Ansible Rolling Updates

Updating every server at the same moment risks a full outage. A broken release takes the whole site offline. A rolling update changes a few servers at a time, checks that they work, and moves on. Think of repainting a row of shop windows: the owner covers one window while the others stay open for customers. Ansible supports this style through serial, delegate_to, run_once, and max_fail_percentage.

Diagram: Rolling in Batches

  Load Balancer
       │
  ┌────┼────┬────┬────┬────┬────┐
  ▼    ▼    ▼    ▼    ▼    ▼    ▼
 web1 web2 web3 web4 web5 web6 web7

 serial: 2

 Batch 1 ▶ [web1 web2]  update, test ✔
 Batch 2 ▶ [web3 web4]  update, test ✔
 Batch 3 ▶ [web5 web6]  update, test ✔
 Batch 4 ▶ [web7]       update, test ✔

 Servers outside the active batch keep serving visitors.

The Serial Keyword

Add serial at the play level. Ansible finishes all tasks on one batch before touching the following batch.

- name: Update web servers safely
  hosts: webservers
  become: true
  serial: 2
  tasks:
    - name: Install the new package version
      ansible.builtin.package:
        name: shop-app
        state: latest

The value accepts several formats.

  • serial: 1 updates one host at a time
  • serial: "25%" updates a quarter of the group per batch
  • serial: [1, 3, "100%"] starts with a canary host, continues with three hosts, and ends with all that remain

The Canary Pattern

A canary host takes the update before anyone else. Staff watch the result, and the rest of the fleet follows after a clean report. The list form of serial creates this pattern in one line.

Delegation

The delegate_to keyword runs a task on a different machine while facts and variables still belong to the current host. A typical case removes a web server from the load balancer before the update.

tasks:
  - name: Remove this server from the load balancer
    ansible.builtin.command: /usr/local/bin/lb-drain {{ inventory_hostname }}
    delegate_to: lb1

  - name: Update the application
    ansible.builtin.package:
      name: shop-app
      state: latest

  - name: Add this server back to the load balancer
    ansible.builtin.command: /usr/local/bin/lb-enable {{ inventory_hostname }}
    delegate_to: lb1

Use delegate_to: localhost for work that belongs on the control node, such as sending a chat notification or calling a web service.

Run Once

The run_once: true keyword executes a task on a single host in the batch. Database migrations, release tagging, and ticket updates fit this keyword.

- name: Run the database migration once
  ansible.builtin.command: /opt/shop/migrate
  run_once: true
  delegate_to: web1

Stop When Things Go Wrong

Set max_fail_percentage to abort the play when too many hosts fail. A value of 0 halts on the first failure inside a batch.

- hosts: webservers
  serial: 3
  max_fail_percentage: 0

The any_errors_fatal: true keyword stops the whole play across all hosts after any single failure.

Health Checks Between Batches

post_tasks:
  - name: Confirm the site responds
    ansible.builtin.uri:
      url: "http://{{ inventory_hostname }}/health"
      status_code: 200
    register: health
    retries: 5
    delay: 10
    until: health.status == 200

A failed health check halts the rollout before the following batch starts, which limits the damage to a small group of servers.

Strategy Options

StrategyBehavior
linearDefault. Every host finishes a task before any host starts the following one
freeEach host runs at its own speed through the task list
host_pinnedKeeps a worker on one host until it finishes

Try This

Build an inventory with four hosts that all point to localhost through different aliases. Run a play with serial: 2 and a debug task that prints inventory_hostname. Watch Ansible process two aliases per batch. Change the value to "50%" and compare the output.

Summary

Rolling updates protect uptime by changing small batches of servers. The serial keyword defines batch size, and a list creates canary stages. The delegate_to keyword runs a task on another machine, and run_once runs a task a single time. Health checks and max_fail_percentage stop a bad release early.