Eliminating QA Bottlenecks with Argo CD and Linkerd: Building PR Preview Environments

Based on the Kubernetes migration, a structure was built to automatically create independent Preview environments for each PR, in order to eliminate the bottleneck caused by QA piling up on a single development environment.

The core was the Argo CD ApplicationSet Pull Request Generator. When a preview label is attached to a PR, an application in the form inflearn-courses-fe-pr-<number> is automatically created, and through Helm templates the Preview environment excludes unnecessary Ingress and minimizes Pods and replicas to reduce cost.

For access, cookie-based routing was chosen instead of subdomain separation or URL path separation. Both of those approaches had the problem of reverting to the production environment when moving between services, and also required application code modifications.

The existing ingress-nginx was disadvantageous for dynamic routing when resources with the same Host and Path were duplicated, but by leveraging the HTTPRoute/Traffic Split configuration of Linkerd—already in use for mTLS and monitoring—it was possible to safely separate PR-specific routing rules at the resource level. By simply putting the PR number into the browser cookie, users could access the corresponding PR environment using the same URL they normally use.

As a result, even when multiple features are developed simultaneously, parallel QA became possible in independent test environments for each, and communication overhead such as "is it okay to deploy now?" was greatly reduced.