Back to all posts
Quality7 min read·October 25, 2026

How to Plan Load Testing Before a Software Launch

TB
ThynkBlox Team
Engineering

To plan load testing before launch, choose the customer workflows that matter, define the workload they must handle and agree on acceptable response times, failures and recovery. Run those workflows with representative data while measuring both customer outcomes and system behaviour. Use the results to support a launch decision within the conditions tested, with clear limits on what remains unproven.

What should a load test prove?

Write a testable statement before choosing a tool or starting a run.

“Support launch traffic” leaves too much open. Specify what people will do, how much activity to expect, how long it will last and what counts as acceptable.

For example: “Customers must be able to browse, search and submit an order throughout the planned launch window, within the agreed response-time and failure limits.”

Make correctness part of that statement. A quick response is insufficient if the order disappears, appears twice or never reaches its next processing step.

Treat load testing as one part of software testing and quality assurance. Keep functional, security and usability acceptance criteria alongside the performance plan.

Which workflows belong in the test?

Choose workflows according to customer importance and expected launch activity. Include the actions that complete a business outcome, rather than testing only individual pages.

For an ordering product, a useful selection might include:

  • Signing in and opening an account.
  • Browsing and searching the catalogue.
  • Viewing a product and checking availability.
  • Submitting an order and viewing its confirmation.
  • Completing background work needed to make the order usable.

Define the journey through completion. If submission starts background processing, measure how long that processing takes as well as how quickly the submission is acknowledged.

Include different account types, permissions and realistic pauses. Have the product owner review the scripts: they should describe recognisable customer behaviour.

How do you define concurrent activity and traffic?

Concurrent activity describes work happening at the same time. An arrival rate describes how often new work begins. Record both when they matter to the launch scenario.

“Concurrent users” needs a precise definition. Does it mean signed-in sessions, people actively navigating or requests currently being processed? A signed-in customer reading a page creates a different workload from someone repeatedly submitting searches.

Specify:

  • How new sessions or journeys begin.
  • What each session does and how often it pauses.
  • How activity is divided between workflows.
  • Whether arrivals continue when responses slow down.
  • How traffic increases, holds steady and falls away.

That fourth point matters to the interpretation. If the script waits for each journey to finish before starting another, inspect whether it still generates the planned arrival rate during slow periods.

Record intended and achieved traffic. A run that fails to produce its stated workload cannot establish that the application passed at that workload.

What could a launch workload look like?

The following numbers are illustrative planning choices, not recommended targets or measured results.

Part of the planIllustrative definition
Arrival patternStart 30 new customer journeys each minute
Journey mixFor every 10 journeys, use 6 browsing, 3 searching and 1 ordering
Concurrent activityRecord active journeys and requests in progress throughout
Traffic shapeIncrease gradually, hold for 40 minutes, then reduce arrivals
DataUse 50,000 products and accounts with varied order histories
Response limitAt least 95 out of every 100 searches finish within 2 seconds
Correctness limitNo duplicate orders or acknowledged orders that disappear
Recovery limitOutstanding background work clears within 10 minutes

Explain where the real targets come from: campaign plans, existing usage, customer commitments or explicit assumptions. Test a plausible expected case and a separate higher-demand case.

Keep the scenarios distinct so the launch decision can say which workload passed.

How realistic should the test data be?

Use data that represents the size, variety and relationships expected at launch. Specify more than a total record count.

Consider catalogue size, account history, attachment sizes, common searches, rare searches and the largest supported customer account. Include popular records that many users access together, alongside journeys that use different records.

Set up enough independent accounts and records for the planned activity. Decide which actions may share data and which must remain separate.

Use synthetic or appropriately sanitised data. Keep personal information and live credentials out of test scripts and reports.

Document the starting state before each comparison. If one run begins with an empty order history and another with substantial accumulated data, explain that difference before attributing changed results to a code change.

Which limits should you agree before testing?

Agree on limits that describe customer experience and business outcomes.

Measure response time for each important workflow. Include a limit for slower responses, such as the time within which 95 out of every 100 actions must finish. Define exactly where timing starts and stops.

Also specify:

  • Failures: Which failed actions count, and how retries are reported.
  • Correctness: Which records and outcomes must be checked.
  • Background completion: How long accepted work may remain unfinished.
  • Recovery: How quickly performance and outstanding work must return to agreed levels.
  • Cost: What spending limit applies to the run and the tested operating conditions.

Set stop conditions before execution. Stop for data corruption, unexpected effects on real users, unauthorised external traffic or spending beyond the agreed limit.

How do you find the bottleneck?

Collect customer-facing results and supporting measurements over the same time window.

When a workflow slows down, inspect where its time is spent: application processing, database work, waiting for an available connection, background queues or an external dependency.

Check the test generator too. Confirm that it delivered the intended workload and had enough resources to run the scripts.

Treat each explanation as a hypothesis. For example, if searches slow while database waiting increases, investigate the relevant queries before deciding that more application servers will solve the problem.

Change one suspected cause at a time where practical, then repeat the comparable scenario. Preserve the workload, data conditions and measurements needed to judge whether the change helped.

For broader design questions, see how to build scalable and robust architectures.

How should third-party services be handled?

List every external dependency used by the tested journeys, including payments, messaging, identity and document processing.

Before sending test traffic, confirm the provider's permitted testing arrangements, applicable limits, charges and approved test facilities. Do not assume permission to run a high-volume test against a live service.

Where direct testing is restricted, use a controlled substitute with explicit response times, failures and limits. Label those assumptions in the report.

Separate what each run establishes. A substitute can support a test of your application's handling of an assumed dependency response. It cannot establish the provider's actual production performance.

Include agreed scenarios for slow replies, rejected requests and temporary unavailability. Check the resulting customer messages, retries and unfinished work.

What common mistakes weaken the result?

Testing only easy reads. Include the important writes and background completion steps.

Reporting averages alone. Show slower responses and results for each critical workflow.

Ignoring retries. Report the original failed attempt as well as the eventual outcome.

Changing the target after the run. Agree acceptance limits first. Record any later revision and its business reason.

Ending measurement with traffic generation. Continue observing recovery and outstanding work.

Using a short run for a long launch window. Match the duration to the question you need answered, and state what remains untested.

Treating a quiet launch as an excuse to skip planning. A small product may need a modest test. Define what that test must establish.

How much production capacity does a passing test prove?

Report the result within its tested conditions: workload, data, software version, resource settings, dependency behaviour and duration.

Do not turn “this scenario passed” into “the product supports this many customers.” Registered customers, active sessions, journey arrivals and completed transactions describe different things.

Document differences between the test setting and production. Treat any capacity estimate beyond the measured workload as an assumption requiring validation.

Use this final checklist:

  • Critical workflows passed their agreed limits.
  • The intended workload was achieved.
  • Data correctness and background completion were checked.
  • Dependency restrictions and substitutions were recorded.
  • Recovery was observed.
  • Untested conditions and launch monitoring were documented.
  • The launch decision names the workload the evidence supports.

Frequently asked questions

When should load testing start?

Start planning when launch workflows and demand assumptions are clear enough to describe. Leave time for investigation, fixes and repeat runs before the launch decision.

How many concurrent users should we test?

Derive the target from expected behaviour. Define active sessions, journey arrivals and pauses first, then check the concurrent activity the test actually produces.

Should we test beyond expected demand?

Include a separate higher-demand scenario when you need to understand behaviour beyond the launch target. Agree its safety limits and purpose in advance.

Does passing mean launch performance is guaranteed?

No. It establishes a result for the conditions tested. Keep production monitoring, response ownership and a plan for reducing demand or limiting affected features ready for launch.

Once the software is live, see what to monitor after launch.


*Need help defining what your launch should withstand? Start a conversation →*

Ready to build?

Let's turn these ideas into your next product.

Start your project