When playing the optimization game, defense matters
Experimentation is an indispensable process for optimizing the metrics that matter to your business. Sometimes this process can be deceptively simple: formulate a good set of hypotheses, start A/B testing them and you will almost certainly see your metrics move in the right direction over time. What can possibly go wrong?
Not so fast. Other teams in your organization are undoubtedly busy testing their own ideas. There is a good chance they don’t share the same goals as you. So while you are on the offensive, new features can come in at any moment and adversely impact the metrics you tirelessly try to improve.
Time to think about your defense strategy…
Case in point: Bing Performance
I faced this problem when I joined Microsoft 8 years ago as a newly-minted program manager for the Bing Performance team. Back then, our search results page was perceivably slower to load than the competition’s. We ran experiments that added artificial delay to the page, which proved that slower page load times (PLT) hurt user engagement and revenue (amounting to millions of dollars lost for every 100 milliseconds or 1/10th of a second added to the page).
An engineer that improves server performance by 10ms (that’s 1/30 of the speed that our eyes blink) more than pays for his (or her) fully-loaded annual costs. Every millisecond counts. – Online Controlled Experiments at Scale, 2013
Our leadership also said publicly that in the world of search engines:
“two hundred fifty milliseconds, either slower or faster, is close to the magic number for competitive advantage”. — for impatient web users, an eye blink is just too long to wait, New York Times, 2012
Consequently, our team was tasked with the urgent mission to be within the blink of an eye to competition (that’s 250 milliseconds give or take). We assembled a formidable group of performance engineers and got to work.
Every millisecond matters… but not for all teams
A few months in, we had shipped a lot of improvements validated by A/B tests. Yet, things were going south (or, to be precise, north, given we were dealing with page load times). Our overall PLT metric continued to move in the wrong direction.
We soon realized why. While we were preoccupied with our work, other teams shipped features that often impacted PLT.
For example, our results relevance team was laser-focused on optimizing clicks on web results. This usually meant running more time intensive ranking models on the server. Similarly, our Multimedia team favored shipping features that added visual richness to the page, which in turn improved engagement on the image and video results. But each of these improvements added more weight to the page, making it slower to load. And the list goes on. Over time, these changes collectively added a considerable penalty to PLT.
We needed a better plan. We turned to A/B testing again, but this time with a defensive mindset.
Fortifying our defenses
Product teams at Bing have an experiment-first philosophy for every big or small feature they want to ship. Conveniently, this allowed us to track features that impacted PLT while they were being A/B tested. We started executing a new strategy that entailed several preventive measures.
1) Experimentation served as an impact prevention mechanism at scale
Every A/B results page included PLT as a guardrail metric. Teams could not ship a feature that hurt PLT without sign-off from the performance team. Experiment alerts also warned us about A/B tests that impacted PLT and allowed us to react with proper corrective action.
2) Performance budgets constrained the impact from other teams
We introduced performance budgets per team. For example, our Multimedia team was allowed to add up to +X milliseconds to the page per release cycle (roughly 6 months in duration). They had to get executive-level approval to go beyond that, or find a way to cut milliseconds elsewhere.
3) Experts and dedicated tools assisted with debugging issues
We assigned experts to help teams mitigate the performance impact of their features. The same folks built new investigation tools that allowed us to pinpoint the culprit code that increased PLT in the confines of an A/B test.
4) Executive Reports created urgency and brought accountability across the organization
Finally, we began sending regular organizational-wide reports that highlighted our progress and often acknowledged the efforts of other teams. The reports reinforced the sense of urgency around performance and promoted cross-team accountability towards meeting our goals.
Ultimately, this defensive strategy proved to be a winning recipe. We exceeded our targets; and this would not have been remotely true had we focused only on shaving milliseconds off the page and not also on preventing teams from impacting PLT.
Takeaways
I focused on web performance in this blog post, but the lessons apply to any optimization goal that is not equally shared by teams in your organization. My main takeaways from this effort:
- A/B testing, particularly when widely adopted within your organization, is an effective method for monitoring at scale the impact of new features or changes on the metrics you care about.
- When necessary, assign team budgets to constrain the impact by other teams.
- Don’t leave other teams helpless. Set them up for success by providing guidance and tools to overcome issues related to your metrics.
- Promote cross-team accountability in meeting your goals. Acknowledge the effort of other teams that are helping you reach your objectives.
Do you have a similar experience to share? Let me know in the comments section!