How Can I Effectively Load Test Single Page Applications with Thousands of Users?

Full-scale headless browsers struggle to accurately replicate thousands of real user interactions. A better strategy is to mix your methodology: use Puppeteer or Playwright for a handful of human-like sessions to add realism, while employing API-level testing tools like k6 or Locust for the bulk of your load testing. Distributing the load across multiple machines and introducing small variations in recorded user flows will help accomplish more realistic metrics.

Why simulate mouse movements or clicks at all? That approach isn't necessary since it's mostly about the server-side performance. Instead, load testing the backend by sending a high volume of requests can give you better insights.

Have you considered synthetic monitoring from real edge locations? Instead of overwhelming your infrastructure with simulated users, it might be better to test the API layer separately where the metrics are more relevant. This way, you can achieve a more genuine understanding of performance.

Stick with k6, but avoid using it in browser mode. Focus on real user metrics, such as HTTP requests per second, to align with actual concurrent users. This helps determine the user-to-workload ratio. In k6, try to replicate the same backend requests that the real users would make, ensuring that your simulated users match the load of actual users. Mimicking 5,000 browser sessions can be resource-intensive, so concentrate on the server-side traffic instead.

When testing SPAs, remember that thousands of real users are operating from thousands of browsers, each with their own interactions. Therefore, it’s more effective to focus on testing the backend of the application rather than simulating all the clicks and mouse movements that users would perform. You might want to consider simulating API requests which more closely represent real-world usage.