Run Jobs on Preemptible ECI Pods to Cut Costs - ACK Serverless

Batch jobs and fault-tolerant workloads can be expensive to run continuously on on-demand instances. Preemptible instances let you run these workloads on Elastic Container Instance (ECI)-based Pods at significantly lower cost, with the Kubernetes Job controller automatically replacing any Pod that is reclaimed before the job completes.

Suitable workloads

Use preemptible instances for workloads that can tolerate interruption:

Suitable	Not suitable
Batch jobs and one-off tasks	Stateful applications
Stateless, fault-tolerant applications such as scalable web services	Long-running services that cannot handle mid-task interruption
Short-lived compute tasks where restarts are acceptable	Workloads whose startup or shutdown time exceeds the 5-minute pre-release window

How it works

When you create a preemptible instance, you specify a bid policy. The instance is created when your bid price exceeds the real-time market price and the resource inventory for that instance type is sufficient.

Protection period: After creation, the instance runs for a 1-hour protection period (default), billed at the market price at the time of purchase.

After the protection period: The system checks the market price and inventory every 5 minutes. If the market price exceeds your bid price, or inventory runs out, the instance is released. A SpotToBeReleased event is sent to the Kubernetes event list approximately 5 minutes before release — use this window to stop inbound traffic or checkpoint your workload.

Job continuity: When a Pod is released before the job completes, the Kubernetes Job controller automatically creates a new Pod to continue the job. The original Pod is retained with a BidFailed status.

Limitations

By default, the protection period is 1 hour.
After the protection period, instance availability depends on market price and resource inventory — there is no uptime guarantee.
The SpotToBeReleased event arrives approximately 5 minutes before release. If your workload takes more than 5 minutes to shut down cleanly, preemptible instances may not be suitable.

Prerequisites

Before you begin, ensure that you have:

An ACK Serverless cluster
kubectl configured to connect to the cluster

Run a job on a preemptible instance

Step 1: Create the job manifest

Create a file named spot_job.yaml with the following content:

apiVersion: batch/v1
kind: Job
metadata:
  name: pi
spec:
  template:
    metadata:
      annotations:
        k8s.aliyun.com/eci-use-specs: ecs.c5.large,2-4Gi  # Multiple instance types improve creation success rate
        k8s.aliyun.com/eci-spot-strategy: SpotAsPriceGo    # Use the real-time market price as the bid price
    spec:
      containers:
      - name: pi
        image: registry-vpc.cn-beijing.aliyuncs.com/ostest/perl
        command: ["perl", "-Mbignum=bpi", "-wle", "print bpi(2000)"]
      restartPolicy: Never

The k8s.aliyun.com/eci-spot-strategy annotation sets the bid policy:

Value	Description
`SpotAsPriceGo`	Automatically uses the market price at the time of purchase as the bid price.
`SpotWithPriceLimit`	Uses a maximum hourly price that you set. Add the `k8s.aliyun.com/eci-spot-price-limit` annotation to specify the price.

The k8s.aliyun.com/eci-use-specs annotation lists multiple instance types. Specifying multiple types improves the chance that a preemptible instance is created when inventory for one type is low.

Step 2: Create the job

kubectl create -f spot_job.yaml

Step 3: Monitor the Pod status

kubectl get pod

The Pod transitions through the following statuses:

Running — the job is in progress:

NAME       READY   STATUS    RESTARTS   AGE
pi-frmr8   1/1     Running   0          35s

Completed — the job finished successfully:

NAME       READY   STATUS      RESTARTS   AGE
pi-frmr8   0/1     Completed   0          2h

BidFailed — the Pod was released before the job completed. The Job controller has already created a new Pod (pi-kp5zx) to continue:

NAME       READY   STATUS      RESTARTS   AGE
pi-frmr8   1/1     BidFailed   0          4h53m
pi-kp5zx   1/1     Running     0          3h45m

Note

A SpotToBeReleased event is emitted approximately 5 minutes before a Pod enters BidFailed. To view the event, run kubectl get events. For details on handling pre-release events, see Release of preemptible instances.

What's next

Create a preemptible elastic container instance — configure advanced bid policies and instance type selection strategies.