// How it works

Scale to Zero
No Cold Starts

How Architect checkpoints, migrates, and wakes running pods. No code changes, no custom CNI, no analytics suite to babysit.

Try it in your cluster>>Read the docs>>
● ● ●loopholelabs · -zsh · 80×24
$ helm install architect
/~\ architect deployed: revision 1

$ kubectl get pods
NAME                     STATUS    CPU    MEMORY
frontend-7d4b8c6-x9k2    Running   0m     0Mi  # <- hibernated
api-server-5f7c9d-h3j7   Running   95m    512Mi  # <- active
ml-model-8c9d2f-p4m8     Running   0m     0Mi  # <- hibernated

$ curl frontend.example.com
200 OK (47ms)  # <- instant wake, no cold start

~ Scaling idle pods to zero

> 96 pods · 96 running · most of them idle

1.

Always on, mostly idle

Typical clusters pay a capacity tax to absorb spikes and dodge cold starts. Every pod stays scheduled, provisioned, and billed, even when nothing is happening.

2.

Deploy Architect

One-line install, zero code changes. Architect begins continuous checkpoint/restore on every pod, captured at the containerd shim layer.

3.

Idle pods hibernate in place

Architect hibernates idle pods automatically on network or exec activity. You tune the behavior. Hibernated pods consume no CPU or memory but stay scheduled.

4.

Cluster rightsizing & bin-packing

With idle pods hibernated, the cluster bin-packs onto fewer nodes. Drop node count or size to reclaim the freed capacity.

5.

Traffic arrives: wake in under 50ms

Architect restores the full pod state and reattaches storage. The eBPF router holds the incoming request until the pod is live, then delivers it. No cold start, no dropped requests.

# Surviving spot reclamation

Spot reclamation isn't a sudden crash. The cloud fires an eviction warning first, and Architect starts migrating the moment it lands.

AWS warns2 min
GCP / Azure warn30s
Architect needs~10s
1Freeze. Architect checkpoints the container's full running state, including its established TCP connections. <1s
2Transfer. The checkpoint moves to another spot node. An eBPF router buffers every packet bound for the container during the move. 3–8s
3Resume. The container resumes on the new node, and the router drains the buffered packets to it. <1s

TRADITIONAL RECOVERY ~60s

detect → schedule → pull image → load snapshot → warm cache → reconnect

↳ downtime, writes lost

ARCHITECT MIGRATION ~10s

freeze → transfer state → resume

↳ zero data loss, clients stay connected

You've seen how it works.
Now try it in your cluster.

Stateful workloads on spot, idle pods at zero, sub-50ms wakes, no code changes. A 10-minute install and Architect starts managing your pods.

Book a demo>>