Use case · PKE Flex

A cluster for every pull request. Gone when it merges.

Open a pull request, get a real Kubernetes cluster in your pool in about a minute. Close it, the cluster is gone.

Start with Flex See it step by step

Own API server, CRDs, RBAC

Production version and policies

URL, TLS, quotas, clean-up included

acme/shopOpen

feat: new checkout flow #1428

feat/checkout → main

P pke-bot · Preview ready in 58 s
pr-1428.preview.pocteo.tn

Your Flex pool · Team8 vCPU · 16 GB

pr-1428

pr-1431

ci-882

demo-q4

free

4 live · 2 freeenv=preview

pr-1428.preview.pocteo.tn

Checkout · new

Lifetime

09:12 opened 09:13 ready 11:40 merged 11:40 deleted

2 h 28 min of capacity, not a month

~1 minto a ready cluster

2 callscreate on open, delete on close

1 poolany number of previews

0environments left behind

The problem

One staging environment, one queue

Every branch waits for the same shared staging. It idles most of the week and drifts from production.

Beforeshared staging

#1428 checkout flowdeploying

#1431 VAT roundingwaiting

#1436 wishlistwaiting

#1440 bump postgresblocked

used40 of 168 h

Afterone cluster per pull request

pr-1428

pr-1431

pr-1436

pr-1440

four reviews in parallel2–4 h each

Reviews wait in line

Paid 168 h a week, used 40

Drift with every manual fix

Definition

Short-lived by design

A production-like environment that exists only while someone needs it.

Mirrors productionsame version, same policies

Made by the pipelinefrom a template, never by hand

Own datafixtures seeded on every start

One URLanyone on the team can open it

How long one lives

CI run

20 min

pull request review

2–4 h

QA campaign

2 days

workshop

1 week

permanent staging

720 h a month

Step by step

Pull request to preview URL in three minutes

Auto-plays · click a step.

Current state: Creating

Platform log · 09:12Pull request opened

$ pke flex create pr-1428 --pool team --ttl 72h \
    --tag env=preview,app=shop --branch feat/checkout
Pool team: 8 vCPU, 16 GB. 4 clusters live, room for 2 more.
Creating virtual cluster pr-1428...

pr-1428

API server · starting Application · not yet URL · reserved Policies · pending

env=previewapp=shoppr=1428ttl=72h

Capacity is already reserved: creating a virtual cluster is scheduling pods, not booting machines.

Isolation

A cluster per pull request, not a namespace

Namespaces share everything cluster-wide. Dedicated clusters cost like it. Flex sits in between.

Namespace per pull request

one shared control planeno CRDs, no own versionblast radius: the whole cluster

PKE Flex: a virtual cluster each

own API server, CRDs, RBACproduction version by tagabout a minute, from your pool

Dedicated cluster each

own nodes, own control plane15–25 min to createone bill per environment

Division of labour

Two calls from your pipeline. The rest is the product.

Your pipeline

Trigger on open and close

Build the image

pke flex create pr-1428 \
  --pool team --ttl 72h \
  --tag env=preview,app=shop

pke flex delete pr-1428

PKE Flexincluded

Cluster in about a minute

Own API server, CRDs, RBAC

pr-N.preview.your-domain

TLS on the wildcard

Version and policies by tag

Quotas from the pool

TTL, sleep, delete

SSO kubeconfig per cluster

Fixtures or masked snapshot

gVisor sandboxes

Cost

Pay for the pool, not per environment

Reserve a pool once. Previews come and go inside it; the invoice does not move.

How many previews fit in your pool?

vCPU per preview: 1 vCPU Memory per preview: 2 GB Lifetime: 3 hours

6previews at once in a Team pool

With previews that live about 3 hours, that is room for roughly 12 pull requests a day, 264 a month, with no change to the pool.

vCPU is the limit here (1 + 0.3 per cluster). Control plane of each virtual cluster included. Pools on the PKE Flex page.

One week, hour by hour paid, idle · 168 hstaging used · 40 ha preview · 2–4 h

1 poolreserved, billed monthly

0fees per environment

Liveupgrade, no restart

Data

Every preview gets its own data

Seeded by the template, deleted with the cluster.

Database with fixturesseeded in seconds, syntheticdefault

Masked snapshot from Primeproduction-shaped, PII maskedregression tests

Schema onlymigrations, no rowsmigration reviews

Shared staging databasepreviews contaminate each otheravoid

Beyond pull requests

The same pattern, six more ways

Create from a template, tag it, use it, delete it.

Tests per pipeline runa fresh cluster each run

QA on demandan afternoon on the release branch

Demo environmentsone per prospect, gone after the call

Training and workshopsa cluster per trainee

Coding-agent sandboxesown cluster, hard TTL

Hotfix rehearsalproduction's exact version

Wiring

Two calls from your pipeline, or none with Argo CD

Every Flex cluster is registered in your Argo CD with its tags the moment it exists.

1Open → create the cluster

2Push → redeploy the same cluster

3Close → delete it. The TTL does it anyway.

.github/workflows/preview.yml

# .github/workflows/preview.yml
on:
  pull_request:
    types: [opened, synchronize, closed]

jobs:
  preview:
    runs-on: ubuntu-latest
    env:
      CLUSTER: pr-${{ github.event.number }}
    steps:
      - uses: actions/checkout@v4
      - name: Create the preview cluster (idempotent)
        if: github.event.action != 'closed'
        run: |
          pke flex create $CLUSTER --pool team --ttl 72h \
            --tag env=preview,app=shop \
            --branch ${{ github.head_ref }}
      - name: Deploy this commit
        if: github.event.action != 'closed'
        run: |
          helm upgrade --install shop ./deploy/chart \
            --kubeconfig $CLUSTER.kubeconfig \
            --set image.tag=$CLUSTER-${{ github.sha }} \
            --set ingress.host=$CLUSTER.preview.pocteo.tn
      - name: Comment the URL
        if: github.event.action == 'opened'
        run: gh pr comment ${{ github.event.number }} \
          --body "Preview ready: https://$CLUSTER.preview.pocteo.tn"
      - name: Delete it when the pull request closes
        if: github.event.action == 'closed'
        run: pke flex delete $CLUSTER

Create on open, redeploy on every push, comment the URL once, delete on close. pke flex create is idempotent, so the same job serves open and synchronize.

Getting started

A working week to previews on every pull request

Day 1One service, one label

Day 2Data fixtures

Day 3Tests on the preview

Day 4Quotas, TTL, sleep

Day 5Every repository

Give your next pull request its own cluster

Reserve a pool, add the workflow, open a pull request.

Choose a Flex pool From the developer portal