There will be an IRC meeting for the CherryPy 2.2 roadmap on Thursday, 6pm GMT. Things I want to discuss (most-important items first):
More as I think of them...
[10/25/05 Update: changed some terms to make it more clear.]
At its most basic, a web-application server can be said to map a set of URI's to a set of handlers. From Roy Fielding's REST dissertation:
The resource is a conceptual mapping -- the server receives the identifier (which identifies the mapping) and applies it to its current mapping implementation (usually a combination of collection-specific deep tree traversal and/or hash tables) to find the currently responsible handler implementation and the handler implementation then selects the appropriate action+response based on the request content.
In an HTTP server, the "identifier" is the URI (which includes the query string, as I learned recently). The "handler implementation" is almost always a function in some programming language; for many HTTP servers written with scripting languages, these handlers will be written in the same language as the server. CherryPy 2.1, Django 1.0, and Quixote 2.3 are Python examples of this. mod_python 3.1 is an example of a Python web-application tool where the HTTP server is written in some other language (Apache, written in C). In a moment, we'll take a look at how each of these manages URI-to-handler mappings, which we'll call "dispatch".
Every web-application server, whether tied to a larger framework (Django) or not (CherryPy, Quixote, mod_python), must also address the need for customization. By "customization", I mean modifications to the per-request behavior of the server. I do not mean the behavior of the application, although the same techniques are often employed for both. I also do not mean end-user settings, which are properly stored in an application database. I also want to make it absolutely clear that I don't mean "data which exist in configuration files"—the concept of customization is distinct from the medium.
Now let's look at our four servers, and see how they manage URI-to-handler dispatch, and how they provide customization:
CherryPy takes the "deep tree traversal" route in order to map URI's to handlers. There is a
cherrypy.root object which the developer creates, which always maps to the root URI "/". Subpaths are attached as attributes to the root object. Since the path portion of a URI is also heirarchical, there is a relatively straightforward mapping.
CherryPy allows some flexibility by providing a
default method; if the mapper reaches the end of its search without finding a matching handler, it will then reverse direction, looking for a parent method named "default", which it then calls, passing any child path info as arguments. That is, a URI of "/path/to/parent/child/repr?color=red", if handled by
cherrypy.root.path.to.parent.default, will be called as
default("child", "repr", color="red").
CherryPy manages customization primarily via an internal Python dict (a key->value map); each key is a URI, and each value is another dict of (name, setting) pairs. This is often specified in an "ini-style" config file.
Django might be said to epitomize the "hash tables" approach to handler dispatch, using an ordered set of regular expressions. The urlpatterns object is a tuple of tuples, where each inner tuple is of the form: ("pattern", "handler"). The pattern-handler pairs are evaluated in order until the URI matches a pattern, at which point the handler is looked up (converted from a string to the function which it identifies) and called. By using regular expressions, Django is free to map any set of URI's to a given handler.
Django keeps global customization data in a
settings module; each global variable in that module can be used as a named value. Per-request customization, however, is managed entirely within the handlers, in code.
Quixote has a mapping strategy apparently designed for maximum flexibility. Applications create a Publisher object, to which the server passes each HTTP request. The default Publisher will call
self.root_directory._q_traverse(), passing the value of the
PATH_INFO environment variable (split into chunks by each "/" in the URI). The
_q_traverse method may then "do what it likes" with that path info; the common Directory object tries to map URI's to local methods, or to
_q_traverse methods of successive child objects.
Quixote manages global customizations with a Config class; each attribute of that class is a constant which the server uses to customize its request-response process. The data can be read from a file (more properly, executed from a Python file). But like Django, per-request customization is managed entirely within the handlers, in code.
Mod_python plugs right into Apache, and can control much more of the HTTP conversation than most of the other frameworks. Here, let's just talk about the PythonHandler directive; it's used as follows:
<Location /myapp> SetHandler python-program PythonHandler wsgiref.modpython_gateway::handler </Location>
That is, the mapping between URI's and handlers is performed with Apache's
Files, and similar directives. That's usually not the whole story, however; many modpython applications define few Handlers, or even just one, choosing instead to implement their own additional dispatch and customization layers within those handlers. I believe this tendency is one of the factors which have led so many Python web-framework developers, even the above three, to build on top of modpython as a deployment option.
The only generic, server-provided means of managing customization data for mod_python is the
PythonOption key value directive (although other directives exist and may even be inspected; much of the customization in a modpython application is done entirely within Apache, or via other modules). Each PythonOption applies to the same set of URI's for which the given handler will be invoked.
All of the above designs, as described, have an additional detail in common: they can map multiple URI's to a given handler, but cannot (or tend not to) do the reverse: map a given URI to multiple handlers . This is a surjective, not injective, mapping (click on the image to learn more):
The central question then arises: is per-request customization data bound to the URI's, or to the handlers? Let's answer that for each of our four examples:
In the two frameworks (CherryPy and mod_python) where per-request customization data is associated more closely with URI's, the implementation is declarative as opposed to imperative; the server is free to use the data as it sees fit, in order to meet the perceived goal of the user. In the other two servers, Django and Quixote, the implementation is ad-hoc; developers may choose to use declarative implementations (for example, global constants), or they may "hard-code" the behavior.
This difference shouldn't be a surprise. Django is already a full-stack framework like Ruby on Rails or Spring. CherryPy, in contrast, was designed to (optionally) act as a base component for larger, "full-stack" frameworks like Subway or TurboGears. That is, CherryPy must be customizable both by end-applications and by intermediate frameworks. That would be difficult to achieve with imperative customization. Mod_python goes even further, since Django, Quixote, and even CherryPy can optionally use it to connect to Apache.
We need to pause, here, and make a distinction between application developers and application deployers. For many small applications, these two roles are played by the same person (who cannot understand why everyone is so picky about config architecture ). But for larger applications or megaframeworks, the two are very distinct. Frequently, the following division of roles is expected:
|Server code||Framework developer||App developer|
|Application code||App developer||App developer
(often default values)
Much of this is a direct result of the state of programming tools and languages. For example, "imperative server code" is the domain of framework developers, because only they have CVS/SVN privileges; if anyone else makes changes to that code, they fear losing their changes on the next update (although distributed RCS like Arch or Bazaar, can help ameliorate this a little). Similarly, config files exist outside the CVS/SVN of the application code, and are therefore the only domain of the deployer. But note that config files which use the same language as the application code are often assumed to be too difficult for non-programmers to use.
When developing applications (both new or existing), many developers tend to start with all behaviors embedded in imperative code. After a time, the developers notice a need for varying behaviors, and decide to provide a switch, in code, for it. This may take the form of constant values or subclassing or composition or some other pattern. Once the behavior set is reduced to a small number of variants, control over its customization may be placed in a config file. This results in a fairly predictable vector:
Imperative Code (IC) -> Declarative Code (DC) -> Declarative Text (DT)
Server and framework authors do the same, of course, preferring to start with imperative implementations in code, and moving slowly, but predictably, to declarative implementations in text. And they are right to move slowly; the decisions about where to store and retrieve such data are critical to proper isolation and encapsulation, key ingredients of multi-layered software.
However, what is often not addressed is that the different mechanisms not only implement access control, but directly affect program readability and server architecture, as well. For example, a server author who wishes to make a new, customizable feature available has several options. They may:
In my limited experience, their decision will most likely be motivated by the roles defined in the previous section, and by the "Maturity Vector" above. I'd like to hear from some other authors about their experience. But for now, let's move on to the architectural implications.
CherryPy and Apache-configured-mod_python share a common weakness: customization data is stored in config files, in a declarative language which is not that of the application code. However, developers like to think of customizations as applying to handlers, not to URI's, and they often gripe when the effects of configuration files are divorced in time and space from their corresponding handler code.
For example, CherryPy has a plugin mechanism consisting of filters: classes with a set of methods which are called at various points in the request process. Until version 2.1, all filters were declared in code; each object on the
cherrypy.root tree could define its own
_cpFilterList attribute, a list of filter instances. Such filters apply to that object and its children, and therefore any URI's which map to that object or its children. CherryPy provided some filters in the standard distribution, but many were created by app developers to meet their specific needs.
Beginning in version 2.1, however, the builtin filters changed their declaration method, from an in-code list to configuration files, and therefore, changed from being associated with handlers to being associated with URI's. Interestingly, user-defined filters did not. Developers, especially CherryPy veterans, therefore, become confused between the two mechanisms. When configuration is bound to URI's instead of handlers, it is easy to delay the actual handler dispatch; in some cases, it may be delayed too long, limiting the customization which developers can perform, both in the handlers themselves and via any filter/plugin hooks provided by the server.
The effect is not limited to the filters themselves. Since some app developers see filters as a catch-all for customization needs, they place customizations in filters which don't really belong there, because they naively expect user-defined filters to be automatically declarable and configurable via the config file (they're not). On the other hand, filters are perceived to be black magic by many app developers, and some behaviors are hard-coded in handlers which belong more properly in filters (Python decorators seem to be a Strange Attractor for this).
Another problem arises because text-config-file declarations in both of these tools map to URI's, and not handlers. Server and app developers are at a loss in those rare cases when they need to allow deployers to specifically customize a handler as opposed to a URI. For example, a deployer for a site which internally redirects an arbitrary number of paths (e.g. http://www.site.com/~[user]/help -> /help) to a common, customizable handler would much rather write a single config entry, but the config file format forces them to write one entry for each virtual path.
Django and Quixote, by contrast, seem to handle all per-request customization in code, imperatively. At least, there is no central, server-managed repository for declarative settings—any app developer could decide to implement their own; some do. Some simply expect customization to be done directly on the source files (example).
The first problem with this approach is that deployers (who are not programmers) have a significant psychological, if not educational, hurdle to overcome when configuring their copy of the application. [This isn't a religious treatise, so I'll stop there.]
The second problem is that it's difficult to really extend the framework itself. It's not expected, of course, that anyone would write a framework on top of Django, but neither is there any facility for extending per-request Django behavior, other than in imperative code. That is, if a site needed to gzip only some of their HTTP responses, a developer would have to implement that behavior, for each handler, in imperative code.
Finally, when customization is bound to handlers instead of URI's, the timing of the handler dispatch is of utmost importance; it must occur very early in the request-handling process, so that any per-request customization data can be available as soon as possible. Often, the mapping from URI to handler must be done absolutely first, so that the server itself has access to such data, even before calling the main handler. Django, for example, resolves the URI to the handler right away; the only serious action it takes first is to call global middleware (which has no per-request config). Quixote takes a hybrid approach; the
_q_traverse methods act as "server code" (providing dispatch, and possibly configuration) but are instantiated in application objects.
The choice of where to store per-request customization data in a web application server is never a trivial one. It is constrained by project maturity and social expectations, as well as the architecture of both the server and each application. These concerns often compete; occasionally, they produce unresolvable conflicts.
When designing a web application server, the design of any configuration system is of utmost importance, and will affect the design of the entire server and any applications or frameworks which are built on the server. If a configuration system is not flexible, it may resist (or deny) applying the server to some application domains, or limit the extensibility of the server.
The customization system must also be designed to work in concert with the handler-dispatch mechanism. Customizations of all kinds should be analyzed and explicitly designed to be bound to either handlers or URI's. Pretending that handlers and URI's are synonymous will only hide implementation conflicts, delaying them from design time to deployment time.
Any feedback on this document is welcome. I'd like to learn a lot more about this from other server/framework authors' experiences, and from the analysis of any developers and/or deployers. Add your comments below, and I'll work on keeping this document updated.
 Mod_python has the best facilities for doing this, since many Apache handlers are designed to be run in series, or to cascade. An enterprising Quixote developer could, in theory, write a Publisher which called multiple handlers (but then, any of these tools' handlers could implement their own additional layer of dispatch, as well). But the vast majority of applications tend to remain surjective.
Sylvain just reminded me of one of Ryan Tomayko's early rants on HTTP and REST, On HTTP Abuse. It was probably the one post that jump-started my exploration of REST, which has been guiding my contributions to CherryPy.
In that post, he presented a short list of things which a "web framework" should provide:
For instance, which frameworks ...
- ... help implement content negotiation properly?
- ... provide facilities for implementing a smart caching strategy for dynamic content? (proper use of If-Modified-Since, Expires, Cache-Control, etc.)
- ... make dealing with media types easy?
- ... make dealing with character encodings easy?
- ... encourage the use of standard HTTP authentication schemes?
- ... have a sane mechanism for attaching different behavior to different verbs for a single resource?
- ... help ensure that URIs stay cool?
- ... make dealing with transfer encodings (gzip, compress, etc.) easy?
- ... help you use response status codes properly? (e.g. Nearly all dynamic content returns either a 200 or 500).
Even if CherryPy isn't a framework, it should do most of these. The latest release of CherryPy, version 2.1, addresses some of these (3, 4, 5, 8, and 9). The others are possible, but not as easy as they could be; 1, 2, and 6 are top priorities (for me, anyway) to work into version 2.2. Items 3 and 4 could use more work, too.
I just discovered I can set the name of a bookmark to nothing. This should free up a bunch of my Browser Toolbar real-estate, for sites that are nice enough to make a distinctive favicon. There's certainly no need for me to name the bookmark for Questionable Content as "QC", when the icon is already a black box with "QC" in white letters. I've also done away with names for Google, OneLook, and CherryPy.
Guido van Rossum recently wrote:
Python, in its design philosophy, tries hard not to be a framework. (This sets it apart from Java, which is hostile to non-Java code.) Python tries to be helpful when you want to solve part of your problem using a different tool. It tries to work well even if Python is only a small part of your total solution. It tries to be agnostic of platform-specific frameworks, optionally working with them (e.g. fork and pipes on Unix) but not depending or relying on them. Even threads are quite optional to Python.
Oddly enough, this is how I feel about CherryPy, that it tries hard not to be a framework. It tries to be helpful, recognizing that it's most likely only part of your solution. It tries to be agnostic of templating and persistence systems, and has little to say about markup languages, content-types, site architecture, or RPC formats.
Guido was responding to Phillip J. Eby, who wrote:
A Pythonic framework shouldn't load you down with new management burdens and keep you from using other frameworks. It should make life easier, and make your code more interoperable, not less. Indeed, I've pretty much come to agreement with the part of the Python developer community that has says Frameworks Are Evil.
Not wanting to be Evil, I've tried to make CherryPy 2.1 a system which doesn't load you down with new management burdens. Instead, it exposes the functionality of HTTP by presenting it in a Pythonic way. I and many others think it makes life easier—CherryPy appears to have an underscore-shaped learning curve. And as for interoperability, CherryPy was one of the first Python web-application servers to grow a WSGI interface.
I suppose that CherryPy will always have to bear the moniker of "framework", if only because it calls your code, instead of the other way around. But let's keep it a Pythonic framework as long as we can.
Thanks, Kevin! Can we get a new one when CP 2.1 final is released? Please?
And there was much rejoicing.
Ryan (my fellow IT worker), Chris (his girlfriend), and I took a well-deserved five days and went backpacking on the Pine Ridge Trail in the Los Padres National Forest (Ventana Wilderness).
On Thursday, we left San Diego at 1:00 AM so that we could get a full day's hiking in. It's over 10 miles to Sykes Camp, plenty of time to reacquaint oneself with California's gorgeous trees:
Once we arrived at Sykes, we partook of the hot springs. Here's Chris standing in a very small one:
The next day, we didn't move camp—just wandered up the Big Sur River in search of cool things. We found quite a lot of them, but I don't have any pictures to show you, and I probably would hide them from you if I had any. No sense making such a fantastic place too popular. Ryan found a turtle, maybe this pic will tide you over:
Saturday, we decided to hike on over to Cienega Camp, which sits above the North Fork of the Big Sur River. Here's a panorama of the North Fork valley:
That day...wasn't the best. Cienega is six miles away from Sykes, but the last three miles are a tough fight through thick brush. Wear pants if you ever try it. I didn't. Once we got there, we found Cienega to be little more than a wide spot in the trail, so we turned around and backtracked the three miles to Redwood Creek Camp, which was much nicer:
In the second picture, I hope you can see that the forest floor is a long way down. The "small" trees in the gaps are the same size as the near trees, just much further away.
Sunday, we continued our return trip, this time passing Sykes and stopping for the night at Barlow Flats Camp. Once there, Ryan and I decided to wade down the Big Sur River. I wanted to see the point where Logwood Creek fed out into the river, and once there, I cajoled Ryan into scrambling up its whole length until it rejoined the trail. Sure glad we did; this waterfall was 20 feet high, and fed into a pool at least 20 feet deep:
This is only one of the many cool waterfalls we navigated on our way up.
In Barlow Flats Camp, unlike the other camps, the nearly-full moon was visible, and astoundingly bright. I spent a lot of time trying to get the perfect moon shots:
On Monday, we headed back to the parking lot, having traveled about 40 miles in all. Here's a last pic of the trail which I particularly liked, due to its spiral structure:
We left the park about 1:30 PM, stopped for pizza in Monterey, and tried to head home. After road closures, rain, flash floods, accidents, wrong turns, and too many leftover snacks, we hit San Diego about midnight! But I got a neat picture of the moon over Tehachapi. Note that, for all of these moon pictures, I used anywhere from a 4 to 15 second exposure, and never had a tripod—just me holding the camera as steady as I could.
To Dan, I just have to say, we should have gone to Los Padres instead of Kern last month.
|<< <||> >>|