Add support for OPTIONS['context_processors'] to Jinja2 template backend

[Carl Meyer]

Conclusion from IRC discussion: There are advantages to being explicit about request-dependence in the template (which context processors are not), but there's currently no convenient alternative to context processors for computing a non-trivial request-dependent value that you want computed only once per template render (and not at all if no template is rendered). And even if we don't like context processors as an API, they have strong precedent in Django, so people will probably find a way to use them regardless (e.g. using django-jinja instead of the built-in Jinja2 backend).

They could be added as a top-level cross-backend feature, but since in 1.8 they already exist as a DTL-specific feature in OPTIONS, they should be added to the Jinja2 backend the same way.

Full IRC discussion transcript below. It's long, but seems useful to more fully document the thinking here for posterity:

18:35 <+carljm> mYk: I'm upgrading a project to Django 1.8, and I am wondering if we were wrong to not provide context processors for Jinja.
18:35 <+carljm> IIRC the rationale was something like "you don't need them in Jinja because you can just put stuff into the env globals 
                directly", but this isn't really true - the point of context processors is that they provide request-specific values, and 
                there's no obvious way to do that.
18:37 <+carljm> You can emulate by putting a function into the context that returns some value, but then you have inefficiency if that function 
                does something non-trivial and its value is used in multiple places.
18:37 <+carljm> For a concrete example, this project wants a serialized (i.e. to primitive values), JSON-serializable representation of 
                request.user to be available in all templates.
18:38 <+carljm> In 1.7 this was done with a context processor that provided that representation in a `me` variable.
18:39 <+carljm> In 1.8, without some context-processor equivalent, that would have to become `me(request)` or `request|me` or something, and to 
                avoid serializing multiple times needlessly, it would often turn into `{% set me = me(request) %}`
18:40 <+carljm> In looking into this, I also noticed that Flask does provide context processors (just like Django's) atop Jinja2.
18:41 <+carljm> At the moment, I'm implementing context processors for the Django Jinja2 backend. It's not hard, but unfortunately requires 
                copying most of the implementation of the Jinja2 backend and its Template wrapper (they aren't really factored for subclass 
                customization)
18:41 <+carljm> I'd be very interested to hear if you think that I'm Doing it Wrong and there's a better option available for this type of use 
                case.
18:43 <+carljm> Or if I'm not wrong, perhaps we should consider adding support for OPTIONS['context_processors'] to the Jinja2 backend for 1.9.
18:46 <+carljm> Or at least make it easier to subclass the backend and add that support. Specifically, this would mean factoring out an 
                "instantiate_template" (name to-be-bikeshedded) internal method on the backend class, which would be called by both from_string 
                and get_template, and which would allow a subclass to easily swap out both the Template wrapper class used and the arguments 
                passed to it.
18:50 <+carljm> (Actually with @jinja2.contextfunction you could remove the explicit naming of the request in the above examples, but the other 
                issues still apply.)
18:51 <+jezdez> carljm: doesn’t django-jinja support ctx processors natively?
18:52 <+jezdez> not that it answers your questions above, just something that may buy you quicker results
18:54 <+carljm> jezdez: So does jingo (which is what we were using on 1.7)
18:55 <+carljm> Unfortunately jingo isn't 1.8-compatible (due to its use of template internals that changed in 1.8); do you know if 
                django-jinja is?
18:55 <+carljm> In any case, part of the goal of this upgrade was to switch from an external Jinja support package to the native support, so 
                I'll probably continue on this path. Already done with adding context proc support, anyway :-)
18:55 <+jezdez> carljm: right, I think I’d like to move away from jingo in general since it does a few things oddly
18:55 <+carljm> Thanks though!
18:56 <+jezdez> yeah, I heard django-jinja is 1.8 compatible
18:56 <+MarkusH> I'd probably look into what they are doing to provide context processors and may adapt some of the monkey patches
18:56 <+carljm> jezdez: What are you thinking of specifically? I was pretty happy with jingo as a pre-1.8 solution.
18:56 < pmachine> carljm: django-jinja provides it's own backend class for 1.8
18:56 <+carljm> Ah, nice!
18:56 <+jezdez> carljm: the helpers api is clunky and the monkey patches just smell like tech debt
18:56 <+jezdez> I don’t see why we can’t have one and only one implementation going forward now that 1.8 is out
18:57 <+jezdez> carljm: but yeah, I get what you’re saying, it’d be neat to just move away from any library
18:57 <+jezdez> maybe in 2.0
18:57 <+jezdez> (not kidding)
18:57 <+carljm> Yeah; with 1.8 having a Jinja2 backend, I'd like to see the "one and only implementation" be that one.
18:58 <+carljm> as in, if using the built-in backend is painful enough that most people are using an external library on top of it still, we 
                should fix those pain points.
18:58 <+jezdez> so maybe we should spend more time on implementing translation support and ctx processors?
18:58 <+carljm> yep :-)
[snip]
19:05 < pmachine> btw, I'm +1 on enabling context processors and making the backend easier to subclass
19:07 <+carljm> pmachine: cool. I'll file an issue and PR for it, but I'd like to see if mYk has any comments first.
19:08 <+carljm> (or more likely, two issues and two PRs)
19:14 <+mYk> am I being summoned?
19:14 <+mYk> I'm reading the backlog
19:15 <+mYk> carljm: part of the decision was: "it's easier not to provide context processors at first and add them later than the opposite"
19:16 <+mYk> if Flask provides them, that's a strong argument
19:16 <+carljm> yes, definitely better to start minimal
19:16 <+carljm> Flask, and it seems every third-party Django-Jinja adapter, too.
[snip discussion of Jinja backend subclassibility]
19:21 <+mYk> it seems more straightforward to me to just add support for context processors
19:22 <+mYk> the question is -- should it be a "standard" option (like DIRS and APP_DIRS) or a "backend-specific" option (which goes inside 
             OPTIONS)?
19:23 <+mYk> since 1.8 was released already and context_processors is a "backend-specific" option of the Django backend, the latter option 
             seems less disruptive
19:23 <+carljm> yes, that was my feeling too
19:23 <+mYk> to be clear -- this is mostly a convenience & performance thing, right?
19:24 <+mYk> you want to write {{ context_processor_output }} instead of {{ context_processor(request) }} in templates
19:24 <+carljm> precisely, yes. it doesn't enable anything that is impossible otherwise.
19:24 <+carljm> it's really the same thing you're doing with csrf_input and csrf_token in the current backend
19:25 <+carljm> only there you've used yet another approach (Lazy), which is one I'd dearly love to avoid :-)
19:25 <+mYk> well I had to in that case
19:25 <+mYk> coming back to context processors
19:25 <+mYk> I really like the explicitness of {{ context_processor(request) }}
19:26 <+carljm> yes, in that case it does have to be lazy so CSRF token isn't triggered unnecessarily - but that approach does solve the same 
                cases
19:26 <+mYk> I think it fits well with Django's insistence on passing the request to each view function
19:26 <+carljm> that in itself doesn't bother me much
19:26 <+carljm> but for a value which may be used many places, the efficiency concern does.
19:27 <+mYk> it also clarifies whether you can render a given template without a request (ie outside of the request-response cycle)
19:27 <+carljm> you can do {% set somevar = context_processor(request) %}
19:27 <+carljm> but for ubiqitous values (the only kind I would ever use a context proc for anyway), that's ugly boilerplate to have in all 
                templates
19:28 <+mYk> do you often have computationally expensive things that you use multiple times in a given template?
19:28 <+mYk> you could also cache the value in a private attribute of the request, but that's ugly
19:28 <+carljm> yes, I considered that, too
19:29 <+mYk> if we step back for a minute, my general position is:
19:30 <+mYk> * real-life usage of context processors (starting with django's built-in processors) leads me to think that they aren't a very 
             good API
19:30 <+carljm> in the current project, I have four context processors I would use in Jinja. Only one of them (serialized user) is expensive 
                and widely-used enough that I'd be concerned about calling it at every use site.
19:31 <+mYk> * they implicitly depend on the request, which is a net negative for me, because I'm a strong proponent of dumb code
19:31 <+mYk> * if we come up with use cases for which they're clearly the best answer, I'm happy to add them, but I'll still be afraid that 
             they will be misused
19:33 <+mYk> based on what you've described, I'm wondering if some sort of caching isn't a better answer to your goal: optimizing performance 
             by avoiding repeated calls to an expensive function
19:34 <+carljm> ...thinking...
19:34 <+mYk> (and that will avoid paying the cost of calling that function on pages that don't use it, if any)
19:35 <+mYk> one last idea: context processors look to me like response middleware in disguise
19:35 <+mYk> if you have something complicated (or expensive) to do in each response, perhaps doing it explicitly in a middleware would be a 
             good design
19:36 <+carljm> Yes, that's a variant of "cache the value in a private attribute of the request", where the value instead becomes a public 
                attribute of the request, attached in a middleware.
19:37 <+carljm> I agree that solves the use case, and preserves the explicit dependence on the request.
19:38 <+mYk> the various ideas I've suggested are merely shifting the executing of this function and the caching of its results at various 
             points in the request/response cycle
19:38 <+carljm> It means the value is computed for all requests, even when a template is not rendered.
19:38 <+mYk> I'm trying to optimize for "there should be one obvious way to do it" and "explicit is better than implicit"
19:39 <+carljm> Yes. The question is whether "compute once when a template is rendered, and make globally available in the template context" 
                has enough good use cases to provide API for.
19:39 <+mYk> it means the value is computed for all requests => maybe -- off the top of my head, I'm not sure how to tell if a given response 
             is going to render a template
19:39 <+carljm> My meta-concern is that it doesn't really matter whether context processors are good API; Django is stuck with them. Unless we 
                deprecate them for DTL, the result of not providing them for Jinja may be that people just use external Jinja backends instead 
                which provide them.
19:40 <+mYk> yes...
19:41 <+mYk> being technically correct serves us well if everyone keeps using django-jinja...
19:41 <+mYk> that reminds me of mistakes I made in the time zone APIs, which were technically correct at first, and are slowly becoming usable 
             by mortals
19:42 <+carljm> I think for my use case (which I think is the prototypical best use case for a context processor), the best alternative would 
                be a cached_property on the request.
19:42 <+mYk> carljm: a realistic solution may be to provide context processors for users who are used to them and don't want to think about it, 
             and discussthe alternatives in the docs
19:42 <+carljm> Thus not computed unless accessed, but computed only once when accessed.
19:42 <+carljm> Unfortunately Django makes that hard to do - trying to do the equivalent in a middleware really requires use of Lazy
19:43 <+carljm> because you don't get to subclass Request, you just get to monkeypatch it.
19:43 <+mYk> yup, AuthMiddleware has horrific code to set up request.user
19:44 <+carljm> yes, horrific code which has been rewritten at least once because the previous version was both horrific and broken.
19:45 <+mYk> I'm not taking bets on whether the current version is also horrific and broken :-)
19:45 <+carljm> Well, my conclusion at this point is that while I see the arguments for explicit dependence on the request, I don't yet see a 
                practical alternative to a context processor that I would be happy with for this use case.
19:45 <+mYk> okay
19:46 <+carljm> And given that, plus the precedent for context-processors in Django and the likelihood that people will find an implementation 
                of them one way or another, I think I'd remain in favor of adding them.
19:46  -!- f3lp <~f@179.218.60.168> has joined #django-dev
19:46 <+carljm> (Perhaps with discussion of downsides and alternatives in the docs.)
19:46 <+mYk> that works for me
19:47 <+mYk> keeping them as a "backend-specific" option will also de-emphasize them a bit
19:47 <+mYk> (even though we're just doing this because the ship has sailed with 1.8)
19:47 <+carljm> A cached function is the closest thing to an alternative that I'd want to use, but then you have to decide between caching in a 
                private request attr (ick) vs some other caching approach (which feels wrong because the value is logically request-scoped)
19:48 <+carljm> sure
19:48 <+carljm> And in practice unless we provide a seamless convenience API for the request-scoped-cached-function approach, I don't think 
                anyone will choose to use it over context processors.
19:50 <+carljm> Ok, I will submit an issue (now) and PR (later).
19:51 <+carljm> Thanks for the thought-provoking discussion!