Pick the encoding before the tool
Why some charts work and others fail is not a question of taste. People judge position along a common scale accurately, length almost as well, and angle, area and colour intensity progressively worse. Every chart type is a wager on one of those channels. A pie chart wagers on angle, asking a reader to compare wedges separated by a few degrees, which is why it fails as soon as there are more than a handful of categories or any two sit close together. The replacement is not a prettier pie. It is a bar chart laid out horizontally and sorted by value, which uses length instead of angle, gives long category names room to be legible, and puts the ranking into the order of the rows where it can be read at a glance. The same reasoning disposes of most of the gallery: donuts, radar plots, anything rendered in three dimensions, and the word cloud, which encodes frequency as area and is decoration wearing the clothes of evidence.
- Comparing categories: horizontal bars, sorted by value rather than alphabetically, unless the categories carry an order of their own such as age bands.
- Change over time: a line, with time running left to right. Begin the axis at zero when the quantity has a meaningful zero; when it does not, truncate openly and say so on the chart.
- Relationship between two measures: a scatter plot. Add a fitted line only if you can defend the model behind it.
- Distribution: a histogram, or a box plot when several groups are being compared. A chart showing only averages has hidden the spread that the argument probably turns on.
- Composition: bars again, or a stacked bar when there are two or three parts and the total matters. Stacked areas with many series are unreadable above the bottom band, because only that band has a flat baseline to measure from.
- Too many series to hold in the head: small multiples. A grid of identical little charts is better than one crowded chart, and better than a second vertical axis every time.
That leaves the question glossaries dodge by retreating into principles: what to learn. Split it by audience. For charts you are drawing in order to think, use Matplotlib inside the notebook where the data already sits, accept that the defaults are plain, and refuse to spend an afternoon styling something you will delete before lunch. For charts other people will open by themselves, learn one business intelligence product properly rather than three of them badly: Power BI if the organisations you are aiming at run on Microsoft, Tableau otherwise, and the second takes about a week once the first is fluent. D3 is not a charting library and should never be chosen as one. It is a programming project for building a bespoke interactive graphic, appropriate when a newsroom would reach for it and almost never when a company reports on itself. Underneath all three sits the same craft: subtraction. Most charts improve when the gridlines fade, the legend is replaced by labels sitting on the lines themselves, and the second idea moves to a chart of its own. A plain graphic with an unbroken axis and a title stating what happened outdoes an elaborate one in any tool you care to name.