feat(back): #1101 globbed attrs #1102
Conversation
dsalaza4
commented
Jun 2, 2023
dsalaza4
commented
- Collect attr paths using python globs for initialization performance
- Pass them as attrPaths to evaluator
- Redefine evaluator logic for building outputs from attrs
- Initialization for universe takes less than 3 seconds now
|
@kamadorueda Any architectural comments regarding this? |
|
Instead of concatenating them, I will dump all paths into a json string and then parse them in nix. I don't want to use files to avoid IO |
Not really, I understand the reason for the change. One failure mode to be aware of is that the OS restricts the length of command (between 32kB to 2MB depending on the OS). But at that point, we can switch to a file Another way to optimize with a more aggressive trade-off is using the fact that people use the build system more often than they modify the build system: so we could add a "generate manifests" step after modifications that produces the list of commands (and everything else needed for the CLI), and which gets versioned in the repo. This way the CLI can consume a "manifest" or "lock" file that already contains the list of outputs, instead of computing it every time. It adds one step to modifications, but saves one step for the average user and also enables faster command auto-completion (as stuff is always ready to consume). The other option is making consumers "initialize the project": Something like "makes install", where before using the project you need to install it, and in this installation step we generate some manifests for the CLI that are reused on all future invocations in general, if this PR makes it fast enough, I think that's good enough |
- Collect attr paths using python globs for initialization performance - Pass them as attrPaths to evaluator - Redefine evaluator logic for building outputs from attrs - Initialization for universe takes 1 second now Signed-off-by: Daniel Salazar <[email protected]>
|
Kudos, SonarCloud Quality Gate passed!
|
