Load Testing: Parallel Requests or Pseudo-Random Order?

Question

Highload

Good afternoon, colleagues.
I apologize for the banal question.
1) There is a service, browser clients access it.
2) You need to find out how many simultaneous users the service can hold.
What is the correct technique for conducting NT? Load the API with parallel requests or simulate requests by spreading them in a pseudo-random order over time (for example, within an hour) ?
If you give competitive requests for API, then this is more like a stress test. If spread over time, then a situation of weak competition may arise and some problems will be hidden. How to conduct NT correctly?

Answer 1

First, no load tests will replace the real load. The behavior of the service on a real load can be fundamentally different from the simulated load, and not for the better.
Secondly, you must understand that people use your service. Therefore, it is wrong to load the API with any one request, and even with one set of requests. Your users generate a certain load. Each user calls the API in a specific (for its type only) order.
That. want to simulate loads - know your user.
As soon as you model the user himself, you will immediately get adequate results of load testing, and the question voiced in the title will lose its meaning for you.

Answer 2

Try jmeter, it can record your actions on the site (what requests are sent and where, it collects data if you turn it on as a proxy server for the browser in which you create a test scenario) and then repeat them in multiple threads. Thus, you can create a scenario similar to real actions.
And it’s also good to test through tsung, I use tsung to test the performance of my comet serverer