Swarovski – the leading producer of cut crystal in the world- relies on its eCommerce store as much like other companies in the highly competitive eCommerce environment. Swarovski’s story is no different from others in this space: They started with “Let’s build a website to sell our products online” a couple of years ago and quickly progressed to “We sell to 60 million annual visitors across 23 countries in 6 languages”. There were bumps along the road and they realized that it takes more than just a bunch of servers and tools to keep the site running.
Why APM and why you do not just need a tool?
Swarovski relies on Intershop’s eCommerce platform and faced several challenges as they rapidly grew. Their challenges required them to apply Application Performance Management (APM) practices to ensure they could fulfill the business requirements to keep pace with customer growth while maintaining an excellent user experience. The most insightful comment I heard was from René Neubacher, Senior eBusiness Technology Consultant at Swarovski: “APM is not just about software. APM is a culture, a mindset and a set of business processes. APM software supports that.”
René recently discussed their Journey to APM, what their initial problems were and what requirements they ended up having on APM and the tools they needed to support their APM strategy. By now they reached the next level of maturity by establishing a Performance Center of Excellence. This allows them to tackle application performance proactively throughout the organization instead of putting out fires reactively in production.
This blog post describes the challenges they faced, the questions that arose and the new generation APM requirements that paved the way forward in their performance journey:
Swarvoski had traditional system monitoring in place on all their systems across their delivery chain including web servers, application servers, SAP, database servers, external systems and the network. Knowing that each individual component is up and running 99.99% of the time is great but no longer sufficient. How might these individual component outages impact the user experience of their online shoppers? WHO is actually responsible for the end user experience and HOW should you monitor the complete delivery chain and not just the individual components? These and other questions came up when the eCommerce site attracted more customers which was quickly followed by more complaints about their user experience:
Questions that had no answers
In addition to “Who is responsible in case users complain?” the other questions that needed to be urgently addressed included:
- How often is the service desk called before IT knows that there is a problem?
- How much time is spent in searching for system errors versus building new features?
- Do we have a process to find the root-cause when a customer reports a problem?
- How do we visualize our services from the customer‘s point of view?
- How much revenue, brand image and productivity are at risk or lost while IT is searching for the problem?
- What to do when someone says ”it‘s slow“?
The 10 Requirements
These unanswered questions triggered the need to move away from traditional system monitoring and develop the requirements for new generation APM and user experience management.
#1: Support State-of-the-Art Architecture
Based on their current system architecture it was clear that Swarovski needed an approach that was able to work in their architecture, now and in the future. The rise of more interactive Web 2.0 and mobile applications had to be factored in to allow monitoring end users from many different devices and regardless of whether they used a web application or mobile native application as their access point.
#2: 100% transactions and clicks – No Averages
Based on their experience, Swarovski knew that looking at average values or sampled data would not be helpful when customers complained about bad performance. Responding to a customer complaint with “Our average user has no problem right now – sorry for your inconvenience” is not what you want your helpdesk engineers to use as a standard phrase. Averages or sampling also hides the real problems you have in your system. Check out the blog post Why Averages Suck by Michael Kopp for more detail.
#3: Business Visibility
As the business had a growing interest in the success of the eCommerce platform, IT had to demonstrate to the business what it took to fulfill their requirements and how business requirements are impacted by the availability or the lack of investment in the application delivery chain.
#4: Impact of 3rd Parties and CDNs
It was important to not only track transactions involving their own Data Center but ALL user interactions with their web site – even those delivered through CDNs or 3rd parties. All of these interactions make up the user experience and therefore ALL of it needs to be analyzed.
#5: Across the lifecycle – supporting collaboration and tearing down silos
The APM initiative was started because Swarovski reacted to problems happening in production. Fixing these problems in production is only the first step. Their ultimate goal is to become pro-active by finding and fixing problems in development or testing—before they spill over into production. Instead of relying on different sets of tools with different capabilities, the requirement is to use one single solution that is designed to be used across the application lifecycle (Developer Workstation, Continuous Integration, Testing, Staging and Production). It will make it easier to share application performance data between lifecycle stages allowing individuals to not only easily look at data from other stages but also compare data to verify impact and behavior of code changes between version updates.
#6: Down to the source code
In order to speed up problem resolution Swarovski’s operations and development teams require as much code-level insight as possible — not only for their own engineers who are extending the Intershop eCommerce Platform but also for Intershop to improve their product. Knowing what part of the application code is not performing well with which input parameters or under which specific load on the system eliminates tedious reproduction of the problem. The requirement is to lower the Mean Time To Repair (MTTR) from as much as several days down to only a couple of hours.
#7: Zero/Acceptable overhead
“Who are we kidding? There is nothing like zero overhead especially when you need 100% coverage!” – Just the words from René when you explained that requirement. And he is right: once you start collecting information from a production system you add a certain amount of overhead. A better term for this would be “imperceptible overhead” – overhead that’s so small, you don’t notice it.
What is the exact number? It depends on your business and your users. The number should be worked out from the impact on the end user experience, rather than additional CPU, memory or network bandwidth required in the data center. Swarovski knew they had to achieve less than 2% overhead on page load times in production, as anything more would have hurt their business; and that’s what they achieved.
#8: Centralized data collection and administration
Running a distributed eCommerce application that gets potentially extended to additional geographical locations requires an APM system with a centralized data collection and administration option. It is not feasible to collect different types of performance information from different systems, servers or even data centers. It would either require multiple different analysis tools or data transformation to a single format to use it for proper analysis.
Instead of this approach, a single unified APM system was required by Swarovski. Central administration is equally important as they need to eliminate the need to rely on remote IT administrators to make changes to the monitored system, for example, simple tasks such as changing the level of captured data or upgrading to a new version.
#9: Auto-Adapting Instrumentation without digging through code
As the majority of the application code is not developed in-house but provided by Intershop, it is mandatory to get insight into the application without doing any manual code changes. The APM system must auto-adapt to changes so that no manual configuration change is necessary when a new version of the application is deployed.
This means Swarovski can focus on making their applications positively contribute to business outcomes; rather than spend time maintaining IT systems.
#10: Ability to extend
Their application is an always growing an ever-changing IT environment. Where everything might have been deployed on physical boxes it might be moved to virtualized environments or even into a public cloud environment.
Whatever the extension may be – the APM solution must be able to adapt to these changes and also be extensible to consume new types of data sources, e.g., performance metrics from Amazon Cloud Services or VMware, Cassandra or other Big Data Solutions or even extend to legacy Mainframe applications and then bring these metrics into the centralized data repository and provide new insights into the application’s performance.
The Solution and the Way Forward
Needless to say that Swarovski took the first step in implementing APM as a new process and mindset in their organization. They are now in the next phase of implementing a Performance Center of Excellence. This allows them moving from Reactive Performance Troubleshooting to Proactive Performance Prevention.
Stay tuned for more blog posts on the Performance Center of Excellence and how you can build one in your own organization. The key message is that it is not about just using a bunch of tools. It is about living and breathing performance throughout the organization. If you are interested in this check out the blogs by Steve Wilson: Proactive vs Reactive: How to prevent problems instead of fixing them faster and Performance in Development is the Chief Cornerstone.