Kernel not shut down on page reload in Firefox

[jtpio]

Reloading the page in Firefox doesn't seem to be shutting the kernel down, while it does in Chrome. It is the same issue as the one mentioned in #55 (comment), but seems to be specific to Firefox at the moment.

Tested with:

• Firefox Developer Edition 73.0b10
• Firefox 72.0.1
• Chromium 79

[vidartf]

Does the network page of the developertools reveal any special caching?

[jtpio]

Investigating this further. It looks like the cache is not involved, but instead it might be related to Adblock.

With Adblock enabled, the shutdown request doesn't show up in the network logs:

And with Adblock disabled, the shutdown request (DELETE) is logged:

With Chrome, the DELETE request is fired whether Adblock is enabled or not:

[jtpio]

Moreover, the request doesn't show up in the Adblock tab in the dev tools. This might be because the page is reloaded and the list of blocked requests deleted.

Interestingly, whitelisting localhost doesn't let the DELETE request go through either.

[jtpio]

It looks like it can be reproduced by enabling other browser extensions, such as HTTPS Everywhere (and disabling ad-blockers).

[vidartf]

I'm looking into if it makes sense to start a ServiceWorker here for managing shutdown submissions. A short postMessage in onbeforeunload would hopefully be more reliable.

[vidartf]

Or possibly sendBeacon ?

[jtpio]

sendBeacon looks interesting and could be a substitute for what is currently in #612 (although it seems to be for POST requests only).

[maartenbreddels]

Even better! Learning every day :) Thanks @vidartf !

[vidartf]

The beacon sends an HTTP request via the POST method, with all relevant cookies available when called.

Good point @jtpio. 🤷

[jtpio]

Maybe we would then need an extra voila POST endpoint for that specific purpose. Something like /voila/shutdown with the kernel id in the data (could be worth a try).

[vidartf]

That would hopefully work, although the endpoint name suggest shutting down the server at the moment 😉 Maybe /api/voila/shutdown/{kernel_id} ?

[jtpio]

For sure. Although in some way it still sounds like a workaround and a "duplicate" of the existing jupyter server shutdown endpoint?

[vidartf]

It is definitely a workaround and duplicate. It might be that the XHTTP async=false is the best approach, as a DELETE request should not be too slow. The sendBeacon is probably more relevant to use when you actually need to transfer some amount of data that might lead to otherwise unreasonably slow behavior.

[vidartf]

At least the sendBeacon docs outline what the alternatives are 😉