go-mysql-server can no longer claim to be "pure Go"

[srabraham]

The repo description and README both say that this package is "pure Go". That's not true anymore, because #3220 forces the use of cgo to use this library. That's a significant breaking change for anyone using dolthub/go-mysql-server directly.

I'm not sure of exactly the reasoning behind the big change in https://github.com/dolthub/go-icu-regex, but if it were me running the show, I'd find a way for github.com/dolthub/go-mysql-server to be built selectively with either the cgo or non-cgo version of https://github.com/dolthub/go-icu-regex. Go offers build tags that allow that sort of thing.

[reltuk]

Unfortunately an issue with one or more of:

1. wazero
2. go-icu-regex's build of icu for wasm
3. go-icu-regex's integration with wazero
4. recent changes to the go runtime

meant that we had a very hard to root cause customer breaking issue that resulted in a hard stall of the entire go runtime, since calls into wazero's JIT compiled functions are not preemptible.

We are happy to accept to PRs that disable ICU4C regex support in go-mysql-server when CGO is not available here, or which convincingly fix the root cause of the wazero/go-icu-regex misbehavior. But we are unlikely to revert this change entirely, since this change makes regexes significantly faster compared to the wazero bridge and it makes tooling around their implementation, like debuggers, much more ergonomic.

Given the choice between:

1. Adding a cgo dependency
2. Removing the functionality
3. Leaving a latent bug which completely stalled the calling process

We chose 1.

[zachmu]

Another option would be alternate implementations of the regexp_like and related expressions that use the go native regex libraries. This would mean the behavior of those expressions differs from MySQL in some cases, but that could be a reasonable tradeoff for many customers using the library. We would be happy to consider a patch for this option as well.

The only option not on the table would be a return to using the wazero implementation of go-icu-regex, which causes runtime hangs. The risk there is just too great for our production customers.

[reltuk]

Relevant PR updating the readme: #3223

[srabraham]

Thanks, folks. I appreciate the conversation here!

[reltuk]

Thanks, folks. I appreciate the conversation here!

Thanks for the feedback. We appreciate it. We do realize this is a difficult change for our consumers to update their processes for. It's helpful for us to see an immediate reaction like this as well. We ourselves have many upstreams of go-mysql-server – updating CI processes, container environments, toolchains, etc., to all be compatible with this new Cgo dependency has been a non-trivial process.

I hope we can drive towards something that works for our existing use cases but which is also not too disruptive for people that want to stick to pure Go or want to take on the new Cgo dependency incrementally or only in certain contexts.