Read a sample: 1. Thinking Like a Deployment Decision-Maker
A deployment decision begins before anyone chooses a button, command, or AWS service. It begins with a question: What must remain true while the application changes?
For one team, the answer might be that customers can continue placing orders throughout the release. For another, it might be that a small internal tool is updated quickly during an agreed maintenance window. Both teams could be deploying the same application package. Their best release plans could still look very different.
This distinction matters in production and in certification study. A deployment strategy is not good because its name sounds sophisticated. It is good when its behavior fits the application’s constraints, provides useful evidence that the release is working, and gives the team a credible way to recover when it is not.
This chapter introduces a repeatable way to make that judgment. You will follow one release through several business situations, then build a framework for comparing options without losing sight of the application, its users, or its data.
Imagine that a small team maintains Harbor Orders, a web application used to submit and track customer orders. The team has prepared a release that improves order search and changes some application logging. The release has passed the team’s functional tests, and the team can identify both the new build and the currently running build.
Harbor Orders runs on multiple application instances behind a load balancer. For this introductory scenario, assume the release does not require a database schema change and that both application versions can use the existing data format. Those assumptions simplify the decision; they do not make the release risk-free. A code defect, unexpected production traffic pattern, or configuration mistake could still cause problems.
The team needs to choose how to introduce the new version. Consider three different operating conditions.
In the first situation, Harbor Orders is used by a small internal operations team. Users have agreed to a maintenance window, and a short interruption is acceptable. The deployment team wants a simple process and has someone available to check the application immediately afterward.
An all-at-once deployment may be reasonable. The team updates all application instances during the window, runs its validation checks, and restores the previous build if the new one fails.
The appeal is not that all-at-once is universally safer or faster. It is that the business has accepted temporary unavailability, making a simple release process a plausible fit. The team should still know how long restoration is likely to take and what it will do if the previous build cannot be redeployed promptly.
Now suppose customers use Harbor Orders throughout the day. A complete interruption would be costly, but the team has limited budget for duplicate infrastructure. The application can tolerate a temporary reduction in serving capacity, and its old and new versions can run side by side.
A rolling deployment may fit. The team updates instances in batches, leaving the remaining instances to serve requests. After each batch, it checks whether the updated instances are healthy before continuing.
This reduces the need to replace the entire fleet at once, but it introduces other questions. Can the remaining instances handle demand while a batch is unavailable? Are the health checks meaningful? Will a request behave correctly if it reaches the old version once and the new version next? If a later batch fails, what is the procedure for returning the updated instances to the previous build?
Rolling deployment trades an abrupt change for a period of mixed versions and potentially reduced capacity. Whether that trade is worthwhile depends on conditions the team must establish, not on the strategy’s name.
In the third situation, Harbor Orders supports a major sales event. The business wants to limit customer exposure to an unproven release and is willing to pay for additional capacity and closer observation.
The team might prepare a separate environment for the new version, validate it, and then shift traffic from the old environment. Alternatively, it might direct a small portion of traffic to the new version first, observe the results, and increase exposure only if the signals are acceptable.
These approaches are often discussed as blue-green and canary releases. Blue-green emphasizes separate environments and a traffic transition between them. Canary emphasizes limited initial exposure and gradual progression. Depending on the platform and implementation, a team can combine elements of both.
The extra environment or traffic controls can improve isolation and make certain recovery actions quicker. They also add cost and operational complexity. A traffic shift back to the old environment helps only if that environment is still usable and compatible with any changes made while the new version was serving requests.
The important lesson is not that the third option is the most advanced and therefore the best. It is that each situation makes different costs acceptable. The application package is the same; the business constraints are not.
When faced with a deployment scenario, resist the urge to select a familiar AWS service immediately. First describe the decision in plain language.
For Harbor Orders, a useful statement might be:
> Introduce a tested application build while keeping customer disruption within an agreed limit, detecting meaningful failures, and preserving a practical route back to the previous build.
That statement identifies a change, an operating requirement, a validation need, and a recovery goal. It does not yet prescribe a tool. The same discipline helps with exam scenarios: identify what the…