A probe that did not answer is not a no
We asked whether one page’s burst of calls queues inside the server. The probe landed on five different machines. Inconclusive was recorded as inconclusive.
The back office's Menu screen fires eight calls when it opens. The speed board showed its slowest call taking far longer than the same call takes on its own, and one explanation would be that the eight were queueing behind each other inside the server. If they were, the fix is to combine them into fewer calls. If they were not, the fix is somewhere else. So we tried to measure it.
The probe
Six calls to a route known to make one database hop, fired at once, then six fired one after another, reading the server's own timing header on each. If the concurrent six read slower than the serial six, the server is queueing them. Simple enough.
Except the six concurrent calls left our test machine through a proxy, and the proxy spread them across five different data centres on two continents. Two landed in London, one in Paris, one in Frankfurt, one in Atlanta, one in New Jersey. They never shared a server, so they could not have queued on one. A browser's eight calls all go to the one data centre nearest the person, which is exactly the situation we could not reproduce. The European calls read the same timing either way, which tells you nothing about the question, because the question was never asked.
The temptation
The easy sentence to write at that point is: no queueing observed. It is true, in the way that a locked door you never tried is unopened. Written into the notes it would become a reason the next person trusts, and the next time the Menu screen came up on the board they would skip the combining fix because somebody had checked. Nobody had.
So the notes say what happened: the probe was inconclusive on the question asked, and why, and that the test machine cannot stand where the customer stands. The second measurement that morning, counting the statements each of the screen's routes issues, did answer something. One route was four reads in a line, every one keyed on the venue, none needing another. That was the fix shipped that day. The queueing question stayed open, and is written down as open.
Why this matters past one screen
A measurement has three outcomes, not two. It can say yes, it can say no, and it can fail to ask. The third one is the dangerous one to record, because it looks like a no and is read as one. In a kitchen the same thing happens when a dish is marked as checked for allergens by someone who could not find the recipe. The honest record is that nobody checked. The same discipline applies to a line in an engineering log.