onDestroy hook not called on SSR since version 3.39.0 in production mode

[cudr]

Describe the bug

According documentation onDestroy hook must be called on server when component is destroyed. It looks like since version 3.39.0 it never gets called.

Reproduction

Clone repo: https://github.com/cudr/svelte-onDestroy-bug-repro then install dependencies and build.
Start project npm run start, go to main page and check terminal output (there are no console here).

If svelte is downgraded to 3.38.3 this will work fine and the onDestroy hook will be called.

Logs

No response

System Info

System:
    OS: Linux 5.8 Ubuntu 20.04.2 LTS (Focal Fossa)
  Binaries:
    Node: 16.6.1 - ~/.nvm/versions/node/v16.6.1/bin/node
    Yarn: 1.22.11 - ~/.nvm/versions/node/v16.6.1/bin/yarn
    npm: 7.20.3 - ~/.nvm/versions/node/v16.6.1/bin/npm
  Browsers:
    Firefox: 91.0.1
  npmPackages:
    svelte: ^3.42.2 => 3.39.0

Severity

blocking an upgrade

[cudr]

Maybe related #6416

[dummdidumm]

Lifecycle hooks should not be called during SSR/on the server, they are meant for the browser only. What's your use case for relying on them getting called on the server? I'm honestly a bit surprised, I don't think they ever were called on the server.

[bluwy]

I had a usecase for needing onDestroy called in SSR from a blog post I wrote. TLDR, I'm using using onDestroy as a hook to do cleanup after the app is rendered in the server, and before Sapper serializes the session, because I was doing some hacky workarounds to get Apollo Client's cache to work. I think it's a breaking change that we remove that functionality now though.

[dummdidumm]

Interesting ... As you see I'm not that well-versed with SSR + Hooks, but if onDestroy was called before but isn't know then that's somewhat of a breaking change, yes. Does this happen for other hooks as well?

[bluwy]

Should be onDestroy only. Per the documentation it's the only hook that's called in the server-side.

[dummdidumm]

@Rich-Harris @benmccann this seems to have slipped while designing the noops for SvelteKit. Is this something we need to revert or do we go "works as designed for SvelteKit"-route here and adjust the documentation?

[dominikg]

i think this one slipped through, the original request describing the feature explicitly mentions onMount, beforeUpdate and afterUpdate #6372

#6677 looks like the proper fix to restore the documented behavior

[benmccann]

I'm okay to revert it. The original motivation was to not call onMount on the server. I think it just slipped by us all and no one realized onDestroy was called on the server