This problem is inherent to page caching. Workarounds to avoid varying
by cookie for anonymous users are conceptually incorrect. If a single
URL can give different responses depending on who's viewing it, then
it varies by cookie. Preventing CSRF is inherently session-variable as
well. Loading the token via a separate AJAX call is possible, but
there are simpler solutions.
If you want to cache pages with small portions that vary by user, then
you want edge site includes and something like Varnish to process
them. If you want a much slower, pure-python solution that doesn't
require a separate service running somewhere, then you want
armstrong.esi[1].
- Niran
[1] <https://github.com/texastribune/armstrong.esi>. armstrong.esi
isn't part of Armstrong proper yet, but if you want to know more about
the project, head to <http://armstrongcms.org/> and
<https://github.com/armstrong/armstrong>.
> This problem is inherent to page caching. Workarounds to avoid varying
> by cookie for anonymous users are conceptually incorrect. If a single
> URL can give different responses depending on who's viewing it, then
> it varies by cookie. Preventing CSRF is inherently session-variable as
> well. Loading the token via a separate AJAX call is possible, but
> there are simpler solutions.
You may in fact be correct, but I'm not convinced by what you're saying here (not that there is any onus on you to convince me of anything of course).
I"m suggesting that all anonymous users *could* receive an identical page from the server, theoretically, since the same URL does *not* need to return a different response depending on which (anonymous) user is viewing it. CSRF is obviously a trickier problem, and it's not really worth solving the anonymous user problem if CSRF isn't solved as well. But if both problems were somehow solvable, then we're in a position where per-site cache would be viable for many common scenarios such as the one I described in my original post.
If these two problems are in fact unsolvable or not worth solving because simpler alternatives exist, that's fine and understandable. Perhaps per-site/per-view caching are indeed exceptionally limited tools that are beneficial in a very limited number of use cases, and perhaps the "solution" here is tidying up the outstanding bugs and perhaps clarifying the documentation as needed to make the limitations more explicit.
>
> If you want to cache pages with small portions that vary by user, then
> you want edge site includes and something like Varnish to process
> them. If you want a much slower, pure-python solution that doesn't
> require a separate service running somewhere, then you want
> armstrong.esi[1].
Thanks. This post wasn't really about what *I* need btw; I can definitely sort out my caching strategies in other areas as I need to. The post only relates to "me" because I sat down yesterday and said, "Gee, I wonder if I could make use of Django's per-site caching feature for this project I'm working on." I turned it "on" to test it out and then spent the next 6 hours delving into the source code, IRC, ticket tracker, Google etc. to figure out why it wasn't working at all and why @cache_page was, and then after finally sorting it out and grokking all of the moving parts etc, realizing that there was extraordinarily limited value in a per-site/view caching strategy that caches per unique visitor, which is pretty much unavoidable for most common usage patterns.
So yeah, maybe it's me and I'm looking at things the wrong way, but needless to say it wasn't a particularly pleasant or worthwhile experience. Not looking for pity btw, but just wondering what I/we can or should do to make it better.
Jim
For PyLucid i made a simple cache middleware [1] simmilar to Django per-site
cache middleware [2]. But i doesn't vary on Cookies and don't cache cookies. I
simply cache only the response content.
Of course: This doesn't solve the problem if "csrfmiddlewaretoken" in content.
Here some pseudo code from [1]:
-----------------------------------------------------------------
def process_request(self, request):
if not self.use_cache(request):
return
response = cache.get(cache_key)
if response is not None:
return response
def process_response(self, request, response):
if not self.use_cache(request):
return response
# Cache only the raw content
response2 = HttpResponse(
content=response._container, status=200,
content_type=response['Content-Type']
)
patch_response_headers(response2, timeout)
cache.set(request.path, response2, timeout)
return response
-----------------------------------------------------------------
[1]
https://github.com/jedie/PyLucid/blob/master/pylucid_project/middlewares/cache.py
[2] https://docs.djangoproject.com/en/1.3/topics/cache/#the-per-site-cache
Mfg.
Jens D.
Hi Jim,
This is a really useful summary of the current state of things, thanks
for putting it together.
Re the anonymous/authenticated issue, CSRF token, and Google Analytics
cookies, it all boils down to the same root issue. And Niran is right,
what we currently do re setting Vary: Cookie is what we have to do in
order to be correct with respect to HTTP and upstream caches. For
instance, we can't just remove Vary: Cookie from unauthenticated
responses, because then upstream caches could serve that unauthenticated
response to anyone, even if they are actually authenticated.
Currently the Django page caching middleware behaves pretty much just
like an upstream cache in terms of the Vary header. Apart from the
CACHE_MIDDLEWARE_ANONYMOUS_ONLY setting, it just looks at the response,
it doesn't make use of any additional "inside information" about what
your Django site did to generate that response in order to decide what
to cache and how to cache it.
This approach is pretty attractive, because it's conceptually simple,
consistent with upstream HTTP caching, and conservative (quite unlikely
to serve the wrong cached content).
It might be possible to make it "smarter" in certain cases, and allow it
to cache more aggressively than an upstream cache can. #9249 is one
proposal to do this for cookies that aren't used on the server, either
via explicit setting or (in a recently-added proposal) via tracking
which cookie values are accessed. If we did that, plus special-cased the
session cookie if the user is unauthenticated and the session isn't used
outside of contrib.auth, I think that could possibly solve the
unauthenticated-users and GA issues.
However, this (especially the latter) would come with the cost of making
the cache middleware implementation more fragile and coupled to other
parts of the framework. And it still doesn't help with CSRF, which is a
much tougher nut to crack, because every response for pages using CSRF
come with a Set-Cookie header and probably with a CSRF token embedded in
the response content; and those both mean that response really can't be
re-used for anyone else. (Getting rid of the token embedded in the HTML
means forms couldn't ever POST without JS help, which is not an option
as the documented default approach). You can mark some form-using views
that are available to anonymous users as csrf-exempt, which exposes you
potentially to CSRF-based spam, but isn't a security issue if you aren't
treating authenticated submissions any differently from
non-authenticated ones.
Generally, I come down on the side of skepticism that introducing these
special cases into the cache middleware really buys enough to be worth
the added complexity (though I could be convinced that #9249 is worth it).
I do think we should improve the cache middleware documentation so its
limitations are outlined more clearly upfront, and point people towards
existing solutions for caching mostly-but-not-entirely-anonymous pages:
edge-side-includes, two-phase-render, and JS/AJAX fetch.
#15855, on the other hand, is a bug that really does need to be fixed. I
still don't see a better fix than the one I outlined in the ticket
description: requiring some middleware to be in MIDDLEWARE_CLASSES for
the cache_page decorator to work, and not doing the actual caching until
we hit that middleware. Or alternatively, adding an implicit "cache any
responses that had cache_page used on them" phase to response
processing, after all middleware. I think those are both ugly fixes,
though; maybe someone has a better idea. The last time I know of that
this was discussed in-depth was in
http://groups.google.com/group/django-developers/browse_frm/thread/f96e982254fbe5c3/2b02361fd6e706f4
Carl
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.10 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/
iEYEARECAAYFAk6gxKkACgkQ8W4rlRKtE2dnggCfeNOeAw8g4/Y5Zu6iM73HFK0m
V6EAn0mGvzLzOs0daC1UZWQp6hZnxvH8
=La3y
-----END PGP SIGNATURE-----
Thanks Carl. This is definitely a good, clarifying response to what I was mulling around about.
A few thoughts of my own to add here:
* You and Nihan are certainly right about upstream caches. Regardless of what we do here, we'll have to vary by cookie in the response header. This makes sense for a site that offers authentication: Django needs to check on every page view to see if the user is authenticated, so we can't have the upstream cache holding on to a page for us.
* Agreed about how the "smartness" comes at the cost of brittleness if the implementations are too tightly coupled. That said, I can squint and sort of see an implementation that could thread the needle here. It would require something like:
- An API in the cache middleware instructing it to ignore certain cookies for the purposes of caching (i.e. something along the lines of #9249).
- Some kind of "pre-fetch" hook in the cache middleware. Whether it's a flag in the request object, a signal or something else, give other systems the ability to look at a request before it hits the FetchFromCacheMiddleware and either allow or prevent the response from being pulled from the cache. E.g if there was a flag request.invalidate_cache that defaults to False, the contrib.auth app could, in combination with the above, pull the session id from consideration in the cache key and do an authentication check on its own, invalidating the cache on its own if the user is authenticated. The core idea is what you already suggested, I'm more illustrating here that this can conceivably be implemented as an API, making it less brittle.
- Some kind of "post-fetch" hook in the cache middleware, combined with a retooling of the CSRF middleware. This is getting in the clouds here a bit, but a hook on the opposite end of the fetch operation could allow the CSRF app to add its token after the response was pulled from the cache. I say we're in the clouds here because for something like this to work the CSRF would have to do a little two-step dance. Before the UpdateCache step the CSRF would had to insert something that looked like a server-side template tag, which gets cached, and then after that step the CSRF would have to insert it's actual value. On the fetch side, the CSRF would have to make use of the post fetch hook to pull the cached paged rendered with the server side template tag thingy and then add the correct value on its way out the door. Essentially, we're talking about a poor man's two phase rendering system.
This barely qualifies as a thought exercise let alone a proposal, but my main underlying suggestion here is that if the cache middleware correctly implemented hooks of some kind in the right locations, it might well be possible for systems like auth and CSRF to do what they would need to do without coupling all these systems together in a giant ball of twine.
> I do think we should improve the cache middleware documentation so its
> limitations are outlined more clearly upfront, and point people towards
> existing solutions for caching mostly-but-not-entirely-anonymous pages:
> edge-side-includes, two-phase-render, and JS/AJAX fetch.
>
> #15855, on the other hand, is a bug that really does need to be fixed. I
> still don't see a better fix than the one I outlined in the ticket
> description: requiring some middleware to be in MIDDLEWARE_CLASSES for
> the cache_page decorator to work, and not doing the actual caching until
> we hit that middleware. Or alternatively, adding an implicit "cache any
> responses that had cache_page used on them" phase to response
> processing, after all middleware. I think those are both ugly fixes,
> though; maybe someone has a better idea. The last time I know of that
> this was discussed in-depth was in
> http://groups.google.com/group/django-developers/browse_frm/thread/f96e982254fbe5c3/2b02361fd6e706f4
>
> Carl
My thinking right now as far as moving forward:
1. Fixing #9249 and #15855. I hear your philosophical concerns about #9249 but the ubiquity of Google Analytics means we must do fine some way to fix it (IMO). Addressing these two tickets would at least ensure page caching wasn't actually broken. I'll try to jump in on those if I have time later next week. #9249 in particular seems quite close.
2. Clarifying the documentation. I think an admonition in the page caching section of the docs which outlined the present challenges a developer might face implementing it would probably have done the trick for me when I was first glancing at it. I can open a ticket on that next week, again if I have time.
It'd be great if these two got in for 1.4.
3. Addressing the other stuff is I guess for now a sort of "some day" goal. I continue to feel strongly that it's a worthy goal, particularly given that CSRF and contrib.auth are such fundamental parts of most projects and that they really are the only two things that stand in the way of page caching being a viable option in many projects. If anyone else gets inspired by this goal let me know, otherwise I'm content for the time being to let it stew.
Thanks all for listening.
The following is from the stupid ideas department: Maybe there could be a "reverse cache" template tag, such that you would mark the places where you want changing content as non-cacheable. You would need two views for this, one which would construct the "base content" and then another which would construct the dynamic parts. Something like:
page_cached.html:
... expensive to generate content ...
{% block "login_logout" non_cacheable %}
{% endblock %}
... expensive to generate content ...
You would generate the base page by a cached render view:
def page_view_cached(request, id):
if cached(id):
return cached_content
else:
... expensive queries ...
return cached_render("page_cached.html", context, ...)
The above view would not be directly usable at all, you would need to use a wrapper view which would render the non-cacheable parts:
def page_view(request, id):
# Below would return quickly from cache most of the time
cached_portions = page_view_cached(request, id)
return render_to_response("page.html", context={cached: cached_portions, user:request.user})
where page.html would be:
{% extends cached %}
{% block login_logout %}
{% if user.is_authenticated %}
Hello, user!
{% else %}
<a href="login.html">login</a>
{% endif %}
{% endblock %}
That seems to be what is really wanted in this situation. The idea is quite simply to extend the block syntax to caching. A whole another issue is how to make this easy enough to be actually usable, and fast enough to be actually worth it.
- Anssi
________________________________________
From: django-d...@googlegroups.com [django-d...@googlegroups.com] On Behalf Of Jim Dalton [jim.d...@gmail.com]
Sent: Friday, October 21, 2011 16:02
To: django-d...@googlegroups.com
Subject: Re: The state of per-site/per-view middleware caching in Django
Thanks all for listening.
--
You received this message because you are subscribed to the Google Groups "Django developers" group.
To post to this group, send email to django-d...@googlegroups.com.
To unsubscribe from this group, send email to django-develop...@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/django-developers?hl=en.
> I do not know nearly enough about caching to participate fully in this discussion. But it strikes me that the attempt to have CSRF protected anonymous page cached is not that smart. If you have an anonymous submittable form, why bother with CSRF protection? I mean, what is it protecting against? Making complex arrangements in the caching layer for this use case seems like wasted effort. Or am I missing something obvious?
First issue is that CSRF can matter for anonymous users. From here http://www.squarefree.com/securitytips/web-developers.html#CSRF:
Attacks can also be based on the victim's IP address rather than cookies:
• Post an anonymous comment that is shown as coming from the victim's IP address.
...
• Perform a distributed password-guessing attack without a botnet. (This assumes they have a way to tell whether the login succeeded, perhaps by submitting a second form that isn't protected against CSRF.)
So two very common uses cases for anonymous forms are log in forms and anonymous comment forms, both of which are potentially vulnerable. I guess I feel like it's quite common to have forms on a page these days even for anonymous users.
Second is -- and I don't know about this -- but I don't know how well CSRF handles authentication conditionally. Like if I have a page and let's say that page has forms in it for logged in users but nothing for anonymous user, can I conditionally exempt the formless page from CSRF? I have no idea, but buy default I presume it's on and I presume the cache is varying on it.
So, yes, you could probably optimize a lot of this to sort of skip around the CSRF issue and it's not a deal breaker. But my main argument has been the ubiquity of CSRF + user authentication in Django projects to me means a solution to both of these is a requirement for page caching to become easy and applicable in most scenarios.
>
> The following is from the stupid ideas department: Maybe there could be a "reverse cache" template tag, such that you would mark the places where you want changing content as non-cacheable. You would need two views for this, one which would construct the "base content" and then another which would construct the dynamic parts. Something like:
>
Your idea sounds a lot like the "server side include" or "two phased template rendering" approach that I know some people are doing. Here's an excellent example of this approach being used in EveryBlock (from two years ago):
http://www.holovaty.com/writing/django-two-phased-rendering/
And looks like some core devs have been involved at some point in this implementation of that concept:
https://github.com/codysoyland/django-phased
That looks almost exactly like your idea: "django-phased contains a templatetag, phased, which defines blocks that are to be parsed during the second phase. A middleware class, PhasedRenderMiddleware, processes the response to render the parts that were skipped during the first rendering."
I guess that was sort of what I was hinting at in my previous discussion about how to handle CSRF. In the link he is taking it to the next level (where even logged in users get the page and that stuff is done after.
Anyhow it's obviously a sensible conceptual approach. It would be a stretch to fit that into the existing page caching approach of Django obviously.
On 10/21/2011 07:02 AM, Jim Dalton wrote:
> 1. Fixing #9249 and #15855. I hear your philosophical concerns about
> #9249 but the ubiquity of Google Analytics means we must do fine some
> way to fix it (IMO). Addressing these two tickets would at least
> ensure page caching wasn't actually broken. I'll try to jump in on
> those if I have time later next week. #9249 in particular seems quite
> close.
>
> 2. Clarifying the documentation. I think an admonition in the page
> caching section of the docs which outlined the present challenges a
> developer might face implementing it would probably have done the
> trick for me when I was first glancing at it. I can open a ticket on
> that next week, again if I have time.
>
> It'd be great if these two got in for 1.4.
Agreed - any work you're able to put in on any of these is very welcome.
> 3. Addressing the other stuff is I guess for now a sort of "some day"
> goal. I continue to feel strongly that it's a worthy goal,
> particularly given that CSRF and contrib.auth are such fundamental
> parts of most projects and that they really are the only two things
> that stand in the way of page caching being a viable option in many
> projects. If anyone else gets inspired by this goal let me know,
> otherwise I'm content for the time being to let it stew.
I take your point that it might be possible to do a cache-tweaking API
that could allow the cache to be more aggressive around auth and CSRF
with less coupling (though you'd still end up sprinkling cache-specific
stuff into auth and CSRF with your approach). I remain pretty skeptical
about whether this is a good idea; it seems like it could significantly
increase the surface area for bugs in the cache middleware
implementation, and just generally make the implementation harder to
maintain with correct behavior. (I have some painful experience in this
area: CACHE_MIDDLEWARE_ANONYMOUS_ONLY is the one existing, and
relatively simple, instance of the type of enhanced caching logic you're
talking about, and I made some fixes to it in the 1.3 cycle that I then
later had to fix again due to unanticipated side effects of the first
change). But at this point this is all kind of hand-waving without code
to look at.
You might also consider what's possible to do outside of core as a
third-party alternative to Django's caching middleware. When you're
proposing major and somewhat experimental changes, that can be a
powerful way to demonstrate that the idea is workable, and makes it a
lot easier to pick up users and advocates; people are generally more
willing to try out a third-party tool than to run or test with a patched
Django.
Carl
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.10 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/
iEYEARECAAYFAk6hvEkACgkQ8W4rlRKtE2edBACfdsW7IHoDKBrpwzwGGMx+ww5g
U+AAoLZLn1CA6c1644kzsnZRZ6xaW60B
=QBjT
-----END PGP SIGNATURE-----