Spike: performance impact of "read only layers"

[natalieparellano]

Description

See comment: buildpacks/rfcs#155 (comment)

We would like to understand the performance impact of the proposal in order to inform the RFC discussion.

[mboldt]

I can dive in to this.

[mboldt]

I've done an initial experiment and captured it here: https://github.com/mboldt/lifecycle-703-read-only-layers

In short, hashing the layers directories after each buildpack runs doubled the (cached) build time for spring-petclinic-rest with the paketo buildpacks. This seems unacceptable.

More details in the experiment repo.

I propose giving a file system watcher a try as an alternate approach, to see if that might be less costly.

[mboldt]

Implemented the file system watcher, and it is much more efficient. I've updated my experiment repo with this approach and the results.

I think this shows a performant approach to detect violation of the buildpack spec where two buildpacks modify the same layer, and therefore can close out this issue/spike.

[natalieparellano]

Amazing work @mboldt ! buildpacks/rfcs#155 has all the context for anyone looking to continue the conversation.