| 21 May 2025 |
|  sinan changed their profile picture. | 11:59:58 |
| sinan changed their profile picture. | 12:00:45 |
 emily | is there a way to trace imports of an evaluation that isn't as horribly slow as nix eval -v -v or scopedImport { import = x: builtins.trace x (import x); }? | 13:26:03 |
| emily | actually I guess the latter needs to be recursive which would make it even slower | 13:27:14 |
 Sergei Zimmerman (xokdvium) | In reply to @emilazy:matrix.org
is there a way to trace imports of an evaluation that isn't as horribly slow as nix eval -v -v or scopedImport { import = x: builtins.trace x (import x); }? Hm, do you need the trace to be super precise? Would a flamegraph suffice? | 14:33:28 |
| emily | yes, precision is the aim (dependency tracking) | 14:34:24 |
| emily | though only a subset of paths (beneath a given root) are required | 14:34:39 |
| emily | possibly I can use that to help short-circuit things | 14:35:04 |
 hacker1024 | Is modifying the implementation of `import` in the Nix source itself an option? | 14:36:37 |
| emily | I suppose I'd really rather not :) | 14:46:21 |
| emily | I don't really know why -v -v is slow in the first place | 14:46:35 |
| Sergei Zimmerman (xokdvium) | Can you share a repro for the slowdown? | 14:59:40 |
| emily | just import a fetchTarball of Nixpkgs and do anything with it. but yeah I can put one up in a couple hours | 16:43:13 |
| emily | I'm on Darwin which does have slower FS perf but presumably it shouldn't be amplifying FS operations | 16:43:33 |
| 22 May 2025 |
| emily | Sergei Zimmerman (xokdvium): whoops, sorry for forgetting about this. hm, a minimal reproducer seems way less slower than expected. let me try and see what i was observing | 05:44:49 |
| emily | I can't reproduce. I guess -v -v is great :) | 05:46:34 |
| emily | I wish it would trace other file accesses though | 05:46:45 |
| emily | like builtins.readFile | 05:46:54 |
| emily | does anyone understand why nix eval .?rev=X#foo will sometimes produce "copying …" messages even when the rev did not change? | 09:35:24 |
 Mic92 | https://discourse.nixos.org/t/nix-2-29-0-released/64609 Nix 2.29.0 released | 13:09:10 |
| Mic92 | I think if you do a lot of -v than it will print the flake cache that causes the copying. | 13:10:53 |
| Mic92 | https://github.com/numtide/nix-eval-cache | 13:12:30 |
| Mic92 | currently limited to Linux, but could be ported to darwin as well with the right environment variables for LD_PRELOAD | 13:13:09 |
|  Perchun Pak [don't ping; dm instead] changed their display name from Perchun Pak to Perchun Pak [don't ping; dm instead]. | 13:50:46 |
| emily | the LD_PRELOAD equivalent doesn't work super reliably on Darwin | 13:58:02 |
| emily | though with Nix it might work fine since I guess it probably doesn't go through SIP stuff | 13:58:11 |
| emily | does this handle builtins.readFile? | 13:58:17 |
| emily | -v -v doesn't actually seem to slow things down appreciably after re-testing so it seems likely adequate for my usecase | 13:58:29 |
| emily | (using scopedImport to catch readFile takes 1.5× longer to eval :( ) | 14:00:53 |
| emily | I guess I could PR tracing readFile in -v -v too 🙃 | 14:01:07 |
