Deployment strategies and rollbacks: DVA-C03 study notes and practice questions

What to know for the exam

These questions describe a release constraint, such as zero downtime, a small blast radius or a fast undo, and ask for the strategy that meets it.

The strategies

• All at once moves all traffic in one step: fastest and riskiest.
• Canary shifts a small share of traffic first, waits, then shifts the rest. Use it when a short test period on real traffic is the goal.
• Linear shifts traffic in equal steps at fixed intervals. Use it when problems should show up gradually.
• Blue/green runs the new version beside the old one, then switches traffic. The old environment stays available, so rollback is a fast switch back.
• Rolling replaces instances in batches. Capacity dips during the update, and old and new versions serve traffic together for a while.

CodeDeploy

• Lambda deployments use deployment configurations that shift traffic between two versions of an alias, such as all at once, canary or linear. CodeDeploy can run hooks before and after traffic shifts, and it rolls back when an alarm fires.
• ECS blue/green deployments through CodeDeploy create a replacement task set, can send a test listener’s traffic to it first (the AfterAllowTestTraffic hook), shift the production listener, and terminate the old set after a wait period. ECS also has its own built-in blue/green strategy with a bake time and lifecycle hooks.
• EC2 and on-premises deployments support in-place and blue/green. In-place updates instances where they run, so capacity drops during the deployment. CodeDeployDefault.OneAtATime, HalfAtATime and AllAtOnce, or a custom minimum healthy hosts value, set how many instances update at once.

Elastic Beanstalk

• All at once and Rolling reduce capacity during the update. Rolling with additional batch launches an extra batch first, so capacity stays full.
• Immutable launches a fresh set of instances and swaps them in, so no existing instance runs a mix of versions.
• Traffic splitting sends a percentage of traffic to fresh instances for a test period, like a canary. It needs an Application Load Balancer.
• Blue/green uses a second environment and a CNAME swap, so the public URL stays the same.

Rollback

• Automatic rollback happens when a deployment fails or a CloudWatch alarm goes into alarm during the deployment. CodeDeploy and AppConfig both support this.
• Manual rollback means redeploying the last good revision, or moving an alias back.
• CloudFormation rolls back a failed stack update automatically. Rollback triggers add CloudWatch alarms that cause a rollback during the update.
• AppConfig rolls back a configuration when an alarm fires during the deployment or the bake time that follows it.

How to choose

Common traps

• Choosing all at once when the question asks for a limited blast radius.
• Choosing in-place when the question says capacity must stay full. In-place reduces capacity on each instance as it updates.
• Expecting automatic rollback without alarms. Without a configured alarm or failure, CodeDeploy does not roll back on its own.
• Confusing canary and linear. A canary has one early step, linear has equal steps.

Try the 2 practice questions ↓