[CD] Prevent unnecessary building during crate publish process

[tjtelan]

Related to: https://github.com/infinyon/fluvio/issues/2590

Our publish process is not efficient. It is driven by these two items.

• A manually curated script of ordered crates.
  • https://github.com/infinyon/fluvio/blob/master/release-scripts/publish-list
• A script that always builds, then attempts to publish all the crates from the list until all the crates in the repo have been published, or the amount of retries is exhausted.
  • https://github.com/infinyon/fluvio/blob/master/release-scripts/publish-crates.sh

The complexity comes from the latency from when a new crate is published to crates.io, and the when it is made available. This delay causes publish of crates further down the list to fail due to their dependencies not being available.

(The delay is at least a minute, but I'm not quite sure what it is anymore at the time of this writing.)

Typically the issue resolves itself when the build cycle loops back from the beginning since the builds provide further delay for crates.io to refresh.

But a consequence is when there is an issue that won't resolve by waiting (like https://github.com/infinyon/fluvio/issues/2550), it will take a long time to fail (>1 hr) because the script will rebuild all the crates 3 times before failing.

---

We want this process to fail fast if we're aware of the conditions. And to minimize the building of crates if we know the publish is going to fail because no publish is required.

Extra points if we can accommodate the crates.io refresh delay before attempting to push crates with dependencies that have been updated.

[mfdorst]

Do you think this is also a candidate for rewriting in Rust, or do you think it makes more sense as a shell script?

[tjtelan]

Do you think this is also a candidate for rewriting in Rust, or do you think it makes more sense as a shell script?

The only nice thing about the shell script is not needing to compile it. But I don't have a preference for this being in Rust or shell script.

[mfdorst]

Given that the script will take a lot of time regardless I don't see that as a huge problem. I'll work on it a bit and try to decide which way is better.

[mfdorst]

And to minimize the building of crates if we know the publish is going to fail because no publish is required.

Under what circumstances would no publish be required?

[mfdorst]

Also, do you know of a good way to test this code that doesn't involve actually publishing crates?

[sehz]

I believe you can test publishing by performing dry run

[tjtelan]

And to minimize the building of crates if we know the publish is going to fail because no publish is required.

Under what circumstances would no publish be required?

• If a crate version is already published, and the crate in the repo is the "exact same" as the latest crate available on crates.io.
  • So source code hasn't changed, and no Cargo.toml changes.
• If a crate's Cargo.toml version is 0.0.0 and/or the crate is marked publish = false

And this last case is an error case. The publish is needed, but it will never succeed.

• If a crate has a dependency that has failed, and it is unlikely it will be resolved by trying again.
  • If we're unable to compile
  • If we're unsuccessful with publishing a dependency

Also, do you know of a good way to test this code that doesn't involve actually publishing crates?

cargo publish --dry-run will work for most of the testing.

Otherwise, maybe you can try to publish a crate you're not authorized to publish for testing internal dependencies. (See https://github.com/infinyon/fluvio/issues/2550)

I'm not sure if this is how cargo publish would work, but maybe you can go through all the build process but the last step would fail bc you'd be unable to complete the publish. Though it's a little bit of a hack.

I did the initial testing as part of our typical release, with publishing involved.

[github-actions[bot]]

Stale issue message