Performance benchmarks We tried to break it. Here is what happened.

A store that is fast when it is quiet proves nothing. We load CSSorcery Commerce until it is overwhelmed, pull the plug in the middle of payments, and check that every order and every cent is still right.

  • 50 /s

    orders placed per second

    with 95% of checkouts finishing in under one second

  • 6,356 /s

    storefront requests served

    95% of them answered within 23 milliseconds

  • 0

    duplicate orders

    from 705 checkouts that were each retried 15 times

  • 18,000 +

    automated backend tests

    plus more than 10,000 frontend tests and 809 browser tests

How to read these numbers

Measured in November 2026 on one developer laptop (Intel Core i7-7700HQ, 4 cores, 16 GB of memory) that ran the application, the database and the load generator together.

Payment gateways and couriers were replaced by local stand-ins, on scratch databases. Treat these figures as what that one laptop managed, not as a guarantee for your hosting. A proper server does better.

Speed

Fast for shoppers, fast at the checkout

Public pages are cached and compressed. Orders are written in one short, safe step.

  • Checkout throughput

    50 orders a second, with 95% completing in under a second.

    Adding to the basket and placing the order, measured on one laptop with SQL Server.

  • Browsing speed

    6,356 storefront requests a second, 95% answered within 23 ms.

    A mix of public shop pages, while 6,400 requests a second were being sent.

  • Small pages

    The data for a page of 24 products is sent as 869 bytes instead of 32,972.

    Answers that are asked for often are kept ready, already compressed.

  • Two databases

    SQL Server and PostgreSQL both pass the same end-to-end run and all 53 integrity checks.

    A fresh database on each, same code.

Overload

When too many people arrive at once

A promotion, a product drop or the last minute of an auction can bring thirty times the normal traffic. The platform is built to stay correct, serve who it can and turn the rest away quickly and politely.

  • Pushed past its limit

    At 9,600 requests a second, 8,737 were served and the rest were refused within 50 ms.

    Excess traffic gets a fast "busy" answer instead of a hung page.

  • Checkout waiting room

    All 4,898 of 4,898 shoppers placed their order. None gave up, expired or saw an error.

    A surge of 90 shoppers a second against a single processor core.

  • Impatient shoppers

    705 checkouts, each retried 15 times, produced zero duplicate orders.

    Identical requests sent over and over, as happens on a slow connection.

  • A sell-out rush

    A product with 300 units sold exactly 300. Not one more.

    30 shoppers a second all buying the same item.

  • Auctions closing together

    50 auctions closing at the same moment all closed within 6 seconds, each to the highest bidder.

    While 15,000 page reads a second were hitting one processor core.

  • Every storm, same result

    No server errors, no lost or duplicated orders, no overselling, and every payment applied once.

    Across all the overload tests, on both databases.

Failure drills

When things break on purpose

Servers crash and services go down. What matters is that no order, payment or email is lost when they do.

  • Server killed mid-payment

    After four hard kills under load, all 1,540 payments matched exactly one paid order.

    Each restart took under 15 seconds, and the recovery jobs settled whatever was in flight.

  • Mail server down

    2,700 orders went through with zero failures, and all 2,508 queued emails went out 106 seconds after mail returned.

    Mail was unreachable for three minutes.

  • Database slowed down

    No server errors. Excess requests got a quick "busy" answer and everything was normal the moment it cleared.

    Busy rows locked for a full minute.

  • Forged payment notices

    Zero forged notices accepted, and every real order was paid exactly once.

    A flood of payment notices where 20% carried a bad signature and 40% were duplicates.

Engineering quality

Tested like software that handles money

Speed is worthless if a change next month breaks the checkout. This is what guards against that.

  • 18,181

    backend tests

    unit, integration and architecture tests

  • 10,361

    frontend unit tests

    across 735 files

  • 809

    browser tests

    including accessibility checks

  • 85.79 %

    coverage of money-handling code

    held above an enforced 75% floor

  • The release pipeline runs a secret scan, security analysis, dependency checks and the full test suites on every change.
  • About 95 sets of architecture rules are checked automatically, so the structure of the code cannot quietly decay.
  • Money-handling code and sign-in code each have their own enforced test-coverage floor.
  • The same tests run against SQL Server and PostgreSQL.

Planning something big?

Tell us about your busiest day. We will show you how the platform handles it and what hosting it needs.