The drawer that says it saved

A slow device and a fast chip sharing one set of wires. This lesson follows a save down them, and asks how far it gets before the program hears “done”.

Before you start. A file is only bytes plus an agreement about how to read them (Everything Is a Number (bits & bytes)), and a program never touches the disk itself: it asks the kernel, the one program with full access to the hardware (Processes, Threads & the Kernel). The disk is slow, so the kernel starts a job on it and does other work until the disk signals it is done (I/O, Interrupts & the Bus).

The idea in 30 seconds

A file is three separate things on disk: its bytes, a numbered record of its size and location called an inode, and a name in a directory’s list that points at that number.
Two names can point at one inode, the record that tracks a file’s bytes. That is a hard link, and deleting one name leaves the file readable through the other.
When a program saves, the kernel copies the bytes into RAM (the page cache) and reports success before the disk has anything. A power cut in that gap loses them.
fsync makes a program wait until the file is on disk. Measured on one Mac, it made saves ten times slower; Apple’s stronger version, which also empties the drive’s memory, 1,400 times slower than not waiting.
A journal writes down each change before making it, so a crash leaves each change done or not done, never half. It cannot save data still in RAM.
The safe way to replace a file is: write a new copy, fsync it, then rename it over the old name.

Read the story · Skip to the numbers