Support FrankenPHP

[artonge]

Current limitations

Critical - For a safe and working worker mode

• Remove static variables
  • Task: Introduce official linting rules for Nextcloud
  • Task: Introduce reflection at container destroy time to detect static variables in classes and fail hard on worker mode production
• Detect open DB transactions at the end of a request
  • Check for open transactions. If any, error and roll it back.
• Smalls adaptation to how Frankenphp is working

Nice to have

• For a debuggable software: check how to use xdebug and blackfire with FrankenPHP
• For a working worker mode after code changes: Reload on code changes (app install/update, config.php)
• To progressively improve the performance: Opt-in/out classes from the per request global reset
  • Task: Introduce an attribute to opt-out classes
  • Task: Introduce "magic" detection if class's constructor argument is bound to the request context (user, session, userFolder, request, response, ...?)

References

Issues

• Add support for RoadRunner PHP webserver #36290

PRs

• feat: frankenphp #58541

cc @CarlSchwan @come-nc @Altahrim @provokateurin for reference. Feel free to amend as you whish.

[come-nc]

• For a working worker mode after code changes: Reload on code changes (app install/update, config.php)

For this one we could use the watch feature: https://frankenphp.dev/docs/config/#watching-for-file-changes
We should test it and make sure it does not impact performances.

[cdvv7788]

• For a working worker mode after code changes: Reload on code changes (app install/update, config.php)

For this one we could use the watch feature: https://frankenphp.dev/docs/config/#watching-for-file-changes We should test it and make sure it does not impact performances.

https://frankenphp.dev/docs/worker/#restart-workers-manually There is this option too. Send a request on demand whenever some specific events are captured. This seems to be a more focused approach.

I think the autoreload one can work, it is based on https://github.com/e-dant/watcher and it seems to be pretty performant, just need to be careful with the extra layer there as it can introduce issues if it is not reloading correctly.

Both approaches seem to restart gracefully:

https://github.com/php/frankenphp/blob/main/caddy/admin.go#L42 <- Caddy admin endpoint
https://github.com/php/frankenphp/blob/main/watcher.go#L36 <- Filesystem watcher
https://github.com/php/frankenphp/blob/main/worker.go#L211 <- Actual stop function

[cdvv7788]

https://gist.github.com/cdvv7788/2604ac65b746a22d86c11638f37bac1f <- I worked on a naive POC for the DB connections issue. Not sure how it will work in practice, but attaching this at the end of the request may be enough for most situations (Worth experimenting with it in #58541).