Change 195826 had a related patch set uploaded (by Gergő Tisza):
[TEST] Wrap arguments to $.Callbacks.add in try..catch

https://gerrit.wikimedia.org/r/195826

The easy but invasive way to do this is to replace $.Callbacks.add() (which will affect at least all promise and animation operations in jQuery); the hard but more targeted way is to decorate $.ajax() and modify the promise it returns, but then we also need to modify any promise that returns (via ˙.promise() and .then`), and so on.

Ori suggested that vbench can be used to check the performance penalty of the first version.

A third alternative is to replace the XHR object with a proxy (that is possible via the xhr property of the AJAX options object) and use that to wrap the callback. That's the most targeted of the three approaches, but somewhat fragile as there is no way to detect when properties of the proxy object are changed (ES6 proxies are only supported by Firefox at this time; defineProperty setters only work with known property names and using them on a native object might result in errors; __defineSetter__ is reasonably well-supported (all major browsers except IE 8-10) but non-standard and again might behave unexpectedly on a native object).

The two approaches that seems the most useful for XHR proxying:

• bail out on browsers which do not support Object.defineProperty (basically just IE 8); define a setter for xhr.onreadystatechange. Worst case this might result on attempts to set the value (and thus all AJAX calls) to error out, but that seems unlikely.
• bail out on browsers which do not support __defineSetter__ (IE 8-10); replace the XHR object with a proxy which uses __defineSetter__ to forward any property changes of the proxy to the real XHR object; set up proxy functions for all methods of the real XHR object. Whenever the real onreadystatechange is called, mirror all properties of the real XHR object to the proxy. This avoids messing with the native object, but involves much more faking which will probably break if someone else tries to mess with the native object.

In T92247#1133722, @Tgr wrote:

define a setter for xhr.onreadystatechange

Unfortunately this breaks in Chrome. This works as expected:

var xhr = $.ajaxSettings.xhr();
xhr.open( 'GET', '/w/api.php?action=query&format=json&meta=siteinfo', true );
xhr.send();
Object.defineProperty( xhr, 'onreadystatechange', { configurable: true, enumerable:true, value: mw.log, writable: true } );

This does not work, even though inspection shows xhr.onreadystatechange to have the correct value:

var xhr = $.ajaxSettings.xhr();
xhr.open( 'GET', '/w/api.php?action=query&format=json&meta=siteinfo', true );
xhr.send();
Object.defineProperty( xhr, 'onreadystatechange', { configurable: true, enumerable:true, get: $.noop, set: $.noop } );
Object.defineProperty( xhr, 'onreadystatechange', { configurable: true, enumerable:true, value: mw.log, writable: true } );

(reported upstream: #469024)

In T92247#1133722, @Tgr wrote:

replace the XHR object with a proxy which uses __defineSetter__

__defineSetter__ also only works on specified property names; must have confused it with PHP magic functions. So there is no point in using it; defineProperty does the same and has better support.

Change 198178 had a related patch set uploaded (by Gergő Tisza):
Make AJAX callback test harder to pass

https://gerrit.wikimedia.org/r/198178

Change 198183 had a related patch set uploaded (by Gergő Tisza):
[WIP] Replace XHR in $.ajax with a proxy object with error catching

https://gerrit.wikimedia.org/r/198183

Change 198178 merged by jenkins-bot:
Make AJAX callback test harder to pass

https://gerrit.wikimedia.org/r/198178

Change 195826 abandoned by Gergő Tisza:
[TEST] Wrap arguments to $.Callbacks.add in try..catch

Reason:
Abandoning, see T92247 #1147500. Might be revisited later.

https://gerrit.wikimedia.org/r/195826

Change 198183 abandoned by Gergő Tisza:
[WIP] Replace XHR in $.ajax with a proxy object with error catching

Reason:
Abandoning, see T92247#1147500. Might be revisited later.

https://gerrit.wikimedia.org/r/198183

At the time of creating this task it was assumed that the server side implementation of error logging would be based on Sentry. We have eventually decided on a different implementation, so de-tagging.