Is code written on a Friday buggier? Not in the Linux kernel

What happened

“Don’t program on Fridays” is the title of a 2005 paper by Śliwerski, Zimmermann and Zeller, and it has outlived the paper. So I tested it on the largest public record I could read: the Linux kernel, whose authors mark a repair with a line naming the commit it repairs, Fixes: <sha>. Nobody has to guess which change caused which bug, because the person who fixed it said so.

I sealed the rule before any commit reached this server. Each commit keeps the weekday of its author date in the author’s own time zone. A commit counts as needing a fix if a later commit names it within two years. Companies and hobbyists work different days and write different code, so the comparison runs inside one author in one year. That leaves the same person’s Friday against their Monday to Thursday.

Friday sits among the ordinary days, and the bar was 1.10. Plain rates tell the same story: 6.79 % of Friday commits needed a fix, and the four other working days ran from 6.74 % to 6.97 %. When I planted a 25 % rise in the Friday rate, the same method caught it 10 times in 10.

The source of the claim

The 2005 paper’s table for Eclipse has fix-inducing changes at 12.2 % on Friday, 12.1 % on Thursday, 11.1 % on Wednesday and 10.4 % on Tuesday, with an average of 11.4 %. Friday is the highest, as the paper says, by a tenth of a point over Thursday. The paper’s own words call this “a first investigation”. The rest of the distance was covered by the title.

I am also not the first to find this. Eyolfson, Tan and Lam, on the Linux kernel, PostgreSQL and the Xorg server, reported that day-of-week differences are “small and inconsistent across projects”. On Linux they found “about the same” bugginess each day, lowest on Sunday and highest on Monday. They inferred buggy commits by another method, on the kernel’s history from 2005 to 2011, so this is a second route to nearly the same place, inside one author.

What the comparison inside one author did

The weekend shows why it matters. In plain rates, Saturday and Sunday commits needed a fix only 0.83 times as often as working-day ones, which looks like a finding about calm weekends. Inside one author and year the figure is 0.97. Most of that gap was who works weekends, not what weekends do. This is printed under the seal and is not a test.

The same paper by Eyolfson and colleagues found commits between midnight and 4 a.m. buggier, and that one I cannot confirm or kill. Hours 00 to 04 against 09 to 16 came out at 1.03, and an interval I added after the verdict runs from 0.97 to 1.10. It would need its own sealed rule.

What this does not show

• A Fixes line records only bugs that somebody traced and tagged. Bugs nobody found, or fixed without naming the cause, count as nothing, and the kernel’s rate of 6.8 % is far below what those authors measured another way (about a quarter of commits).
• The author date is roughly when the patch was first written. Kernel patches then go through review before they reach master, so a Friday slip may be caught there. This says nothing about merging or deploying on a Friday, which is the form of the claim most teams care about.
• I did not control for the size of a commit. The clone holds commit messages without file contents.
• This is one project. Eclipse in 2005 was another world, and day-of-week may differ elsewhere.

The rule, the code and every number are in the pack.