Set the load target from the launch model
The launch-day forecast is not your test target. It is the starting point. A platform that handles the expected average can still fail when clients log in together, news drives order activity, or a price feed reconnects and sends a burst of updates. Test against a credible peak, then test what happens when that peak is exceeded.
Build the forecast from the product and its distribution plan. Estimate funded accounts, concurrent sessions, active traders and orders per active trader per minute. These are different measures. A session count says little about order flow if most users are watching prices, while a relatively small group of highly active clients can dominate gateway traffic.
Use operational evidence where it exists: existing products, demo environments, campaign forecasts and the busiest comparable trading period. Separate normal activity from short bursts.
A launch campaign may create a login surge at the open; a new instrument can concentrate orders in a narrow set of symbols. Include both in the model, and record the assumptions so they can be changed as marketing plans or client mix shift.
Also define what "works" means before running a test. Set service-level objectives for order acceptance, risk checks, price updates and account views.
Measure latency at named points, such as gateway receipt to order acknowledgement, rather than using one vague end-to-end figure. The external execution venue, client network and device add delay that your platform cannot fully control.
Build a workload that follows the order path
A useful test drives transactions through the same components as production: authentication, session management, API or FIX gateway, pre-trade risk, order management, routing, execution handling, positions and reporting. If the test stops at the gateway, it will not expose a slow risk service or a database queue that fills after several minutes.
Reproduce the mix of actions, not just a high rate of identical order submissions. Include market and limit orders, amendments, cancellations, rejects, partial fills and position updates. Mix symbols with different tick rates and account types with realistic limits. For a prop firm, represent the evaluation, risk and account-rule checks that apply to each order. For a broker, include the relevant routing paths and the handling of execution reports.
Use a venue simulator or a controlled test environment for external execution. Do not direct a stress test at a live liquidity provider, exchange or market-data source without explicit agreement. Simulate realistic response times, rejects and partial fills. Otherwise, an unrealistically fast mock can hide back-pressure, while an uncontrolled external service can make test results noisy or cause operational risk.
The figures below are typical planning ranges, not universal standards. Select values against your forecast, architecture and release requirements, then document why they fit.
| Test dimension | Typical planning range | What to verify |
|---|---|---|
| Sustained load | 1.2 to 1.5 times forecast peak for 60 to 120 minutes | Queues, resource use and latency remain stable after warm-up |
| Short order burst | 1.5 to 2 times forecast peak for 1 to 5 minutes | Rate limits, gateway behaviour and recovery after the burst |
| Ramp-up | 10 to 30 minutes to reach target load | Whether gradual growth reveals a bottleneck before saturation |
| Soak run | 4 to 8 hours, with longer runs for overnight processes | Memory growth, connection leaks, log volume and scheduled jobs |
| Recovery check | Return to normal load within 5 to 15 minutes after a fault | Backlog drainage, duplicate handling and service health |
Generate unique users, orders and request identifiers at a rate that does not make the load generator itself the bottleneck. Run generators close enough to minimise network noise, but keep their CPU, memory, connection count and output rate under observation.
A useful run records both offered load and completed work. If the generator sends 10,000 requests but the platform only processes 8,000, reporting the sent rate as capacity is misleading.
For further context on platform components and operating models, the Trading Technology infrastructure coverage can help frame what belongs in the test boundary.
Run tests that expose bottlenecks and failure modes
Use a sequence of controlled runs. Change one major variable at a time where possible. Start below the forecast peak to validate the harness and telemetry, then increase concurrency and order rate in steps. A sudden jump straight to a large target can show that something failed, but it gives little evidence about the point at which performance began to degrade.
- Validate the path at low load. Confirm that test accounts, symbols, risk settings and execution simulation behave as expected. Compare counts across the generator, gateway, order system and downstream services so missing or duplicated events are visible.
- Establish a baseline. Record latency percentiles, error and reject rates, queue depth, CPU, memory, database connections and network throughput at a modest load. Capture per-service metrics, not just host averages. Averages can conceal a hot partition or a single overloaded instance.
- Ramp to the sustained target. Increase load in defined steps, for example 25 per cent increments, and hold each step long enough to observe queues and percentiles. Watch p95 and p99 latency alongside throughput. A rising p99 with a flat average is often an early sign that a small share of requests is waiting behind a saturated dependency.
- Apply a burst and a soak. Push the agreed short burst, then hold sustained traffic for several hours. Check whether queues drain after the burst and whether memory, file descriptors, connection pools or storage use continue to climb during the soak.
- Remove one dependency at a time. Test a delayed execution response, a market-data reconnect, a slow database, or a node loss. Keep faults controlled and agreed with operations. Verify that timeouts and circuit breakers prevent a failed dependency from tying up all worker threads.
- Test recovery and replay. Restore the service and check backlog processing, order-state reconciliation and duplicate protection. Confirm that recovery does not create a second surge that exceeds the original peak. Measure time to a stable state, not only time until a health check turns green.
- Repeat after the fix. Keep the same workload and test data so the before-and-after comparison is meaningful. Then run a regression pass against the previous baseline, because a local optimisation can shift load to another service.
Correlate application traces with infrastructure metrics and event timestamps. For each slow order, find where time accumulated: gateway queue, risk evaluation, database call, routing or execution response.
Retain raw results and configuration with the build identifier. A graph without the test version, data set and workload profile is difficult to use during a release decision.
Keep an eye on contention that synthetic tests often miss. Shared databases, caches and message brokers can be busy with administrative tasks, reporting or overnight jobs. If production uses separate pools or partitions, model them accurately. If it does not, include those competing workloads and decide whether the launch requires a maintenance window.
Turn results into a release decision
Capacity is not a single number. Record the highest load at which the system meets its latency and error objectives, then identify the first component that approaches its limit. CPU headroom does not help if a connection pool is exhausted, and a fast order acknowledgement does not prove that executions and account balances are being updated correctly.
Agree release gates with engineering and operations before the final run. A practical set includes:
- All critical order paths complete with reconciled counts and no unexplained state differences.
- Latency percentiles and error rates stay within the product's defined objectives at the sustained target and during the agreed burst.
- Queues return to their normal range after bursts and faults, with no growing backlog during the soak.
- Failure handling, alerting, rate limits and rollback procedures work in a controlled exercise.
- Capacity has a stated buffer above forecast peak, with the assumptions and scaling actions recorded.
Typical headroom targets vary. A team might plan for 20 to 50 per cent above a credible forecast peak, but the right buffer depends on how quickly capacity can be added, how uncertain adoption is, and whether a bottleneck is shared.
A cloud service that can scale predictably may need a different reserve from a fixed-capacity matching or risk stack. Do not treat autoscaling as proof of capacity: verify its trigger delay, warm-up time, quotas and behaviour under a burst.
Make the release decision conditional on operational readiness as well as throughput. Name who watches the launch, which dashboards and alerts they use, when the team will pause new onboarding, and who can disable a product or route. Set rollback triggers in measurable terms, such as sustained order rejects above an agreed rate or p99 latency outside the objective for a defined period. A rehearsed intervention is safer than an informal expectation that someone will notice.
If this were my operation, I would first build a workload model from the launch forecast and the busiest comparable production period, then validate every hop with low-volume transactions. I would not approve the launch until a sustained run, a burst and a recovery test all pass the agreed gates on the release build.
Frequently Asked Questions
How much load should a trading platform handle before launch?
Use the launch forecast as a baseline, then test above it. Typical planning ranges are 1.2 to 1.5 times forecast peak for sustained load and 1.5 to 2 times peak for a short burst, adjusted for uncertainty and available scaling.
How long should a trading platform load test run?
A sustained run of 60 to 120 minutes can expose queue and latency problems, while a 4 to 8 hour soak can reveal leaks and resource growth. Extend the soak if the platform has overnight processing or scheduled workloads that affect production.
What should be included in broker platform load testing?
Test authentication, gateways, pre-trade risk, order management, routing, execution reports, positions and reporting. Include a realistic mix of order types, cancels, amendments, rejects, partial fills and market-data activity.
Can load testing use a live liquidity provider?
Only with explicit agreement and suitable controls. A venue simulator is usually safer for stress and fault testing, provided it models realistic response times, rejects and partial fills.
Platforms · Software · Providers · Markets