Posts tagged "what survives a crash"

In 2009, ext4 users found empty files after crashes, and ext4 got a heuristic called auto_da_alloc. I watched it work on my laptop: a rename over an existing file starts the writeback at once. It protects against the worst case, but it is not durability.

fsync of a new file does not promise that its name is on the disk. I measured directory fsync on ext4: half a microsecond after a file fsync, about 420 microseconds after a rename, and alone it made the name durable while the data stayed in RAM.

I killed a program 300 times while it saved a file. Writing in place left a cut file 233 times, without any power loss. A temp file, fsync and rename never did. Then I traced how sed, Python, cp, install and git save files.

I measured 1,000 rounds of a 4 KiB write followed by a sync on my laptop: about 540 microseconds for an append, 208 for an overwrite in place, a slowest round eleven times the median, and 3.6 microseconds on tmpfs, where fsync waits for nothing.

When write returns, the kernel has your data, but the disk does not have it yet. I measured it on my laptop: 256 MiB sat in RAM until fsync, kill -9 lost nothing, and a power cut could lose everything for up to about 35 seconds.