Website

Load Testing Before Launch: How to Find Out What Your System Can Take

November 11, 2026 · 7 min read · Etzal Group Team

A company launches a new ordering portal for its 800 dealers on the first of the month. It was tested for weeks by the project team of six. At 09:05, as dealers log in to place month-start orders, pages take forty seconds to load. By 09:20 the database stops accepting connections. The portal is taken down, and dealers are told to fax their orders. Nothing in the software was wrong in the sense of a bug. Nobody had ever seen it with more than six people on it. Load testing is how you see that morning in advance.

What load testing answers

  • How does the system respond at the number of users we expect?
  • At what load does it start to slow, and where?
  • What breaks first?
  • Does it recover when the load drops?
  • Does it stay healthy over many hours?

A functional test asks "does it work?" A load test asks "does it still work when everyone uses it at once?"

Step 1: set targets from real usage

Targets must come from the business, not from a round number.

Estimate the peak.

  • How many users in total? Say 800 dealers.
  • When is the busiest time? The first working morning of the month.
  • What share are active in that hour? Perhaps half: 400.
  • What does each do? Log in, browse the catalogue, add twenty lines, submit an order, check status.

Turn that into a rate. If 400 users each perform about 60 actions in an hour, that is 24,000 requests an hour, around 7 per second on average. Peaks within the hour will be several times the average. Plan for a peak of perhaps 25 to 30 per second.

Set response targets. For example:

  • 95 percent of page loads within 2 seconds.
  • 95 percent of order submissions within 3 seconds.
  • Error rate below 0.5 percent.

Use the 95th percentile, not the average. An average of one second can hide one user in ten waiting fifteen.

Add headroom. Test to at least twice the expected peak. Estimates are often low, and growth is the plan.

Step 2: build realistic scenarios

A test that requests the home page ten thousand times proves little.

  • Model journeys, not single requests. Log in, search, open items, add to cart, submit.
  • Mix them realistically. Seventy percent browsing, twenty percent ordering, ten percent reports.
  • Include pauses. Real users read and think between clicks.
  • Vary the data. Different users, different products, different search terms. Repeating one request tests the cache.
  • Include the heavy operations. The monthly report, the large export, the search with no filter. One of these often does the damage.
  • Use realistic data volume. A database with a hundred test orders behaves nothing like one with three years of history. Load production-sized data into the test environment.

Step 3: prepare the environment

  • As close to production as you can. Same sizes, same configuration, same database volume. If it is smaller, note it: results will not transfer directly.
  • Isolate it from real effects. Disable or stub anything that sends emails, messages or payments, or calls third parties who have not agreed to a test.
  • Tell your hosting provider if required. Some need notice.
  • Set up monitoring before you start. A load test without server measurements tells you that it is slow, not why.

Step 4: run four kinds of test

1. Load test. Ramp up gradually to the expected peak and hold for 30 to 60 minutes. Does it meet the targets?

2. Stress test. Keep increasing beyond the peak until something fails. Where is the limit, what fails first, and how? A system that slows gracefully is better than one that collapses.

3. Spike test. Jump suddenly from low to high load, as at 09:00 on the first of the month. Can it absorb the surge, and does it recover afterwards?

4. Soak test. Hold a moderate load for many hours, ideally overnight. This reveals slow problems: memory that creeps up, connections that are never released, disks filling with logs.

Run them in that order. Fix what the first reveals before moving on.

Step 5: watch five measurements

From the test tool:

  • Response time, at the median, 95th and 99th percentile.
  • Throughput: requests handled per second.
  • Error rate.

From the servers:

  • Resource use: processor, memory, disk and network, on every component: web, application, database, cache.
  • Saturation signals: database connections in use, queue lengths, threads waiting.

Read them together. The pattern tells the story:

  • Response time rises while processor use is low: something is waiting. Usually the database, a lock, or an external service.
  • Throughput stops increasing as load rises: you have reached a ceiling. Find the component at its limit.
  • Errors appear suddenly at a threshold: a pool, a limit or a timeout has been reached.
  • Memory climbs steadily in a soak test: a leak.

The usual bottlenecks

Most business systems fail in a small number of familiar ways.

  • Slow database queries. A missing index, or a query that reads far more than it needs. The commonest cause by a wide margin.
  • Too many queries per page. A list of fifty items that runs fifty-one queries.
  • Connection pool exhausted. Requests queue for a database connection.
  • No caching of data that rarely changes, such as the catalogue.
  • Heavy work done while the user waits: generating a report, sending an email, calling a slow external service. Move it to a background queue.
  • A single component that cannot be scaled.
  • Third-party limits: a payment or messaging provider that throttles you.
  • Large responses: sending thousands of rows when the screen shows twenty.

Fix the largest one, test again, and repeat. Each fix moves the bottleneck somewhere else, and the aim is to move it beyond your target.

Reading results without a specialist

A decision-maker needs four facts from a load test report.

  1. Did it meet the targets at expected peak? Yes or no, with the numbers.
  2. How much headroom is there? At what multiple of the peak did it fail?
  3. What fails first, and what would it cost to raise that limit?
  4. Did it recover by itself after overload?

If the report cannot answer these, ask for them.

When to test

  • Not at the very end. A test in the final week leaves no time to fix what it finds.
  • Early, on the riskiest parts. As soon as the main journeys work.
  • Before launch, as a full rehearsal.
  • After significant changes.
  • Before known peaks: a campaign, a seasonal high, month end.

Keep the test scripts. Rerunning them after changes is cheap, and it shows whether performance has drifted.

Plan for the day it is exceeded

No test guarantees the future. Decide in advance what happens when load goes beyond what was tested.

  • Degrade gracefully. Switch off non-essential features, such as heavy reports, under pressure.
  • Queue or limit. A waiting page is better than a crash.
  • Alerts before the limit, not at it.
  • A way to add capacity quickly, and someone who knows how.
  • A fallback process for the business.

Common mistakes

  • Testing only one page.
  • A tiny test database.
  • Identical requests, which only exercise the cache.
  • No think time, making 100 simulated users behave like thousands.
  • Looking only at averages.
  • No server monitoring.
  • Testing against live third-party services without agreement.
  • Sending real emails or messages from a test.
  • Testing once, at the end.
  • Declaring success at expected load without finding the limit.

We include load testing in the plan for systems where a peak matters, as part of Enterprise Software. If you have a launch date and no idea how the system behaves under real numbers, contact us.

Summary

Derive targets from real usage: users in the peak hour, what they do, and response times at the 95th percentile, with headroom of at least double. Build journeys with realistic mixes, pauses and data volume, in an environment like production with outgoing effects disabled. Run load, stress, spike and soak tests while monitoring every component. Fix the largest bottleneck and repeat. Test early, keep the scripts, and plan what the system does when the limit is passed.

Frequently asked questions

What is load testing?

Load testing means sending a realistic amount of simulated use to a system and measuring how it behaves: how fast it responds, how many requests fail, and how hard its servers work. The purpose is to find the limits and the weak points before real users do, while there is still time to fix them.

How many concurrent users should we test for?

Work it out from expected real use. Take the number of people who will use the system, estimate the share active in the busiest hour, and what they do in that hour. Test at that level, then at about twice it for headroom. A figure chosen without reference to actual usage is not a target.

Can we load test in production?

It can be done with great care, at quiet times, with limits and the ability to stop at once, and never with actions that send real messages or payments. The safer course is a test environment that matches production as closely as possible. Many teams do both: thorough tests in staging and a careful confirmation in production before launch.

Author

Etzal Group Team

Etzal Group Team, helping businesses grow with technology. Contact us.

Related service

Website Development

Fast, SEO-ready websites that turn visitors into customers.