Estimate the effect of a policy by comparing how a treated group changed against how a similar untreated group changed over the same period.
Store every fact once and reads work harder; copy it where it is read and the copies drift apart. Work a day in the shop and see which bill you would rather pay.
Store each fact once and it cannot contradict itself, but reads pay to reassemble it; copy it everywhere and reads get cheap, but a copy you forget can make the data lie.
Splitting one big table across machines only helps if the partition key spreads the load evenly. Pick a bad key and watch one shard melt.
When two data centres lose their link, a system can stay correct or stay available — not both. Cut the cable and watch each side decide.
When the link between two data centres drops, a system either declines what it can’t verify or acts blind — never both.
The eye groups elements by proximity, similarity, and enclosure before it reads a single label. Rearrange a dashboard and watch the groupings form and break.
In a rainbow colormap, equal steps in data are not equal steps to the eye — some bands vanish, others shout. Compare ramps against a colourblind simulation.
Outliers are points that sit far from the normal pattern — by distance, by density, by deviation. Drop a point on the canvas and see why it is or is not flagged.
A detector never learns what fraud looks like — only what normal looks like, and every line you draw pays for the anomalies it catches with the customers it wrongly stops.
Words become vectors, and directions in that space carry meaning: king − man + woman lands near queen. Do the arithmetic and watch the analogy resolve.
Adding calibrated noise to an answer hides whether any one person is in the data, at a measurable cost to accuracy. Turn the privacy budget and watch the trade.