Ferien-Karlsruhe: A Rocky Start on Feb 23

The sales launch was exactly at 5 PM. We knew that, and so did the public. So why should anything go wrong?

It’s not that simple. In fact, it was clear that there would be a rush for the Ferien Portal’s offers. But what does “rush” mean in numbers? How do you dimension server capacities?

SaaS Web has hosted the Ferien Portal since 2015, after the sales launch with the previous provider ended with a “pulled plug.” Since then, visitor numbers and server load have been tracked by SaaS Web, ensuring the portal runs smoothly. How could it happen that the servers were overloaded, even though SaaS Web had managed two sales launches (Spring 2016 and 2017)?

Unlike previous years, when the sales launch began offline and online bookings were only possible a few days later, this time it was entirely online. Consequently, there was no empirical data on user numbers and order frequency. But that’s not all: the portal was completely reprogrammed, meaning each visit, each order, or reservation generated different load values than previously known. Load tests were conducted before the launch, giving us an idea of the capacity the servers should handle. Since these tests did not simulate real ordering processes, the assumption was a rough estimate. It should be noted that thorough load tests involve significant effort and corresponding costs.

Chronology in Summary:

• Shortly before 5 PM, more and more visitors arrive at ferien-karlsruhe.de, but cannot yet order any of the new offers. Immediately before 5 PM, there are approximately 2,500 visitors.
• At 5 PM, the ability to order is automatically enabled.
• Under the load of many orders (as these generate particularly high load), actions such as clicking “Continue” in the shopping cart become impossible. Because the servers react particularly slowly, confusing side effects also become visible, such as a shopping cart icon at the top showing a value, but the shopping cart itself reports being empty.
• During this time, we are monitoring the servers and identifying exactly what is not working as it should. We eventually increase the performance of all components: Load Balancer 1 (out of 2), Web Servers 1, 2, 3, and 4 (yes, there were 4 application servers running), as well as Database Server 1 (out of 3).
• From approximately 6 PM, the situation eases, and reservations flow into the system – by 7 PM, there are approximately 1,200 reservations.