tore g. eschliman

Tested this out on a beta service and not seeing the hoped-for results; noticing that the phased restart is extremely slow (we run 31 workers but it took over 6...

This is close but is also likely to be superseded by upcoming work to convert worker-0 to a non-serving parent process (or "mold"), which I'm hoping to start work on...

@nateberkopec started work at https://github.com/puma/puma/pull/3619/files on this; rough idea is just for worker zero to stop handling requests whenever it's asked to do its first refork (after initialization) and move...

We intend to keep using this on reasonably large deployments at Instacart, and we're fine doing it while it's labeled "experimental". Given the extent of the challenges in implementing it...

We got a basic working `mold` [implementation](https://github.com/toregeschliman/puma/pull/1) up and tested in our preprod environment but were not happy enough with some of the compromises involved (it involves some weird juking...

We have a working mold implementation that we're running in production, early results are good but I'm letting it run for a few weeks to build some confidence in edge-case...

Alright, we've been dogfooding this for a couple weeks and are seeing great results in terms of cold-start performance and shared memory usage, so I put up a PR while...

`on_refork` doesn't apply to `mold_worker` (at least right now) because it doesn't make any sense in the context; the mold process isn't doing anything between reforks that would require a...

@MSP-Greg adding CI coverage and docs is just making the PR bigger; should I be following @joshuay03 's lead and splitting out as much as I can or will larger...

> it's very helpful to have the test commits before the 'lib' commits fully agree, but for the next week or so I'm mostly just celebrating surviving the integration specs......