How I think about log levels

I recently had a moment to think about how I use log levels. I've used logging in various frameworks and standard libraries across many languages. While some levels are standard, not all tools expose the same levels and not all tools have all of the levels I've come to want.

Level	Meaning	Alert	Enabled in Production
trace	Emit practically every variable in scope or "got here" type messages	Never	Only in madness
debug	Emit critical variables or function entry/return	Never	When unable to reproduce behavior in lower environments
info	Critical section entry/exit, expensive section input parameter or return inspection	No	Yes for you code, No for dependencies
warning	Recoverable, accounted-for problems that could be expensive en masse	Yes, when exceeding a frequency threshold, a human should intervene at their convenience	Yes, except to silence an alert when it's been triaged and cannot be immediately fixed
error	Unaccounted-for problems or critical dependency failures, but not grave enough to halt	Yes, a human should intervene ASAP	Of course
fatal OR critical	Unaccounted-for, unexpected failures that must stop the process emitted moments before exit	Absolutely, a human must intervene before the process can restart	Absolutely

I've generally colored them with ANSI colors in log output like this:

• fatal/critical: ⚫🟥 black text on red background
• error: 🔴⬛ red text on black background
• warning: 🟡 yellow or gold text on a transparent background
• info: 🔵 blue text on a transparent background
• debug: 🟢 green text on a transparent background
• trace: 🟠 orange text on a transparent background

These colors work for my non-colorblind eyes. There are color schemes that are probably better to account for those for whom my colors aren't differentiable, so think of them and comment if you have one you like.

I'm doing mostly data pipelines for the last few years but even in my web service days, these were the levels we wanted and rules we followed. One team even looked into aspect-oriented programming in order to get the level of granularity we wanted for trace logging once.

An earlier version was originally posted here.