)

Component Rate Per 2h demo Notes
GKE Autopilot pods €0.02 / pod-h €0.60 15 pods × 2h — billed per declared resource request
PD-balanced for PVCs €0.048 / GB-month €0.01 64 GB × 2h = 0.17 GB-month
Ingress Load Balancer €0.025 / h €0.05 1 LB, 2h — N/A if ADR-0025 stays (no ingress)
Egress (image pulls, OTel) — ~€0.02 rare, a few MB of inter-zone traffic
Cloud SQL (if enabled) — €0 deliberately disabled; in-cluster Postgres
TOTAL per demo — ~€0.70 wall-clock cluster time × component rates

At 5 demos/month (real portfolio cadence)

Idle baseline × 1 month + 5 × per-demo = ~€0.10 + 5 × €0.70 = €3.60/month.

This slightly overshoots the ADR-0022 €2 target because the per-demo pod-hour assumption there was 1h not 2h. Either:

  • Shrink demos to 1h — tight but feasible, runtime is 8min cold start
    • 15min demo-ready window + 37min to spare.
  • Accept €3.60/month, update ADR-0022 with the measurement.

The point of this doc is that either choice is deliberate, not surprise-inducing.

Per-component cost rationale (for reviewers)

The table above is the fact sheet. The reasoning below is why each component earns its line.

GKE Autopilot vs GKE Standard

Autopilot at 15 pods × 2h × 5 demos = 150 pod-hours × €0.02 = €3/month. GKE Standard 1 × e2-small always-on = €30/month. Autopilot wins at our usage because we're <20% duty cycle; it would lose at a real- traffic always-on app. Tracked in ADR-0022.

In-cluster Postgres vs Cloud SQL

Cloud SQL db-f1-micro = €7/month, db-custom-1-3840 = €50/month. In- cluster Postgres on a PVC = included in the pod-hour rate above, so effectively €0 incremental. The trade-off is we lose PITR + managed backups; the demo doesn't need them. ADR-0013.

Cloud Function budget-kill vs always-on listener

Function: €0 at rest. A Compute VM e2-micro listening to Pub/Sub 24/7 = €5/month. The Cloud Function wins because the listen-on-demand pattern is free until invoked; a VM would nullify the point of budget-kill. See docs/ops/cost-control.md for the full rationale.

Artifact Registry vs Docker Hub public

Artifact Registry: €0.05/month for our image. Docker Hub public: free but rate-limits pulls from GKE without a registered account, which has bit us once in CI. €0.05 is cheaper than the hour wasted diagnosing a 429 from Docker Hub.

LGTM bundle (Grafana) vs Grafana Cloud

Self-hosted LGTM (Tempo + Loki + Mimir + Grafana + Pyroscope in one container): included in the 15-pod count, ~€0.40/demo. Grafana Cloud free tier = free up to 10k series / 50 GB logs — easily exceeded with full OTel instrumentation on a real workload. Self-host wins while demo volume is small; Grafana Cloud wins once we have real users.

What doesn't appear here

  • Human time. The script bin/dev/iris-doctor + runbooks exist specifically to make operator time cheap, but they're not a GCP line item.
  • Email / notifications. GCP budget alerts' default email is free.
  • Renovate bot runs. Hosted by Mend, free tier.
  • gitlab.com shared runners. Not used — we run on macbook-local (ADR-0004), zero quota consumed.
  • DuckDNS. External free service for the prod DNS, outside GCP pricing entirely.

When to revisit

  • Any single component forecast crosses €5/month → open an ADR on the trade-off before the change lands.
  • Actual bill crosses €10 → budget-kill fires, cluster destroyed, incident report in docs/ops/runbooks/ with the line item that caused the drift.
  • GCP publishes a price update for europe-west1 → refresh the relevant row here, update bin/budget/budget.sh if the cap changes.