The open vs closed workload model choice decides how a test treats a slow system: in one, clients wait for the system before returning, and in the other they keep arriving no matter what. Beginners often take whatever their tool does by default and never notice, which makes it one of the quieter mistakes in stress testing.
Both models are legitimate tools. They answer different questions, and mixing them up produces numbers that look precise and describe a different test. Each gets its own section below, followed by a short guide to which one fits which job.
Closed model: a fixed set of clients
A closed workload model keeps a set number of clients active, and each one completes its journey before beginning another. Grafana’s k6 documentation describes it as a model where the execution time of each iteration dictates how many iterations the test runs, since the next iteration does not start until the previous one finishes.
That has a cost. When the target slows down, a closed test waits, so iterations take longer and fewer requests reach the system in the same time. The k6 documentation adds that some testing literature calls this problem coordinated omission. A closed test eases off exactly when the system gets slower, which is the moment you most want the pressure to continue.
It also fits poorly when the goal is a certain arrival rate of new virtual users, or a throughput such as requests per second.
Open model: arrivals at a chosen rate
An open workload model starts new journeys at a chosen rate, and new arrivals continue even when earlier requests take longer to complete. In a software performance testing article on workload modelling, this is the model said to reveal growing queues that closed tests may hide. k6 describes the same idea as virtual users arriving independently of iteration completion, and implements it in its arrival-rate executors.
The model also makes capacity arithmetic simple. Take the rate from observed traffic: a recorded peak of 600 journeys per minute is 10 per second. If those journeys last 12 seconds on average, roughly 120 journeys overlap, a number you can compare with pool and queue limits before the test starts.
Either model still needs think time. Real customers read descriptions and reconsider quantities before moving on, so scripts should pause wherever the recorded behavior supports it.
What a mix-up looks like in a stress test
Picture a promotion-day test built as a closed model with a fixed number of clients. Payments start to slow down. Each client now spends longer inside its journey, so fewer new journeys begin every second and the recorded promotional level never fully reaches the system. The report shows a mild slowdown and a green result. An open model would have kept sending journeys at the planned rate, and the growing queue would have shown up in the numbers.
The mistake is a stress test that relies on the system being fast enough to keep up with the test, and a closed test does not announce the problem. It simply sends less. A simple safeguard is to compare the planned arrival rate with the rate the run reached. When the two differ, the test did not deliver the load you meant it to.
Open vs closed workload model: when each one fits
| Situation | Better fit |
|---|---|
| Holding a target arrival rate, such as journeys per second | Open |
| Reproducing a recorded peak expressed as journeys per minute | Open |
| Looking for queues that keep growing while the system slows | Open |
| Running a fixed number of clients that each finish before starting again | Closed |
Different models suit different test aims, as the k6 documentation puts it. For a basic plan, the peak phase is defined by arrivals, so an open model is the natural choice there, and the spike phase follows the same logic. A short warm-up has less to gain from the choice.
A practical way to decide is to look at what you recorded. Logs that describe arrivals per minute point to an open model. Whichever you pick, write it into the test report, because a result is hard to read when nobody knows whether the load kept coming or paused.