https://pantsbuild.org/ logo
#development
Title
# development
w

witty-crayon-22786

03/14/2023, 10:43 PM
related to the above thread, i’ve been investigating whether to proceed with switching to symlinks in https://github.com/pantsbuild/pants/pull/18153, based on a discussion with @enough-analyst-54434:
have created an issue here discussing that: https://github.com/pantsbuild/pants/issues/18507
will post benchmarks when / if i’ve reproduced them in a single
hyperfine
run, but possibly useful already. cc @bitter-ability-32190
e

enough-analyst-54434

03/14/2023, 10:46 PM
What does tmp have to do with anything other than we currently use it? We could be using some dir guaranteed to be file system coresident for hard linking success IIUC.
w

witty-crayon-22786

03/14/2023, 10:48 PM
yea, that’s the suggestion on the issue. a persistent directory is faster than
/tmp
(for an unknown reason) but slower than
tmpfs
e

enough-analyst-54434

03/14/2023, 10:48 PM
All the tmp talk seems confusing.
w

witty-crayon-22786

03/14/2023, 10:48 PM
the relevance is primarily: “we are currently using `/tmp`… and for some users, that might actually have been a `tmpfs`”
b

bitter-ability-32190

03/14/2023, 10:48 PM
One additional metric (for me at least) is my homedir is encrypted. So possibly placing all of the above outside would net me some gain? Another dimension to test for sure
e

enough-analyst-54434

03/14/2023, 10:49 PM
Filesystems and paths aside, it just seems symlink vs hardlink is totally separable from fs perf.
Choosing symlink makes an extra decision for you in Windows. You now force developer mode
Which sounds reasonable on paper, but we have at least one user who can't use that.
w

witty-crayon-22786

03/14/2023, 10:53 PM
yea, i’m skittish about symlinks in general, since they are also much more visible. it’s possible that even for a fallback we will want to copy, to avoid rules needing to care (mentioned on the ticket)