process_view middleware prevents process_exception handling

[Michael Manfre]

If process_view() calls a view that raises and exception, it is handled by django.core.handlers.base.BaseHandler.handle_uncaught_exception() instead of being passed to the middleware for process_exception() handling.

[Jeremy Dunck]

This really needs tests. It's a critical piece of code and there are lots of exception paths around there.

[Michael Manfre]

Patch against trunk rev 13959

[Michael Manfre]

Updated the patch against trunk and hoping to get this in to 1.3

[Ramiro Morales]

So the ticket originally was about exceptions generated in the view itself not being caught an processed by .process_exception().

But the code modification and accompanying tests proposed aren't about exceptions in the view but exceptions originated in a middleware process_view() method.

The first is no longer an issue (I suspect it was fixed in r14393 and r14938) and the second one still happens but I'm not sure the change of behavior in that case should happen. I'm adding a modified and simplified patch demonstrating both scenarios.

Because of that, I'm moving this to design decision needed.

[Ramiro Morales]

Patch modified and simplfied

[Aymeric Augustin]

Change UI/UX from NULL to False.

[Aymeric Augustin]

Change Easy pickings from NULL to False.

[kevin1024]

I think the latest patch is missing the actual behavior modification. It seems to only have tests in it.

[Walter Doekes]

126	+        """
127	+        A <Middleware>.process_view() that generates an exception and another
128	+        middleware whose process_exception() provides a response. Result should
129	+        be that response.
130	+        """
131	+        settings.MIDDLEWARE_CLASSES = (
132	+            'regressiontests.base_handler.middleware.ProcessViewRaise',
133	+            'regressiontests.base_handler.middleware.ProcessExceptionResponse',
134	+        )

Well, I have a common scenario that would benefit from having the exception handler handle the process_view() exception:

django/middleware/csrf.py

calls

    request_csrf_token = request.POST.get('csrfmiddlewaretoken', '')
...
    self._load_post_and_files()
...
    return self._stream.read(*args, **kwargs)

IOError: error reading wsgi.input data

I'd love to ignore that particular IOError, and I can, when it's generated in the view. But I can't when it's generated in the CSRF process_view.

As for how: I'd expect it to be handled just as if it was an exception in the view itself. I.e. surround the _view_middleware calls with a try/except just like the "response = callback(...)" bit.

Regards,
Walter Doekes

[Anssi Kääriäinen]

Another similar issue exists in process_request(). Consider case:

MIDDLEWARE_CLASSES = [
    'TransactionMiddleware',
    'SavesToDB',
    'RaisesException'
]

What happens is that TransactionMiddleware will open a transaction in process_request, savestodb will save to db and then RaisesException will raise an exception. The process_exceptions aren't called in this case. Finally process_response is called for every of the middlewares and thus TranscationMiddleware will commit the transaction. This is not good.

I see two solutions here - either we do the process_exception() dance for all middlewares for which process_request has been called (in process_view() case this means every middleware). Or, we decide that process_* should not ever raise an exception, and we go to panic mode if this happens. The panic mode would mean that we would try to release all resources (close connections etc) and then return a barebone 500 response.

To me the panic mode doesn't seem like a good idea.

I reported this in #19648. I marked that ticket as a duplicate of this one.

[Aymeric Augustin]

This is almost the same problem as #16367 which I closed as wontfix.

Given the amount of discussion above, I'm not going to close as wontfix right now, but that's still my inclination.

[Aymeric Augustin]

#12909 was closed as a duplicate.

[Aymeric Augustin]

Overall, the current design of the request handler assumes that middleware methods won't raise exceptions. They must deal with their own exceptions. Exceptions raised by middleware will end up in handle_uncaught_exception and produce a 500.

This may have some drawbacks, but at least it's consistent. Aborting on the first exception also avoids the problem of deciding which middleware should or shouldn't run after the first exception.

I could consider a different behavior, but that would require a global proposal for consistent exception handling in middleware. It seems quite tricky to design something that is useful, consistent, and simple.

I'm going to close this ticket because it's a piecewise approach of the problem that introduces some inconsistency.

[Walter Doekes]

Should I create a new ticket for the unhandled exception in the csrf middleware then? :)

[Aymeric Augustin]

Yes please. Sorry, I missed that in the comments.

[Christopher Martin]

"the current design of the request handler assumes that middleware methods won't raise exceptions. They must deal with their own exceptions. Exceptions raised by middleware will end up in handle_uncaught_exception and produce a 500."

That would be a great piece of information to include in ​https://docs.djangoproject.com/en/dev/topics/http/middleware/#writing-your-own-middleware