Consize — The Demo Script

The 10-minute story: "Here is a cluster with real waste. Watch Consize find it, fix it, prove the savings — and then prove it can take it all back if something breaks."

This is the script for the recorded demo video and the README's cover story. Every number in it is produced live by the running product — no screenshots of mockups.

0. Setup (before the camera — 20 minutes)

Sandbox GKE cluster with Consize deployed (Terraform + Helm, all from make deploy).
10 fixture workloads with inflated requests (e.g., requests: 8Gi / usage: 300Mi), 2 with realistic usage, 1 excluded (consize.savings.dev/exclude=true), 1 in a protected namespace.
One oversized RDS instance (e.g., db.r6g.2xlarge running at ~9% CPU) with the maintenance window set to tonight.
Slack/alert channel wired to the Consize alerts.

1. The pain (30 seconds)

Show the cluster: kubectl get pods with requests: 8Gi — and the Prometheus graph showing 300 MiB of actual usage. One sentence: "This is normal. This waste exists in every cluster. Nobody fixes it because fixing it manually is a spreadsheet job — and risky."

2. Consize finds it (90 seconds)

Open the dashboard → Recommendations, sorted by savings.

10 compute workloads with recommendations, each with: current vs proposed, usage chart (14-day percentiles), rationale ("p95 = 1.1 Gi; request 1.4 Gi covers 95% of real usage with 20% headroom"), savings per month.
The DB instance: one-class-down recommendation with headroom guarantees shown.
The excluded and protected workloads: present in the list with skipped — policy and the reason — "Consize doesn't touch what you tell it not to."
Total: $2,340/month projected.

3. Consize fixes it (3 minutes)

Click Apply (dry-run) on one workload → show the exact patch diff. "Nothing changed — this is the preview."
Click Apply → watch the rollout in the workload history.
Run the live verification: show the dashboard flipping to verified after the window with baseline vs post-SLI charts flat.
Apply the rest (namespaces are auto-apply labeled; Consize steps each down in ≤30% increments).
DB: attempt apply → "blocked — outside maintenance window" → approve with the window → verify.

4. Consize proves it (60 seconds)

Savings overview: realized $2,260/month — with the per-workload evidence trail and the team attribution. "Projected is a number; realized is an audit trail." Show the apply audit table: every change, who, when, diff, verdict.

5. Consize takes it back (2 minutes — the money shot)

Deploy a latency bug to one workload (pre-staged), then apply its rightsizing. Watch:

Verifier detects error rate +50% / p99 +30%.
Automatic rollback fires — pod requests/limits restored to previous values.
Alert hits Slack: "rollback triggered, evidence link".
Recommendation marked rolled_back with the diff and the SLI charts attached.

One sentence: "Consize only saves money when it's safe. When it isn't, it's the fastest thing in the building at saying 'I was wrong'."

6. Close (30 seconds)

Dashboard summary shot: savings trend, zero rollbacks (except the one we just did), data quality green.
The pitch line: "Find the waste. Remove it safely. Prove it."
Link in the description: repo, docs, architecture diagram.

Recording notes

Screen at 4K; terminal font large; every step reads from the live UI (no edits).
The rollback section is rehearsed 3× before recording — it's the part people remember.
Subtitle the demo: the video is also the README for people who won't read.

Script variants

15-second version: recommendations → apply → verified → savings number.
5-minute version: drop section 1 (pain) and section 6 (close) to a sentence each.