Defending Pyramid’s Design

From time to time, challenges to various aspects of Pyramid design are lodged. To give context to discussions that follow, we detail some of the design decisions and trade-offs here. In some cases, we acknowledge that the framework can be made better and we describe future steps which will be taken to improve it; in some cases we just file the challenge as noted, as obviously you can’t please everyone all of the time.

Pyramid Provides More Than One Way to Do It

A canon of Python popular culture is “TIOOWTDI” (“there is only one way to do it”, a slighting, tongue-in-cheek reference to Perl’s “TIMTOWTDI”, which is an acronym for “there is more than one way to do it”).

Pyramid is, for better or worse, a “TIMTOWTDI” system. For example, it includes more than one way to resolve a URL to a view callable: via url dispatch or traversal. Multiple methods of configuration exist: imperative configuration, configuration decoration, and ZCML (optionally via pyramid_zcml). It works with multiple different kinds of persistence and templating systems. And so on. However, the existence of most of these overlapping ways to do things are not without reason and purpose: we have a number of audiences to serve, and we believe that TIMTOWTI at the web framework level actually prevents a much more insidious and harmful set of duplication at higher levels in the Python web community.

Pyramid began its life as repoze.bfg, written by a team of people with many years of prior Zope experience. The idea of traversal and the way view lookup works was stolen entirely from Zope. The authorization subsystem provided by Pyramid is a derivative of Zope’s. The idea that an application can be extended without forking is also a Zope derivative.

Implementations of these features were required to allow the Pyramid authors to build the bread-and-butter CMS-type systems for customers in the way they were accustomed to building them. No other system save Zope itself had such features. And Zope itself was beginning to show signs of its age. We were becoming hampered by consequences of its early design mistakes. Zope’s lack of documentation was also difficult to work around: it was hard to hire smart people to work on Zope applications, because there was no comprehensive documentation set to point them at which explained “it all” in one consumble place, and it was too large and self-inconsistent to document properly. Before repoze.bfg went under development, its authors obviously looked around for other frameworks that fit the bill. But no non-Zope framework did. So we embarked on building repoze.bfg.

As the result of our research, however, it became apparent that, despite the fact that no one framework had all the features we required, lots of existing frameworks had good, and sometimes very compelling ideas. In particular, URL dispatch is a more direct mechanism to map URLs to code.

So although we couldn’t find a framework save for Zope that fit our needs, and while we incorporated a lot of Zope ideas into BFG, we also emulated the features we found compelling in other frameworks (such as url dispatch). After the initial public release of BFG, as time went on, features were added to support people allergic to various Zope-isms in the system, such as the ability to configure the application using imperative configuration and configuration decoration rather than solely using ZCML, and the elimination of the required use of interface objects. It soon became clear that we had a system that was very generic, and was beginning to appeal to non-Zope users as well as ex-Zope users.

As the result of this generalization, it became obvious BFG shared 90% of its featureset with the featureset of Pylons 1, and thus had a very similar target market. Because they were so similar, choosing between the two systems was an exercise in frustration for an otherwise non-partisan developer. It was also strange for the Pylons and BFG development communities to be in competition for the same set of users, given how similar the two frameworks were. So the Pylons and BFG teams began to work together to form a plan to merge. The features missing from BFG (notably view handler classes, flash messaging, and other minor missing bits), were added, to provide familiarity to ex-Pylons users. The result is Pyramid.

The Python web framework space is currently notoriously balkanized. We’re truly hoping that the amalgamation of components in Pyramid will appeal to at least two currently very distinct sets of users: Pylons and BFG users. By unifying the best concepts from Pylons and BFG into a single codebase and leaving the bad concepts from their ancestors behind, we’ll be able to consolidate our efforts better, share more code, and promote our efforts as a unit rather than competing pointlessly. We hope to be able to shortcut the pack mentality which results in a much larger duplication of effort, represented by competing but incredibly similar applications and libraries, each built upon a specific low level stack that is incompatible with the other. We’ll also shrink the choice of credible Python web frameworks down by at least one. We’re also hoping to attract users from other communities (such as Zope’s and TurboGears’) by providing the features they require, while allowing enough flexibility to do things in a familiar fashion. Some overlap of functionality to achieve these goals is expected and unavoidable, at least if we aim to prevent pointless duplication at higher levels. If we’ve done our job well enough, the various audiences will be able to coexist and cooperate rather than firing at each other across some imaginary web framework DMZ.

Pyramid Uses A Zope Component Architecture (“ZCA”) Registry

Pyramid uses a Zope Component Architecture (ZCA) “component registry” as its application registry under the hood. This is a point of some contention. Pyramid is of a Zope pedigree, so it was natural for its developers to use a ZCA registry at its inception. However, we understand that using a ZCA registry has issues and consequences, which we’ve attempted to address as best we can. Here’s an introspection about Pyramid use of a ZCA registry, and the trade-offs its usage involves.

Problems

The global API that may be used to access data in a ZCA component registry is not particularly pretty or intuitive, and sometimes it’s just plain obtuse. Likewise, the conceptual load on a casual source code reader of code that uses the ZCA global API is somewhat high. Consider a ZCA neophyte reading the code that performs a typical “unnamed utility” lookup using the zope.component.getUtility() global API:

1
2
3
from pyramid.interfaces import ISettings
from zope.component import getUtility
settings = getUtility(ISettings)

After this code runs, settings will be a Python dictionary. But it’s unlikely that any civilian would know that just by reading the code. There are a number of comprehension issues with the bit of code above that are obvious.

First, what’s a “utility”? Well, for the purposes of this discussion, and for the purpose of the code above, it’s just not very important. If you really want to know, you can read this. However, still, readers of such code need to understand the concept in order to parse it. This is problem number one.

Second, what’s this ISettings thing? It’s an interface. Is that important here? Not really, we’re just using it as a key for some lookup based on its identity as a marker: it represents an object that has the dictionary API, but that’s not very important in this context. That’s problem number two.

Third of all, what does the getUtility function do? It’s performing a lookup for the ISettings “utility” that should return.. well, a utility. Note how we’ve already built up a dependency on the understanding of an interface and the concept of “utility” to answer this question: a bad sign so far. Note also that the answer is circular, a really bad sign.

Fourth, where does getUtility look to get the data? Well, the “component registry” of course. What’s a component registry? Problem number four.

Fifth, assuming you buy that there’s some magical registry hanging around, where is this registry? Homina homina... “around”? That’s sort of the best answer in this context (a more specific answer would require knowledge of internals). Can there be more than one registry? Yes. So which registry does it find the registration in? Well, the “current” registry of course. In terms of Pyramid, the current registry is a thread local variable. Using an API that consults a thread local makes understanding how it works non-local.

You’ve now bought in to the fact that there’s a registry that is just hanging around. But how does the registry get populated? Why, via code that calls directives like config.add_view. In this particular case, however, the registration of ISettings is made by the framework itself under the hood: it’s not present in any user configuration. This is extremely hard to comprehend. Problem number six.

Clearly there’s some amount of cognitive load here that needs to be borne by a reader of code that extends the Pyramid framework due to its use of the ZCA, even if he or she is already an expert Python programmer and whom is an expert in the domain of web applications. This is suboptimal.

Ameliorations

First, the primary amelioration: Pyramid does not expect application developers to understand ZCA concepts or any of its APIs. If an application developer needs to understand a ZCA concept or API during the creation of a Pyramid application, we’ve failed on some axis.

Instead, the framework hides the presence of the ZCA registry behind special-purpose API functions that do use ZCA APIs. Take for example the pyramid.security.authenticated_userid function, which returns the userid present in the current request or None if no userid is present in the current request. The application developer calls it like so:

1
2
from pyramid.security import authenticated_userid
userid = authenticated_userid(request)

He now has the current user id.

Under its hood however, the implementation of authenticated_userid is this:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
def authenticated_userid(request):
    """ Return the userid of the currently authenticated user or
    ``None`` if there is no authentication policy in effect or there
    is no currently authenticated user. """

    registry = request.registry # the ZCA component registry
    policy = registry.queryUtility(IAuthenticationPolicy)
    if policy is None:
        return None
    return policy.authenticated_userid(request)

Using such wrappers, we strive to always hide the ZCA API from application developers. Application developers should just never know about the ZCA API: they should call a Python function with some object germane to the domain as an argument, and it should returns a result. A corollary that follows is that any reader of an application that has been written using Pyramid needn’t understand the ZCA API either.

Hiding the ZCA API from application developers and code readers is a form of enhancing domain specificity. No application developer wants to need to understand the small, detailed mechanics of how a web framework does its thing. People want to deal in concepts that are closer to the domain they’re working in: for example, web developers want to know about users, not utilities. Pyramid uses the ZCA as an implementation detail, not as a feature which is exposed to end users.

However, unlike application developers, framework developers, including people who want to override Pyramid functionality via preordained framework plugpoints like traversal or view lookup must understand the ZCA registry API.

Pyramid framework developers were so concerned about conceptual load issues of the ZCA registry API for framework developers that a replacement registry implementation named repoze.component was actually developed. Though this package has a registry implementation which is fully functional and well-tested, and its API is much nicer than the ZCA registry API, work on it was largely abandoned and it is not used in Pyramid. We continued to use a ZCA registry within Pyramid because it ultimately proved a better fit.

Note

We continued using ZCA registry rather than disusing it in favor of using the registry implementation in repoze.component largely because the ZCA concept of interfaces provides for use of an interface hierarchy, which is useful in a lot of scenarios (such as context type inheritance). Coming up with a marker type that was something like an interface that allowed for this functionality seemed like it was just reinventing the wheel.

Making framework developers and extenders understand the ZCA registry API is a trade-off. We (the Pyramid developers) like the features that the ZCA registry gives us, and we have long-ago borne the weight of understanding what it does and how it works. The authors of Pyramid understand the ZCA deeply and can read code that uses it as easily as any other code.

But we recognize that developers who might want to extend the framework are not as comfortable with the ZCA registry API as the original developers are with it. So, for the purposes of being kind to third-party Pyramid framework developers in, we’ve drawn some lines in the sand.

In all core code, We’ve made use of ZCA global API functions such as zope.component.getUtility and zope.component.getAdapter the exception instead of the rule. So instead of:

1
2
3
from pyramid.interfaces import IAuthenticationPolicy
from zope.component import getUtility
policy = getUtility(IAuthenticationPolicy)

Pyramid code will usually do:

1
2
3
4
from pyramid.interfaces import IAuthenticationPolicy
from pyramid.threadlocal import get_current_registry
registry = get_current_registry()
policy = registry.getUtility(IAuthenticationPolicy)

While the latter is more verbose, it also arguably makes it more obvious what’s going on. All of the Pyramid core code uses this pattern rather than the ZCA global API.

Rationale

Here are the main rationales involved in the Pyramid decision to use the ZCA registry:

  • History. A nontrivial part of the answer to this question is “history”. Much of the design of Pyramid is stolen directly from Zope. Zope uses the ZCA registry to do a number of tricks. Pyramid mimics these tricks, and, because the ZCA registry works well for that set of tricks, Pyramid uses it for the same purposes. For example, the way that Pyramid maps a request to a view callable using traversal is lifted almost entirely from Zope. The ZCA registry plays an important role in the particulars of how this request to view mapping is done.
  • Features. The ZCA component registry essentially provides what can be considered something like a superdictionary, which allows for more complex lookups than retrieving a value based on a single key. Some of this lookup capability is very useful for end users, such as being able to register a view that is only found when the context is some class of object, or when the context implements some interface.
  • Singularity. There’s only one place where “application configuration” lives in a Pyramid application: in a component registry. The component registry answers questions made to it by the framework at runtime based on the configuration of an application. Note: “an application” is not the same as “a process”, multiple independently configured copies of the same Pyramid application are capable of running in the same process space.
  • Composability. A ZCA component registry can be populated imperatively, or there’s an existing mechanism to populate a registry via the use of a configuration file (ZCML, via the optional pyramid_zcml package). We didn’t need to write a frontend from scratch to make use of configuration-file-driven registry population.
  • Pluggability. Use of the ZCA registry allows for framework extensibility via a well-defined and widely understood plugin architecture. As long as framework developers and extenders understand the ZCA registry, it’s possible to extend Pyramid almost arbitrarily. For example, it’s relatively easy to build a directive that registers several views all at once, allowing app developers to use that directive as a “macro” in code that they write. This is somewhat of a differentiating feature from other (non-Zope) frameworks.
  • Testability. Judicious use of the ZCA registry in framework code makes testing that code slightly easier. Instead of using monkeypatching or other facilities to register mock objects for testing, we inject dependencies via ZCA registrations and then use lookups in the code find our mock objects.
  • Speed. The ZCA registry is very fast for a specific set of complex lookup scenarios that Pyramid uses, having been optimized through the years for just these purposes. The ZCA registry contains optional C code for this purpose which demonstrably has no (or very few) bugs.
  • Ecosystem. Many existing Zope packages can be used in Pyramid with few (or no) changes due to our use of the ZCA registry.

Conclusion

If you only develop applications using Pyramid, there’s not much to complain about here. You just should never need to understand the ZCA registry API: use documented Pyramid APIs instead. However, you may be an application developer who doesn’t read API documentation because it’s unmanly. Instead you read the raw source code, and because you haven’t read the documentation, you don’t know what functions, classes, and methods even form the Pyramid API. As a result, you’ve now written code that uses internals and you’ve painted yourself into a conceptual corner as a result of needing to wrestle with some ZCA-using implementation detail. If this is you, it’s extremely hard to have a lot of sympathy for you. You’ll either need to get familiar with how we’re using the ZCA registry or you’ll need to use only the documented APIs; that’s why we document them as APIs.

If you extend or develop Pyramid (create new directives, use some of the more obscure hooks as described in Using Hooks, or work on the Pyramid core code), you will be faced with needing to understand at least some ZCA concepts. In some places it’s used unabashedly, and will be forever. We know it’s quirky, but it’s also useful and fundamentally understandable if you take the time to do some reading about it.

Pyramid Uses Interfaces Too Liberally

In this TOPP Engineering blog entry, Ian Bicking asserts that the way repoze.bfg used a Zope interface to represent an HTTP request method added too much indirection for not enough gain. We agreed in general, and for this reason, repoze.bfg version 1.1 (and subsequent versions including Pyramid 1.0+) added view predicate and route predicate modifiers to view configuration. Predicates are request-specific (or context -specific) matching narrowers which don’t use interfaces. Instead, each predicate uses a domain-specific string as a match value.

For example, to write a view configuration which matches only requests with the POST HTTP request method, you might write a @view_config decorator which mentioned the request_method predicate:

1
2
3
4
from pyramid.view import view_config
@view_config(name='post_view', request_method='POST', renderer='json')
def post_view(request):
    return 'POSTed'

You might further narrow the matching scenario by adding an accept predicate that narrows matching to something that accepts a JSON response:

1
2
3
4
5
from pyramid.view import view_config
@view_config(name='post_view', request_method='POST',
             accept='application/json', renderer='json')
def post_view(request):
    return 'POSTed'

Such a view would only match when the request indicated that HTTP request method was POST and that the remote user agent passed application/json (or, for that matter, application/*) in its Accept request header.

Under the hood, these features make no use of interfaces.

Many prebaked predicates exist. However, use of only prebaked predicates, however, doesn’t entirely meet Ian’s criterion. He would like to be able to match a request using a lambda or another function which interrogates the request imperatively. In repoze.bfg version 1.2, we acommodate this by allowing people to define custom view predicates:

1
2
3
4
5
6
7
8
9
from pyramid.view import view_config
from pyramid.response import Response

def subpath(context, request):
    return request.subpath and request.subpath[0] == 'abc'

@view_config(custom_predicates=(subpath,))
def aview(request):
    return Response('OK')

The above view will only match when the first element of the request’s subpath is abc.

Pyramid “Encourages Use of ZCML”

ZCML is a configuration language that can be used to configure the Zope Component Architecture registry that Pyramid uses for application configuration. Often people claim that Pyramid “needs ZCML”.

It doesn’t. In Pyramid 1.0, ZCML doesn’t ship as part of the core; instead it ships in the pyramid_zcml add-on package, which is completely optional. No ZCML is required at all to use Pyramid, nor any other sort of frameworky declarative frontend to application configuration.

Pyramid Uses “Model” To Represent A Node In The Graph of Objects Traversed

The repoze.bfg documentation used to refer to the graph being traversed when traversal is used as a “model graph”. A terminology overlap confused people who wrote applications that always use ORM packages such as SQLAlchemy, which has a different notion of the definition of a “model”. As a result, in Pyramid 1.0a7, the tree of objects traversed is now renamed to resource tree and its components are now named resource objects. Associated APIs have been changed. This hopefully alleviates the terminology confusion caused by overriding the term “model”.

Pyramid Does Traversal, And I Don’t Like Traversal

In Pyramid, traversal is the act of resolving a URL path to a resource object in a resource tree. Some people are uncomfortable with this notion, and believe it is wrong. Thankfully, if you use Pyramid, and you don’t want to model your application in terms of a resource tree, you needn’t use it at all. Instead, use URL dispatch to map URL paths to views.

The idea that some folks believe traversal is unilaterally wrong is understandable. The people who believe it is wrong almost invariably have all of their data in a relational database. Relational databases aren’t naturally hierarchical, so traversing one like a tree is not possible.

However, folks who deem traversal unilaterally wrong are neglecting to take into account that many persistence mechanisms are hierarchical. Examples include a filesystem, an LDAP database, a ZODB (or another type of graph) database, an XML document, and the Python module namespace. It is often convenient to model the frontend to a hierarchical data store as a graph, using traversal to apply views to objects that either are the resources in the tree being traversed (such as in the case of ZODB) or at least ones which stand in for them (such as in the case of wrappers for files from the filesystem).

Also, many website structures are naturally hierarchical, even if the data which drives them isn’t. For example, newspaper websites are often extremely hierarchical: sections within sections within sections, ad infinitum. If you want your URLs to indicate this structure, and the structure is indefinite (the number of nested sections can be “N” instead of some fixed number), a resource tree is an excellent way to model this, even if the backend is a relational database. In this situation, the resource tree a just a site structure.

Traversal also offers better composability of applications than URL dispatch, because it doesn’t rely on a fixed ordering of URL matching. You can compose a set of disparate functionality (and add to it later) around a mapping of view to resource more predictably than trying to get the right ordering of URL pattern matching.

But the point is ultimately moot. If you don’t want to use traversal, you needn’t. Use URL dispatch instead.

Pyramid Does URL Dispatch, And I Don’t Like URL Dispatch

In Pyramid, url dispatch is the act of resolving a URL path to a view callable by performing pattern matching against some set of ordered route definitions. The route definitions are examined in order: the first pattern which matches is used to associate the URL with a view callable.

Some people are uncomfortable with this notion, and believe it is wrong. These are usually people who are steeped deeply in Zope. Zope does not provide any mechanism except traversal to map code to URLs. This is mainly because Zope effectively requires use of ZODB, which is a hierarchical object store. Zope also supports relational databases, but typically the code that calls into the database lives somewhere in the ZODB object graph (or at least is a view related to a node in the object graph), and traversal is required to reach this code.

I’ll argue that URL dispatch is ultimately useful, even if you want to use traversal as well. You can actually combine URL dispatch and traversal in Pyramid (see Combining Traversal and URL Dispatch). One example of such a usage: if you want to emulate something like Zope 2’s “Zope Management Interface” UI on top of your object graph (or any administrative interface), you can register a route like config.add_route('manage', '/manage/*traverse') and then associate “management” views in your code by using the route_name argument to a view configuration, e.g. config.add_view('.some.callable', context=".some.Resource", route_name='manage'). If you wire things up this way someone then walks up to for example, /manage/ob1/ob2, they might be presented with a management interface, but walking up to /ob1/ob2 would present them with the default object view. There are other tricks you can pull in these hybrid configurations if you’re clever (and maybe masochistic) too.

Also, if you are a URL dispatch hater, if you should ever be asked to write an application that must use some legacy relational database structure, you might find that using URL dispatch comes in handy for one-off associations between views and URL paths. Sometimes it’s just pointless to add a node to the object graph that effectively represents the entry point for some bit of code. You can just use a route and be done with it. If a route matches, a view associated with the route will be called; if no route matches, Pyramid falls back to using traversal.

But the point is ultimately moot. If you use Pyramid, and you really don’t want to use URL dispatch, you needn’t use it at all. Instead, use traversal exclusively to map URL paths to views, just like you do in Zope.

Pyramid Views Do Not Accept Arbitrary Keyword Arguments

Many web frameworks (Zope, TurboGears, Pylons 1.X, Django) allow for their variant of a view callable to accept arbitrary keyword or positional arguments, which are filled in using values present in the request.POST or request.GET dictionaries or by values present in the route match dictionary. For example, a Django view will accept positional arguments which match information in an associated “urlconf” such as r'^polls/(?P<poll_id>\d+)/$:

1
2
def aview(request, poll_id):
    return HttpResponse(poll_id)

Zope, likewise allows you to add arbitrary keyword and positional arguments to any method of a resource object found via traversal:

1
2
3
4
5
from persistent import Persistent

class MyZopeObject(Persistent):
     def aview(self, a, b, c=None):
         return '%s %s %c' % (a, b, c)

When this method is called as the result of being the published callable, the Zope request object’s GET and POST namespaces are searched for keys which match the names of the positional and keyword arguments in the request, and the method is called (if possible) with its argument list filled with values mentioned therein. TurboGears and Pylons 1.X operate similarly.

Out of the box, Pyramid is configured to have none of these features. By default, pyramid view callables always accept only request and no other arguments. The rationale: this argument specification matching done aggressively can be costly, and Pyramid has performance as one of its main goals, so we’ve decided to make people, by default, obtain information by interrogating the request object within the view callable body instead of providing magic to do unpacking into the view argument list.

However, as of Pyramid 1.0a9, user code can influence the way view callables are expected to be called, making it possible to compose a system out of view callables which are called with arbitrary arguments. See Using a View Mapper.

Pyramid Provides Too Few “Rails”

By design, Pyramid is not a particularly opinionated web framework. It has a relatively parsimonious feature set. It contains no built in ORM nor any particular database bindings. It contains no form generation framework. It has no administrative web user interface. It has no built in text indexing. It does not dictate how you arrange your code.

Such opinionated functionality exists in applications and frameworks built on top of Pyramid. It’s intended that higher-level systems emerge built using Pyramid as a base. See also Pyramid Applications are Extensible; I Don’t Believe In Application Extensibility.

Pyramid Provides Too Many “Rails”

Pyramid provides some features that other web frameworks do not. These are features meant for use cases that might not make sense to you if you’re building a simple bespoke web application:

These features are important to the authors of Pyramid. The Pyramid authors are often commissioned to build CMS-style applications. Such applications are often frameworky because they have more than one deployment. Each deployment requires a slightly different composition of sub-applications, and the framework and sub-applications often need to be extensible. Because the application has more than one deployment, pluggability and extensibility is important, as maintaining multiple forks of the application, one per deployment, is extremely undesirable. Because it’s easier to extend a system that uses traversal from the outside than it is to do the same in a system that uses URL dispatch, each deployment uses a resource tree composed of a persistent tree of domain model objects, and uses traversal to map view callable code to resources in the tree. The resource tree contains very granular security declarations, as resources are owned and accessible by different sets of users. Interfaces are used to make unit testing and implementation substitutability easier.

In a bespoke web application, usually there’s a single canonical deployment, and therefore no possibility of multiple code forks. Extensibility is not required; the code is just changed in-place. Security requirements are often less granular. Using the features listed above will often be overkill for such an application.

If you don’t like these features, it doesn’t mean you can’t or shouldn’t use Pyramid. They are all optional, and a lot of time has been spent making sure you don’t need to know about them up-front. You can build “Pylons-1.X-style” applications using Pyramid that are purely bespoke by ignoring the features above. You may find these features handy later after building a bespoke web application that suddenly becomes popular and requires extensibility because it must be deployed in multiple locations.

Pyramid Is Too Big

“The Pyramid compressed tarball is almost 2MB. It must be enormous!”

No. We just ship it with test code and helper templates. Here’s a breakdown of what’s included in subdirectories of the package tree:

docs/

3.0MB

pyramid/tests/

1.1MB

pyramid/paster_templates/

804KB

pyramid/ (except for pyramd/tests and pyramid/paster_templates)

539K

The actual Pyramid runtime code is about 10% of the total size of the tarball omitting docs, helper templates used for package generation, and test code. Of the approximately 19K lines of Python code in the package, the code that actually has a chance of executing during normal operation, excluding tests and paster template Python files, accounts for approximately 5K lines of Python code. This is comparable to Pylons 1.X, which ships with a little over 2K lines of Python code, excluding tests.

Pyramid Has Too Many Dependencies

This is true. At the time of this writing, the total number of Python package distributions that Pyramid depends upon transitively is 15 if you use Python 2.7, or 17 if you use Python 2.5 or 2.6. This is a lot more than zero package distribution dependencies: a metric which various Python microframeworks and Django boast.

The zope.component, package on which Pyramid depends has transitive dependencies on several other packages (zope.event, and zope.interface). Pyramid also has its own direct dependencies, such as Paste, Chameleon, Mako WebOb, zope.deprecation and some of these in turn have their own transitive dependencies.

We try not to reinvent too many wheels (at least the ones that don’t need reinventing), and this comes at the cost of some number of dependencies. However, “number of package distributions” is just not a terribly great metric to measure complexity. For example, the zope.event distribution on which Pyramid depends has a grand total of four lines of runtime code.

In the meantime, Pyramid has a number of package distribution dependencies comparable to similarly-targeted frameworks such as Pylons 1.X. It may be in the future that we shed more dependencies as the result of a port to Python 3 (the less code we need to port, the better). In the future, we may also move templating system dependencies out of the core and place them in add-on packages, to be included by developers instead of by the framework. This would reduce the number of core dependencies by about five.

Pyramid “Cheats” To Obtain Speed

Complaints have been lodged by other web framework authors at various times that Pyramid “cheats” to gain performance. One claimed cheating mechanism is our use (transitively) of the C extensions provided by zope.interface to do fast lookups. Another claimed cheating mechanism is the religious avoidance of extraneous function calls.

If there’s such a thing as cheating to get better performance, we want to cheat as much as possible. We optimize Pyramid aggressively. This comes at a cost: the core code has sections that could be expressed more readably. As an amelioration, we’ve commented these sections liberally.

Pyramid Gets Its Terminology Wrong (“MVC”)

“I’m a MVC web framework user, and I’m confused. Pyramid calls the controller a view! And it doesn’t have any controllers.”

If you are in this camp, you might have come to expect things about how your existing “MVC” framework uses its terminology. For example, you probably expect that models are ORM models, controllers are classes that have methods that map to URLs, and views are templates. Pyramid indeed has each of these concepts, and each probably works almost exactly like your existing “MVC” web framework. We just don’t use the MVC terminology, as we can’t square its usage in the web framework space with historical reality.

People very much want to give web applications the same properties as common desktop GUI platforms by using similar terminology, and to provide some frame of reference for how various components in the common web framework might hang together. But in the opinion of the author, “MVC” doesn’t match the web very well in general. Quoting from the Model-View-Controller Wikipedia entry:

Though MVC comes in different flavors, control flow is generally as
follows:

  The user interacts with the user interface in some way (for
  example, presses a mouse button).

  The controller handles the input event from the user interface,
  often via a registered handler or callback and converts the event
  into appropriate user action, understandable for the model.

  The controller notifies the model of the user action, possibly
  resulting in a change in the model's state. (For example, the
  controller updates the user's shopping cart.)[5]

  A view queries the model in order to generate an appropriate
  user interface (for example, the view lists the shopping cart's
  contents). Note that the view gets its own data from the model.

  The controller may (in some implementations) issue a general
  instruction to the view to render itself. In others, the view is
  automatically notified by the model of changes in state
  (Observer) which require a screen update.

  The user interface waits for further user interactions, which
  restarts the cycle.

To the author, it seems as if someone edited this Wikipedia definition, tortuously couching concepts in the most generic terms possible in order to account for the use of the term “MVC” by current web frameworks. I doubt such a broad definition would ever be agreed to by the original authors of the MVC pattern. But even so, it seems most MVC web frameworks fail to meet even this falsely generic definition.

For example, do your templates (views) always query models directly as is claimed in “note that the view gets its own data from the model”? Probably not. My “controllers” tend to do this, massaging the data for easier use by the “view” (template). What do you do when your “controller” returns JSON? Do your controllers use a template to generate JSON? If not, what’s the “view” then? Most MVC-style GUI web frameworks have some sort of event system hooked up that lets the view detect when the model changes. The web just has no such facility in its current form: it’s effectively pull-only.

So, in the interest of not mistaking desire with reality, and instead of trying to jam the square peg that is the web into the round hole of “MVC”, we just punt and say there are two things: resources and views. The resource tree represents a site structure, the view presents a resource. The templates are really just an implementation detail of any given view: a view doesn’t need a template to return a response. There’s no “controller”: it just doesn’t exist. The “model” is either represented by the resource tree or by a “domain model” (like a SQLAlchemy model) that is separate from the framework entirely. This seems to us like more reasonable terminology, given the current constraints of the web.

Pyramid Applications are Extensible; I Don’t Believe In Application Extensibility

Any Pyramid application written obeying certain constraints is extensible. This feature is discussed in the Pyramid documentation chapters named Extending An Existing Pyramid Application and Advanced Configuration. It is made possible by the use of the Zope Component Architecture and within Pyramid.

“Extensible”, in this context, means:

  • The behavior of an application can be overridden or extended in a particular deployment of the application without requiring that the deployer modify the source of the original application.
  • The original developer is not required to anticipate any extensibility plugpoints at application creation time to allow fundamental application behavior to be overriden or extended.
  • The original developer may optionally choose to anticipate an application-specific set of plugpoints, which may be hooked by a deployer. If he chooses to use the facilities provided by the ZCA, the original developer does not need to think terribly hard about the mechanics of introducing such a plugpoint.

Many developers seem to believe that creating extensible applications is not worth it. They instead suggest that modifying the source of a given application for each deployment to override behavior is more reasonable. Much discussion about version control branching and merging typically ensues.

It’s clear that making every application extensible isn’t required. The majority of web applications only have a single deployment, and thus needn’t be extensible at all. However, some web applications have multiple deployments, and some have many deployments. For example, a generic content management system (CMS) may have basic functionality that needs to be extended for a particular deployment. That CMS system may be deployed for many organizations at many places. Some number of deployments of this CMS may be deployed centrally by a third party and managed as a group. It’s useful to be able to extend such a system for each deployment via preordained plugpoints than it is to continually keep each software branch of the system in sync with some upstream source: the upstream developers may change code in such a way that your changes to the same codebase conflict with theirs in fiddly, trivial ways. Merging such changes repeatedly over the lifetime of a deployment can be difficult and time consuming, and it’s often useful to be able to modify an application for a particular deployment in a less invasive way.

If you don’t want to think about Pyramid application extensibility at all, you needn’t. You can ignore extensibility entirely. However, if you follow the set of rules defined in Extending An Existing Pyramid Application, you don’t need to make your application extensible: any application you write in the framework just is automatically extensible at a basic level. The mechanisms that deployers use to extend it will be necessarily coarse: typically, views, routes, and resources will be capable of being overridden. But for most minor (and even some major) customizations, these are often the only override plugpoints necessary: if the application doesn’t do exactly what the deployment requires, it’s often possible for a deployer to override a view, route, or resource and quickly make it do what he or she wants it to do in ways not necessarily anticipated by the original developer. Here are some example scenarios demonstrating the benefits of such a feature.

  • If a deployment needs a different styling, the deployer may override the main template and the CSS in a separate Python package which defines overrides.
  • If a deployment needs an application page to do something differently needs it to expose more or different information, the deployer may override the view that renders the page within a separate Python package.
  • If a deployment needs an additional feature, the deployer may add a view to the override package.

As long as the fundamental design of the upstream package doesn’t change, these types of modifications often survive across many releases of the upstream package without needing to be revisited.

Extending an application externally is not a panacea, and carries a set of risks similar to branching and merging: sometimes major changes upstream will cause you to need to revisit and update some of your modifications. But you won’t regularly need to deal wth meaningless textual merge conflicts that trivial changes to upstream packages often entail when it comes time to update the upstream package, because if you extend an application externally, there just is no textual merge done. Your modifications will also, for whatever its worth, be contained in one, canonical, well-defined place.

Branching an application and continually merging in order to get new features and bugfixes is clearly useful. You can do that with a Pyramid application just as usefully as you can do it with any application. But deployment of an application written in Pyramid makes it possible to avoid the need for this even if the application doesn’t define any plugpoints ahead of time. It’s possible that promoters of competing web frameworks dismiss this feature in favor of branching and merging because applications written in their framework of choice aren’t extensible out of the box in a comparably fundamental way.

While Pyramid application are fundamentally extensible even if you don’t write them with specific extensibility in mind, if you’re moderately adventurous, you can also take it a step further. If you learn more about the Zope Component Architecture, you can optionally use it to expose other more domain-specific configuration plugpoints while developing an application. The plugpoints you expose needn’t be as coarse as the ones provided automatically by Pyramid itself. For example, you might compose your own directive that configures a set of views for a prebaked purpose (e.g. restview or somesuch) , allowing other people to refer to that directive when they make declarations in the includeme of their customization package. There is a cost for this: the developer of an application that defines custom plugpoints for its deployers will need to understand the ZCA or he will need to develop his own similar extensibility system.

Ultimately, any argument about whether the extensibility features lent to applications by Pyramid are good or bad is mostly pointless. You needn’t take advantage of the extensibility features provided by a particular Pyramid application in order to affect a modification for a particular set of its deployments. You can ignore the application’s extensibility plugpoints entirely, and instead use version control branching and merging to manage application deployment modifications instead, as if you were deploying an application written using any other web framework.

Zope 3 Enforces “TTW” Authorization Checks By Default; Pyramid Does Not

Challenge

Pyramid performs automatic authorization checks only at view execution time. Zope 3 wraps context objects with a security proxy, which causes Zope 3 to do also security checks during attribute access. I like this, because it means:

  1. When I use the security proxy machinery, I can have a view that conditionally displays certain HTML elements (like form fields) or prevents certain attributes from being modified depending on the the permissions that the accessing user possesses with respect to a context object.
  2. I want to also expose my resources via a REST API using Twisted Web. If Pyramid performed authorization based on attribute access via Zope3’s security proxies, I could enforce my authorization policy in both Pyramid and in the Twisted-based system the same way.

Defense

Pyramid was developed by folks familiar with Zope 2, which has a “through the web” security model. This TTW security model was the precursor to Zope 3’s security proxies. Over time, as the Pyramid developers (working in Zope 2) created such sites, we found authorization checks during code interpretation extremely useful in a minority of projects. But much of the time, TTW authorization checks usually slowed down the development velocity of projects that had no delegation requirements. In particular, if we weren’t allowing untrusted users to write arbitrary Python code to be executed by our application, the burden of through the web security checks proved too costly to justify. We (collectively) haven’t written an application on top of which untrusted developers are allowed to write code in many years, so it seemed to make sense to drop this model by default in a new web framework.

And since we tend to use the same toolkit for all web applications, it’s just never been a concern to be able to use the same set of restricted-execution code under two web different frameworks.

Justifications for disabling security proxies by default notwithstanding, given that Zope 3 security proxies are viral by nature, the only requirement to use one is to make sure you wrap a single object in a security proxy and make sure to access that object normally when you want proxy security checks to happen. It is possible to override the Pyramid traverser for a given application (see Changing the Traverser). To get Zope3-like behavior, it is possible to plug in a different traverser which returns Zope3-security-proxy-wrapped objects for each traversed object (including the context and the root). This would have the effect of creating a more Zope3-like environment without much effort.

Pyramid Uses its Own HTTP Exception Class Hierarchy Rather Than webob.exc

Note

This defense is new as of Pyramid 1.1.

The HTTP exception classes defined in pyramid.httpexceptions are very much like the ones defined in webob.exc (e.g. HTTPNotFound, HTTPForbidden, etc). They have the same names and largely the same behavior and all have a very similar implementation, but not the same identity. Here’s why they have a separate identity:

  • Making them separate allows the HTTP exception classes to subclass pyramid.response.Response. This speeds up response generation slightly due to the way the Pyramid router works. The same speedup could be gained by monkeypatching webob.response.Response but it’s usually the case that monkeypatching turns out to be evil and wrong.
  • Making them separate allows them to provide alternate __call__ logic which also speeds up response generation.
  • Making them separate allows the exception classes to provide for the proper value of RequestClass (pyramid.request.Request).
  • Making them separate allows us freedom from having to think about backwards compatibility code present in webob.exc having to do with Python 2.4, which we no longer support in Pyramid 1.1+.
  • We change the behavior of two classes (HTTPNotFound and HTTPForbidden) in the module so that they can be used by Pyramid internally for notfound and forbidden exceptions.
  • Making them separate allows us to influence the docstrings of the exception classes to provide Pyramid-specific documentation.
  • Making them separate allows us to silence a stupid deprecation warning under Python 2.6 when the response objects are used as exceptions (related to self.message).

Pyramid has Simpler Traversal Machinery than Does Zope

Zope’s default traverser:

  • Allows developers to mutate the traversal name stack while traversing (they can add and remove path elements).
  • Attempts to use an adaptation to obtain the next element in the path from the currently traversed object, falling back to __bobo_traverse__, __getitem__ and eventually __getattr__.

Zope’s default traverser allows developers to mutate the traversal name stack during traversal by mutating REQUEST['TraversalNameStack']. Pyramid’s default traverser (pyramid.traversal.ResourceTreeTraverser) does not offer a way to do this; it does not maintain a stack as a request attribute and, even if it did, it does not pass the request to resource objects while it’s traversing. While it was handy at times, this feature was abused in frameworks built atop Zope (like CMF and Plone), often making it difficult to tell exactly what was happening when a traversal didn’t match a view. I felt it was better to make folks that wanted the feature replace the traverser rather than build that particular honey pot in to the default traverser.

Zope uses multiple mechanisms to attempt to obtain the next element in the resource tree based on a name. It first tries an adaptation of the current resource to ITraversable, and if that fails, it falls back to attempting number of magic methods on the resource (__bobo_traverse__, __getitem__, and __getattr__). My experience while both using Zope and attempting to reimplement its publisher in repoze.zope2 led me to believe the following:

  • The default traverser should be as simple as possible. Zope’s publisher is somewhat difficult to follow and replicate due to the fallbacks it tried when one traversal method failed. It is also slow.
  • The entire traverser should be replaceable, not just elements of the traversal machinery. Pyramid has a few big components rather than a plethora of small ones. If the entire traverser is replaceable, it’s an antipattern to make portions of the default traverser replaceable. Doing so is a “knobs on knobs” pattern, which is unfortunately somewhat endemic in Zope. In a “knobs on knobs” pattern, a replaceable subcomponent of a larger component is made configurable using the same configuration mechanism that can be used to replace the larger component. For example, in Zope, you can replace the default traverser by registering an adapter. But you can also (or alternately) control how the default traverser traverses by registering one or more adapters. As a result of being able to either replace the larger component entirely or turn knobs on the default implementation of the larger component, no one understands when (or whether) they should ever override the larger component entrirely. This results, over time, in a rusting together of the larger “replaceable” component and the framework itself, because people come to depend on the availability of the default component in order just to turn its knobs. The default component effectively becomes part of the framework, which entirely subverts the goal of making it replaceable. In Pyramid, typically if a component is replaceable, it will itself have no knobs (it will be solid state). If you want to influence behavior controlled by that component, you will replace the component instead of turning knobs attached to the component.

Microframeworks Have Smaller Hello World Programs

Self-described “microframeworks” exist: Bottle and Flask are two that are becoming popular. Bobo doesn’t describe itself as a microframework, but its intended userbase is much the same. Many others exist. We’ve actually even (only as a teaching tool, not as any sort of official project) created one using Pyramid (the videos use BFG, a precursor to Pyramid, but the resulting code is available for Pyramid too). Microframeworks are small frameworks with one common feature: each allows its users to create a fully functional application that lives in a single Python file.

Some developers and microframework authors point out that Pyramid’s “hello world” single-file program is longer (by about five lines) than the equivalent program in their favorite microframework. Guilty as charged.

This loss isn’t for lack of trying. Pyramid is useful in the same circumstance in which microframeworks claim dominance: single-file applications. But Pyramid doesn’t sacrifice its ability to credibly support larger applications in order to achieve hello-world LoC parity with the current crop of microframeworks. Pyramid’s design instead tries to avoid some common pitfalls associated with naive declarative configuration schemes. The subsections which follow explain the rationale.

Application Programmers Don’t Control The Module-Scope Codepath (Import-Time Side-Effects Are Evil)

Please imagine a directory structure with a set of Python files in it:

.
|-- app.py
|-- app2.py
`-- config.py

The contents of app.py:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
from config import decorator
from config import L
import pprint

@decorator
def foo():
    pass

if __name__ == '__main__':
    import app2
    pprint.pprint(L)

The contents of app2.py:

1
2
3
4
5
import app

@app.decorator
def bar():
    pass

The contents of config.py:

1
2
3
4
5
  L = []

  def decorator(func):
      L.append(func)
      return func

If we cd to the directory that holds these files and we run python app.py given the directory structure and code above, what happens? Presumably, our decorator decorator will be used twice, once by the decorated function foo in app.py and once by the decorated function bar in app2.py. Since each time the decorator is used, the list L in config.py is appended to, we’d expect a list with two elements to be printed, right? Sadly, no:

[chrism@thinko]$ python app.py
[<function foo at 0x7f4ea41ab1b8>,
 <function foo at 0x7f4ea41ab230>,
 <function bar at 0x7f4ea41ab2a8>]

By visual inspection, that outcome (three different functions in the list) seems impossible. We only defined two functions and we decorated each of those functions only once, so we believe that the decorator decorator will only run twice. However, what we believe is wrong because the code at module scope in our app.py module was executed twice. The code is executed once when the script is run as __main__ (via python app.py), and then it is executed again when app2.py imports the same file as app.

What does this have to do with our comparison to microframeworks? Many microframeworks in the current crop (e.g. Bottle, Flask) encourage you to attach configuration decorators to objects defined at module scope. These decorators execute arbitrarily complex registration code which populates a singleton registry that is a global defined in external Python module. This is analogous to the above example: the “global registry” in the above example is the list L.

Let’s see what happens when we use the same pattern with the Groundhog microframework. Replace the contents of app.py above with this:

1
2
3
4
5
6
7
8
9
from config import gh

@gh.route('/foo/')
def foo():
    return 'foo'

if __name__ == '__main__':
    import app2
    pprint.pprint(L)

Replace the contents of app2.py above with this:

1
2
3
4
5
import app

@app.gh.route('/bar/')
def bar():
    'return bar'

Replace the contents of config.py above with this:

1
2
from groundhog import Groundhog
gh = Groundhog('myapp', 'seekrit')

How many routes will be registered within the routing table of the “gh” Groundhog application? If you answered three, you are correct. How many would a casual reader (and any sane developer) expect to be registered? If you answered two, you are correct. Will the double registration be a problem? With our Groundhog framework’s route method backing this application, not really. It will slow the application down a little bit, because it will need to miss twice for a route when it does not match. Will it be a problem with another framework, another application, or another decorator? Who knows. You need to understand the application in its totality, the framework in its totality, and the chronology of execution to be able to predict what the impact of unintentional code double-execution will be.

The encouragement to use decorators which perform population of an external registry has an unintended consequence: the application developer now must assert ownership of every codepath that executes Python module scope code. Module-scope code is presumed by the current crop of decorator-based microframeworks to execute once and only once; if it executes more than once, weird things will start to happen. It is up to the application developer to maintain this invariant. Unfortunately, however, in reality, this is an impossible task, because, Python programmers do not own the module scope codepath, and never will. Anyone who tries to sell you on the idea that they do is simply mistaken. Test runners that you may want to use to run your code’s tests often perform imports of arbitrary code in strange orders that manifest bugs like the one demonstrated above. API documentation generation tools do the same. Some people even think it’s safe to use the Python reload command or delete objects from sys.modules, each of which has hilarious effects when used against code that has import-time side effects.

Global-registry-mutating microframework programmers therefore will at some point need to start reading the tea leaves about what might happen if module scope code gets executed more than once like we do in the previous paragraph. When Python programmers assume they can use the module-scope codepath to run arbitrary code (especially code which populates an external registry), and this assumption is challenged by reality, the application developer is often required to undergo a painful, meticulous debugging process to find the root cause of an inevitably obscure symptom. The solution is often to rearrange application import ordering or move an import statement from module-scope into a function body. The rationale for doing so can never be expressed adequately in the checkin message which accompanies the fix and can’t be documented succinctly enough for the benefit of the rest of the development team so that the problem never happens again. It will happen again, especially if you are working on a project with other people who haven’t yet internalized the lessons you learned while you stepped through module-scope code using pdb. This is a really pretty poor situation to find yourself in as an application developer: you probably didn’t even know your or your team signed up for the job, because the documentation offered by decorator-based microframeworks don’t warn you about it.

Folks who have a large investment in eager decorator-based configuration that populates an external data structure (such as microframework authors) may argue that the set of circumstances I outlined above is anomalous and contrived. They will argue that it just will never happen. If you never intend your application to grow beyond one or two or three modules, that’s probably true. However, as your codebase grows, and becomes spread across a greater number of modules, the circumstances in which module-scope code will be executed multiple times will become more and more likely to occur and less and less predictable. It’s not responsible to claim that double-execution of module-scope code will never happen. It will; it’s just a matter of luck, time, and application complexity.

If microframework authors do admit that the circumstance isn’t contrived, they might then argue that real damage will never happen as the result of the double-execution (or triple-execution, etc) of module scope code. You would be wise to disbelieve this assertion. The potential outcomes of multiple execution are too numerous to predict because they involve delicate relationships between application and framework code as well as chronology of code execution. It’s literally impossible for a framework author to know what will happen in all circumstances. But even if given the gift of omniscience for some limited set of circumstances, the framework author almost certainly does not have the double-execution anomaly in mind when coding new features. He’s thinking of adding a feature, not protecting against problems that might be caused by the 1% multiple execution case. However, any 1% case may cause 50% of your pain on a project, so it’d be nice if it never occured.

Responsible microframeworks actually offer a back-door way around the problem. They allow you to disuse decorator based configuration entirely. Instead of requiring you to do the following:

1
2
3
4
5
6
7
8
gh = Groundhog('myapp', 'seekrit')

@gh.route('/foo/')
def foo():
    return 'foo'

if __name__ == '__main__':
    gh.run()

They allow you to disuse the decorator syntax and go almost-all-imperative:

1
2
3
4
5
6
7
8
def foo():
    return 'foo'

gh = Groundhog('myapp', 'seekrit')

if __name__ == '__main__':
    gh.add_route(foo, '/foo/')
    gh.run()

This is a generic mode of operation that is encouraged in the Pyramid documentation. Some existing microframeworks (Flask, in particular) allow for it as well. None (other than Pyramid) encourage it. If you never expect your application to grow beyond two or three or four or ten modules, it probably doesn’t matter very much which mode you use. If your application grows large, however, imperative configuration can provide better predictability.

Note

Astute readers may notice that Pyramid has configuration decorators too. Aha! Don’t these decorators have the same problems? No. These decorators do not populate an external Python module when they are executed. They only mutate the functions (and classes and methods) they’re attached to. These mutations must later be found during a scan process that has a predictable and structured import phase. Module-localized mutation is actually the best-case circumstance for double-imports; if a module only mutates itself and its contents at import time, if it is imported twice, that’s OK, because each decorator invocation will always be mutating an independent copy of the object its attached to, not a shared resource like a registry in another module. This has the effect that double-registrations will never be performed.

Routes Need Relative Ordering

Consider the following simple Groundhog application:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
from groundhog import Groundhog
app = Groundhog('myapp', 'seekrit')

app.route('/admin')
def admin():
    return '<html>admin page</html>'

app.route('/:action')
def action():
    if action == 'add':
       return '<html>add</html>'
    if action == 'delete':
       return '<html>delete</html>'
    return app.abort(404)

if __name__ == '__main__':
    app.run()

If you run this application and visit the URL /admin, you will see the “admin” page. This is the intended result. However, what if you rearrange the order of the function definitions in the file?

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
from groundhog import Groundhog
app = Groundhog('myapp', 'seekrit')

app.route('/:action')
def action():
    if action == 'add':
       return '<html>add</html>'
    if action == 'delete':
       return '<html>delete</html>'
    return app.abort(404)

app.route('/admin')
def admin():
    return '<html>admin page</html>'

if __name__ == '__main__':
    app.run()

If you run this application and visit the URL /admin, you will now be returned a 404 error. This is probably not what you intended. The reason you see a 404 error when you rearrange function definition ordering is that routing declarations expressed via our microframework’s routing decorators have an ordering, and that ordering matters.

In the first case, where we achieved the expected result, we first added a route with the pattern /admin, then we added a route with the pattern /:action by virtue of adding routing patterns via decorators at module scope. When a request with a PATH_INFO of /admin enters our application, the web framework loops over each of our application’s route patterns in the order in which they were defined in our module. As a result, the view associated with the /admin routing pattern will be invoked: it matches first. All is right with the world.

In the second case, where we did not achieve the expected result, we first added a route with the pattern /:action, then we added a route with the pattern /admin. When a request with a PATH_INFO of /admin enters our application, the web framework loops over each of our application’s route patterns in the order in which they were defined in our module. As a result, the view associated with the /:action routing pattern will be invoked: it matches first. A 404 error is raised. This is not what we wanted; it just happened due to the order in which we defined our view functions.

This is because Groundhog routes are added to the routing map in import order, and matched in the same order when a request comes in. Bottle, like Groundhog, as of this writing, matches routes in the order in which they’re defined at Python execution time. Flask, on the other hand, does not order route matching based on import order; it reorders the routes you add to your application based on their “complexity”. Other microframeworks have varying strategies to do route ordering.

Your application may be small enough where route ordering will never cause an issue. If your application becomes large enough, however, being able to specify or predict that ordering as your application grows larger will be difficult. At some point, you will likely need to more explicitly start controlling route ordering, especially in applications that require extensibility.

If your microframework orders route matching based on complexity, you’ll need to understand what is meant by “complexity”, and you’ll need to attempt to inject a “less complex” route to have it get matched before any “more complex” one to ensure that it’s tried first.

If your microframework orders its route matching based on relative import/execution of function decorator definitions, you will need to ensure you execute all of these statements in the “right” order, and you’ll need to be cognizant of this import/execution ordering as you grow your application or try to extend it. This is a difficult invariant to maintain for all but the smallest applications.

In either case, your application must import the non-__main__ modules which contain configuration decorations somehow for their configuration to be executed. Does that make you a little uncomfortable? It should, because Application Programmers Don’t Control The Module-Scope Codepath (Import-Time Side-Effects Are Evil).

Pyramid uses neither decorator import time ordering nor does it attempt to divine the relative complexity of one route to another in order to define a route match ordering. In Pyramid, you have to maintain relative route ordering imperatively via the chronology of multiple executions of the pyramid.config.Configurator.add_route() method. The order in which you repeatedly call add_route becomes the order of route matching.

If needing to maintain this imperative ordering truly bugs you, you can use traversal instead of route matching, which is a completely declarative (and completely predictable) mechanism to map code to URLs. While URL dispatch is easier to understand for small non-extensible applications, traversal is a great fit for very large applications and applications that need to be arbitrarily extensible.

“Stacked Object Proxies” Are Too Clever / Thread Locals Are A Nuisance

Some microframeworks use the import statement to get a handle to an object which is not logically global:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
from flask import request

@app.route('/login', methods=['POST', 'GET'])
def login():
    error = None
    if request.method == 'POST':
        if valid_login(request.form['username'],
                       request.form['password']):
            return log_the_user_in(request.form['username'])
        else:
            error = 'Invalid username/password'
    # this is executed if the request method was GET or the
    # credentials were invalid

The Pylons 1.X web framework uses a similar strategy. It calls these things “Stacked Object Proxies”, so, for purposes of this discussion, I’ll do so as well.

Import statements in Python (import foo, from bar import baz) are most frequently performed to obtain a reference to an object defined globally within an external Python module. However, in normal programs, they are never used to obtain a reference to an object that has a lifetime measured by the scope of the body of a function. It would be absurd to try to import, for example, a variable named i representing a loop counter defined in the body of a function. For example, we’d never try to import i from the code below:

1
2
3
def afunc():
    for i in range(10):
        print i

By its nature, the request object created as the result of a WSGI server’s call into a long-lived web framework cannot be global, because the lifetime of a single request will be much shorter than the lifetime of the process running the framework. A request object created by a web framework actually has more similarity to the i loop counter in our example above than it has to any comparable importable object defined in the Python standard library or in normal library code.

However, systems which use stacked object proxies promote locally scoped objects such as request out to module scope, for the purpose of being able to offer users a nice spelling involving import. They, for what I consider dubious reasons, would rather present to their users the canonical way of getting at a request as from framework import request instead of a saner from myframework.threadlocals import get_request; request = get_request() even though the latter is more explicit.

It would be most explicit if the microframeworks did not use thread local variables at all. Pyramid view functions are passed a request object; many of Pyramid’s APIs require that an explicit request object be passed to them. It is possible to retrieve the current Pyramid request as a threadlocal variable but it is a “in case of emergency, break glass” type of activity. This explicitness makes Pyramid view functions more easily unit testable, as you don’t need to rely on the framework to manufacture suitable “dummy” request (and other similarly-scoped) objects during test setup. It also makes them more likely to work on arbitrary systems, such as async servers that do no monkeypatching.

Explicitly WSGI

Some microframeworks offer a run() method of an application object that executes a default server configuration for easy execution.

Pyramid doesn’t currently try to hide the fact that its router is a WSGI application behind a convenience run() API. It just tells people to import a WSGI server and use it to serve up their Pyramid application as per the documentation of that WSGI server.

The extra lines saved by abstracting away the serving step behind run() seem to have driven dubious second-order decisions related to API in some microframeworks. For example, Bottle contains a ServerAdapter subclass for each type of WSGI server it supports via its app.run() mechanism. This means that there exists code in bottle.py that depends on the following modules: wsgiref, flup, paste, cherrypy, fapws, tornado, google.appengine, twisted.web, diesel, gevent, gunicorn, eventlet, and rocket. You choose the kind of server you want to run by passing its name into the run method. In theory, this sounds great: I can try Bottle out on gunicorn just by passing in a name! However, to fully test Bottle, all of these third-party systems must be installed and functional; the Bottle developers must monitor changes to each of these packages and make sure their code still interfaces properly with them. This expands the packages required for testing greatly; this is a lot of requirements. It is likely difficult to fully automate these tests due to requirements conflicts and build issues.

As a result, for single-file apps, we currently don’t bother to offer a run() shortcut; we tell folks to import their WSGI server of choice and run it by hand. For the people who want a server abstraction layer, we suggest that they use PasteDeploy. In PasteDeploy-based systems, the onus for making sure that the server can interface with a WSGI application is placed on the server developer, not the web framework developer, making it more likely to be timely and correct.

Wrapping Up

Here’s a diagrammed version of the simplest pyramid application, where comments take into account what we’ve discussed in the Microframeworks Have Smaller Hello World Programs section.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
from pyramid.response import Response      # explicit response objects, no TL
from paste.httpserver import serve         # explicitly WSGI

def hello_world(request):  # accepts a request; no request thread local reqd
    # explicit response object means no response threadlocal
    return Response('Hello world!')

if __name__ == '__main__':
    from pyramid.config import Configurator
    config = Configurator()       # no global application object.
    config.add_view(hello_world)  # explicit non-decorator registration
    app = config.make_wsgi_app()  # explicitly WSGI
    serve(app, host='0.0.0.0')    # explicitly WSGI

Pyramid Doesn’t Offer Pluggable Apps

It is “Pyramidic” to compose multiple external sources into the same configuration using include(). Any number of includes can be done to compose an application; includes can even be done from within other includes. Any directive can be used within an include that can be used outside of one (such as add_view(), etc).

Pyramid has a conflict detection system that will throw an error if two included externals try to add the same configuration in a conflicting way (such as both externals trying to add a route using the same name, or both externals trying to add a view with the same set of predicates). It’s awful tempting to call this set of features something that can be used to compose a system out of “pluggable applications”. But in reality, there are a number of problems with claiming this:

  • The terminology is strained. Pyramid really has no notion of a plurality of “applications”, just a way to compose configuration from multiple sources to create a single WSGI application. That WSGI application may gain behavior by including or disincluding configuration, but once it’s all composed together, Pyramid doesn’t really provide any machinery which can be used to demarcate the boundaries of one “application” (in the sense of configuration from an external that adds routes, views, etc) from another.
  • Pyramid doesn’t provide enough “rails” to make it possible to integrate truly honest-to-god, download-an-app-from-a-random-place and-plug-it-in-to-create-a-system “pluggable” applications. Because Pyramid itself isn’t opinionated (it doesn’t mandate a particular kind of database, it offers multiple ways to map URLs to code, etc), it’s unlikely that someone who creates something application-like will be able to casually redistribute it to J. Random Pyramid User and have it just work by asking him to config.include a function from the package. This is particularly true of very high level components such as blogs, wikis, twitter clones, commenting systems, etc. The integrator (the Pyramid developer who has downloaded a package advertised as a “pluggable app”) will almost certainly have made different choices about e.g. what type of persistence system he’s using, and for the integrator to appease the requirements of the “pluggable application”, he may be required to set up a different database, make changes to his own code to prevent his application from shadowing the pluggable app (or vice versa), and any other number of arbitrary changes.

For this reason, we claim that Pyramid has “extensible” applications, not pluggable applications. Any Pyramid application can be extended without forking it as long as its configuration statements have been composed into things that can be pulled in via config.include.

It’s also perfectly reasonable for a single developer or team to create a set of interoperating components which can be enabled or disabled by using config.include. That developer or team will be able to provide the “rails” (by way of making high-level choices about the technology used to create the project, so there won’t be any issues with plugging all of the components together. The problem only rears its head when the components need to be distributed to arbitrary users. Note that Django has a similar problem with “pluggable applications” that need to work for arbitrary third parties, even though they provide many, many more rails than does Pyramid. Even the rails they provide are not enough to make the “pluggable application” story really work without local modification.

Truly pluggable applications need to be created at a much higher level than a web framework, as no web framework can offer enough constraints to really make them work out of the box. They really need to plug into an application, instead. It would be a noble goal to build an application with Pyramid that provides these constraints and which truly does offer a way to plug in applications (Joomla, Plone, Drupal come to mind).

Pyramid Has Zope Things In It, So It’s Too Complex

On occasion, someone will feel compelled to post a mailing list message that reads something like this:

had a quick look at pyramid ... too complex to me and not really
understand for which benefits.. I feel should consider whether it's time
for me to step back to django .. I always hated zope (useless ?)
complexity and I love simple way of thinking

(Paraphrased from a real email, actually.)

Let’s take this criticism point-by point.

Too Complex

If you can understand this hello world program, you can use Pyramid:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
from paste.httpserver import serve
from pyramid.config import Configurator
from pyramid.response import Response

def hello_world(request):
    return Response('Hello world!')

if __name__ == '__main__':
    config = Configurator()
    config.add_view(hello_world)
    app = config.make_wsgi_app()
    serve(app)

Pyramid has ~ 650 pages of documentation (printed), covering topics from the very basic to the most advanced. Nothing is left undocumented, quite literally. It also has an awesome, very helpful community. Visit the #pyramid IRC channel on freenode.net (irc://freenode.net#pyramid) and see.

Hate Zope

I’m sorry you feel that way. The Zope brand has certainly taken its share of lumps over the years, and has a reputation for being insular and mysterious. But the word “Zope” is literally quite meaningless without qualification. What part of Zope do you hate? “Zope” is a brand, not a technology.

If it’s Zope2-the-web-framework, Pyramid is not that. The primary designers and developers of Pyramid, if anyone, should know. We wrote Pyramid’s predecessor (repoze.bfg), in part, because we knew that Zope 2 had usability issues and limitations. repoze.bfg (and now Pyramid) was written to address these issues.

If it’s Zope3-the-web-framework, Pyramid is definitely not that. Making use of lots of Zope 3 technologies is territory already staked out by the Grok project. Save for the obvious fact that they’re both web frameworks, Pyramid is very, very different than Grok. Grok exposes lots of Zope technologies to end users. On the other hand, if you need to understand a Zope-only concept while using Pyramid, then we’ve failed on some very basic axis.

If it’s just the word Zope: this can only be guilt by association. Because a piece of software internally uses some package named zope.foo, it doesn’t turn the piece of software that uses it into “Zope”. There is a lot of great software written that has the word Zope in its name. Zope is not some sort of monolithic thing, and a lot of its software is usable externally. And while it’s not really the job of this document to defend it, Zope has been around for over 10 years and has an incredibly large, active community. If you don’t believe this, http://taichino.appspot.com/pypi_ranking/authors is an eye-opening reality check.

Love Simplicity

Years of effort have gone into honing this package and its documentation to make it as simple as humanly possible for developers to use. Everything is a tradeoff, of course, and people have their own ideas about what “simple” is. You may have a style difference if you believe Pyramid is complex. Its developers obviously disagree.

Other Challenges

Other challenges are encouraged to be sent to the Pylons-devel maillist. We’ll try to address them by considering a design change, or at very least via exposition here.