Average vs Median Google Search Latency: Reasoning About a Skewed Distribution

Consider the response time of searches on Google: how long it takes from the moment a query is submitted until its results come back. Looking across a very large number of searches, how would the mean response time compare with the median response time? State which one you expect to be larger, and justify your answer from what makes some searches fast and others slow. ```hint Look at both ends Is there a limit on how fast a search can possibly return? Is there a comparable limit on how slow one can get? Think about what that asymmetry does to each statistic. ``` ### Clarifying Questions - Is response time measured inside the search backend, or end to end as the person experiences it, including their network connection and page rendering? - Does the data include searches that time out or fail, or only searches that returned results? - Is the comparison over all searches worldwide, or within one region, device type or kind of query? ### What a Strong Answer Covers - A clear prediction of which statistic is larger, tied to the shape and direction of skew of the response-time distribution. - Concrete reasons why most searches cluster near a fast typical time while a minority take much longer. - Why a rare slow search moves the mean but barely moves the median, ideally shown with a small numeric example. - How the claim would be checked against real latency data, and the conditions under which the gap would shrink or even flip. - What the gap implies for which latency numbers a search team should track and set goals on. ### Follow-up Questions - After a release, median search latency is unchanged but the mean rose noticeably. What could explain that, and does it matter? - Why might a team set its latency target on the 99th percentile instead of the mean or the median? - Each search fans out to many backend servers and waits for all of them to answer. How does that affect the slow end of the distribution?