A go-live is approaching, or a period of high activity: the opening of an online service, a sales campaign, an accounting close. The question put to you fits on one line: will the application hold up? A load test answers it by measurement. It subjects the system to simulated traffic, close to what you expect, and checks that response times, error rate and resource consumption remain within the limits you have set.
Its value is decided before the first run: scenarios that resemble real usage, an environment whose differences from production are known, acceptance criteria written in advance. Without these prerequisites, the campaign produces curves from which nobody can draw a decision.
What a load test proves, and what it does not tell you
A load test provides bounded proof. For a given distribution of transactions, a given data volume and a given configuration, the system met its objectives, or did not. It also brings to light what functional tests cannot see: a slow query that monopolises the database, a number of connections capped too low, cascading calls to a third-party service, a default setting that does not hold at scale.
It remains silent about what it did not simulate. A forgotten journey, a batch process running at the same time as users, an external partner replaced by an always fast simulator: so many blind spots. Nor does it say anything about holding up beyond the level tested, stability over several days or behaviour when a component fails. A compliant result holds only for its assumptions.
Load testing and performance testing: four distinct questions
“Performance testing” is the generic term. It refers to any test that assesses the behaviour of a system under demand: response time, throughput, resource consumption. Load testing is one type among others. Vocabulary varies from one team to another. The ISTQB Performance Testing syllabus (foundation level, 2018 version) distinguishes in particular load, stress, scalability, spike and endurance tests, to which it adds concurrency and capacity tests.
| Type of test | Question asked | Simulated traffic profile | What it reveals |
|---|---|---|---|
| Load test | Does the system meet its objectives at the expected level of activity? | Gradual ramp-up, then a plateau at the target level | Response times and errors under nominal conditions, bottlenecks |
| Ramp-up test | Where is the limit, and how does the system degrade? | Increasing steps, beyond the expected level | Maximum capacity, first saturated component, effect of adding resources |
| Endurance test | Does the system remain stable over time? | Sustained load over a long period | Memory leaks, progressive saturation, drift in response times |
| Spike test | Does the system absorb a sudden influx, then return to normal? | Sharp and brief rise, then return to the usual level | Reaction of autoscaling, queues, recovery time |
“Ramp-up test” is not a term from the ISTQB syllabus. The expression refers here to the test in increasing steps, carried out beyond the expected level: it brings together what the ISTQB separates into capacity testing (how far the objectives are met), stress testing (what happens beyond) and scalability testing (the ability to absorb more users or data without failing the requirements). Some teams use the same expression as a synonym for load test: have the intended meaning specified before ordering a campaign. These tests do not replace one another. An application can pass a load test and collapse after a few hours of endurance. The choice follows the risk: a ticketing launch calls first for a spike test, a continuously used service for an endurance test, a fast-growing activity for a ramp-up test.
Framing the campaign: scenarios, environment, acceptance criteria
The scenarios. The load model describes who does what, and at what pace. When the application exists, it is deduced from usage data: access logs, traffic statistics, business volumes from the busiest period. When it is new, it rests on assumptions that the business validates in writing. It specifies the distribution between journeys, the think times between two actions and the variety of data entered. Thousands of simulated users viewing the same record mostly test the cache.
The environment. An environment identical to production is rarely available. What counts is to write down its differences: sizing, database volume, network, third-party services. An almost empty database often distorts measurements more than a more modest server. External dependencies are handled explicitly: either they are included with the agreement of their operators, or they are replaced by simulators with realistic latency.
If the data set comes from production and contains personal data, its use in testing remains subject to the GDPR, Regulation (EU) 2016/679. Fictitious data generated from rules avoids the question; failing that, a pseudonymised extract requires an environment protected like production. The choice is detailed in our article on data anonymisation.
The acceptance criteria. They are written before the campaign, with the business and operations: response time per transaction, tolerated error rate, throughput to reach, level of resource usage not to exceed. There is no universal threshold. The right value comes from what your users tolerate and from your service level objectives, as defined by the approach presented in our page SRE and observability.
Reading the results without being reassured by an average
The average response time hides the essential. If one request in a hundred takes ten times longer than the others, the average increases by less than a tenth, while users are waiting. The results are therefore read in percentiles: the median for the usual case, the high percentiles for the share of users least well served.
Three curves are read together: response time, throughput and error rate. A response time that improves while errors increase is not good news. The system is quickly rejecting what it no longer processes. A throughput that plateaus while the injected load continues to grow designates a bottleneck.
The injector’s measurements signal a problem without locating it. The answer comes from the metrics, logs and traces of the system under test. A campaign carried out without server-side monitoring produces an observation without a diagnosis. Finally, an isolated result proves little: a run repeated under the same conditions should give similar measurements.
What distorts a load testing campaign
- Testing on the eve of go-live. The defects revealed often concern architecture, queries or sizing. With no time to correct and rerun, the test only serves to document a risk already taken.
- Confusing simultaneous users with throughput. The number of virtual users says nothing without the pace of their actions. Removing think times manufactures a load that does not exist in production.
- Limiting yourself to the home page. The costly journeys are elsewhere: search, payment, document generation, database writes.
- Not rerunning after a fix. Removing one bottleneck often reveals another, located just behind it.
- Not monitoring the injector. A saturated injection machine caps throughput and lengthens measured times: the bottleneck is then in the test tool and not in the application.
The place of load testing in the delivery cycle
Two rhythms complement each other. The first is continuous: short, automated tests, run by the delivery pipeline on sensitive transactions, and compared with a reference measurement. They do not prove that the system will hold up in production, but they detect a regression at the moment it is introduced. Their integration falls under the pipeline described in our page DevOps and platform engineering.
The second is occasional: a full campaign, on a representative environment, before a go-live, a major change in architecture or an announced peak in activity. After go-live, production measurements feed the load model of the next campaign.
For financial entities subject to the DORA regulation, Regulation (EU) 2022/2554, these campaigns fit together with the digital operational resilience testing programme provided for in its Article 24; its Article 25 expressly cites performance testing.
Key points before ordering a load test. Make sure you can answer four questions. What risk do you want to cover: the expected level, the limit, the duration or the spike? On what activity volumes does the load model rest? What differences separate the test environment from production? By what criteria, set before the run, will you judge the result?
We organise these campaigns alongside the other levels of control, as part of our software quality and testing offering.