Tag Archives: werkzeug

ANFSCD: Revisiting the Web Server

Nearly two years ago, I was considering which Python web framework to use for a user interface to Postgres: CherryPy, Flask, Werkzeug? Not entirely satisfied with the choices, I started reviewing even more frameworks thinking I might want to write my own minimalist framework.

Several months later, somebody (through Planet Python, IIRC) referred me to a presentation by Jacob Kaplan-Moss on the history and future of Python on the web. Surprisingly, halfway through the talk Jacob started raving about Meteor, a pure JavaScript framework, saying “we’re deluding ourselves if we think this [something like Meteor] is not the future of web applications.” This prompted me to take a close look at Meteor and several other JS frameworks.

Tarek Ziadé’s “A new development era” essay reinforced this change in direction. Ultimately, I settled on AngularJS as the (client) framework. Two-way data binding, dependency injection and testability are some of the features that won me over.

Angular opened the door to the Node.js world—which appears somewhat chaotic compared to Python’s (and even more to the staidness of Postgres). Like Python, Node.js has an abundance of web frameworks, templating libraries and other tools to choose from (and master). Aside from that, are there any negatives in continuing down this path?

For one, although Angular is an open source project, unlike Python and PostgreSQL, its destiny is controlled by a behemoth. A saving grace is its large community of contributors. And perhaps some of Angular’s innovations may eventually become part of standard HTML.

Second, in spite of Selena Deckelmann’s recent comments on JS and PG, I’m strongly partial to Python and not fond of JavaScript as an implementation language. It’s liberating not to have to use braces (and semicolons) for code structure! To compensate, CoffeeScript appears to be the obvious alternative.

When it comes to interfacing to Postgres, although I haven’t explored it enough to do justice, node-postgres doesn’t seem to be up to par with psycopg, and I’m not about to throw away the work I’ve done on Pyrseas, in particular the TTM-inspired interface. So Werkzeug may still play a part, as a Postgres-Python-to-JSON service, particularly now that it support Python 3. However, for contrast I will use node-postgres in an early implementation.

Last, the Angular team’s choice for “workflow” tool (Yeoman) did not sit well with me: I don’t care for “scaffolding” and my first experience with Grunt rubbed me the wrong way. Fortunately, in the Node.js “chaos” I found Brunch, which although not without problems, looks suitable for my purposes.

Having addressed the negatives, I’ve started work on this at GitHub, and plan to post more about it later on.

Update: Due to the change in direction, I was wondering whether I should also change the title of this blog to something like “Taming Serpents, Pachyderms and White A’s in Red Shields”, but fortunately I discovered that at least O’Reilly uses a rhinoceros as the JavaScript mascot and rhinos are considered pachyderms. :-)

Database User Interfaces – Pagination

Since it’s been a while from my last post on this subject, let me recap what we’ve covered:

I also took a detour to explore Python web frameworks. One of the first comments inquired about Tornado. I recently had the opportunity to experiment with it. There’s much to admire, particularly in terms of speed. However, the use of one-or-two-method handler classes essentially for each action (see the blog demo, for example) struck me as counterproductive. It reminded me of Jack Diederich’s “Stop Writing Classes” presentation.

Getting back to database UIs, I have added basic pagination to the listing of films. Here is an example of what it looks like:

The number of lines per page is currently hard-coded (see maxlines in FilmHandler.index), but eventually it should be configurable. The principal additions are the count() and slice() methods in the bl/film.py module. The first does a SELECT COUNT(*) FROM films, so that we can determine how many pages will be needed (the upcoming PostgreSQL 9.2 should improve performance of that query). The slice() method does a SELECT like the all() method but uses LIMIT and OFFSET to retrieve the subset of rows needed for the requested page number.

There are further refinements possible. For example, the list of page numbers/links at the bottom could get very long given enough data in the table. For now, correcting this is left as an exercise for the reader.


1 I’m glad to hear that Kenneth Reitz, Armin Ronacher and others are working on “merging” Werkzeug with Requests.

Python Web Frameworks – Application Object

In response to my previous post, someone asked “… what’s the point? How one way is better than the other?” My response was that I’m “not trying to judge these frameworks from the perspective of a developer (at least not completely), but rather from the POV of ‘if I were writing a framework, which features are essential, which nice to have, which I can disregard'” (emphasis added). Of course, I have my own biases as a developer as to what I consider “nice” or essential.

With that clarification out of the way, let’s look at the next feature: the WSGI callable generally known as the application object.

Django doesn’t want developers to be concerned with this artifact. A Django application is the collection of model, views (controllers in standard MVC terminology) and templates (views) that gets run by the server. For those interested, the Django WSGI callable is an instance of WSGIHandler (in django.core.handlers.wsgi).

As best as I can tell, Twisted Web expects the developer to write the application object. It will get called from the WSGIResponse run() method in twisted.web.wsgi. As its documentation says, Twisted Web is “a framework for doing things with the web” but is “not a ‘web framework’ in the same sense” as Django, so it doesn’t fit well in this comparison.

Pyramid‘s Router class (in pyramid.router) is what implements the WSGI callable. The Router instance gets created from the Configurator class’s make_wsgi_app() method. Select Router attributes and methods:

  • logger
  • root_factory
  • routes_mapper
  • request_factory
  • registry (from args)
  • handle_request()

In CherryPy, the Application class is the application object, although it’s not generallly used directly. Instead a user root object is mounted on a Tree which is a registry of these applications. Major attributes and methods:

  • root (from args): root object (application handler)
  • namespaces, config: for configuration
  • wsgiapp(): a CPWSGIApp instance, to handle apps in a pipeline

Being a support library, Werkzeug doesn’t include an application object, but it has a sample in its Shortly example. Major attributes and methods:

  • dispatch_request(): dispatcher
  • jinja_env: template enviroment
  • url_map: Map() taking list of Rule()’s
  • error_404(): error handler
  • render_template(): template renderer

Like Django, Web2py isn’t interested in having developers worry about the application object. Its WSGI callable is the standalone wsgibase() function, in gluon.main.

The web.py application class (yes, not capitalized) implements its WSGI callable. Major attributes and methods:

  • autoreload (from args)
  • init_mapping (from args)
  • request()
  • handle(): matches path to mapping
  • run(): runs the developer server
  • notfound(), internalerror(): error handlers

Flask‘s Flask class is its application object, derived from flask.helpers._PackageBoundObject. Selected attributes and methods (in addition to those in Werkzeug’s example above):

  • static_url_path, static_folder (from args)
  • instance_path (from args or auto-determined)
  • view_functions
  • before/after_xxx_funcs (dicts)
  • logger
  • run()
  • route(): decorator for URL rules
  • error_handler_spec (a dict)
  • errorhandler(): decorator for error handler functions

Like Werkzeug, WebOb doesn’t include an application object, but it has a sample in its wiki example, most of it very specific to the example.

Bottle‘s Bottle class is its WSGI callable. Major attributes and methods:

  • routes (list) and router (class Router instance)
  • config (class ConfigDict instance)
  • mount(): to mount an app at a given point
  • match(): to search for matching routes
  • route(): decorator to bind a function to a URL

Pesto‘s DispatcherApp class is its application object. Major attributes and methods:

  • prefix (from args): a “mount” point
  • matchpattern(), match(): matches URLs (match is decorator)
  • urlfor(): handler/dispatcher to URL converter
  • gettarget(): returns a four-part tuple from the URI
  • status404_application(): error handler

Diva‘s Application class is the object of interest. Select attributes and methods:

  • config (from kwargs), configure()
  • routing_cfg and routing
  • locale_dir (from args)
  • template_dirs (from args)
  • templates(): template loader
  • prepare() and cleanup(): before and after methods

Summary: For developers who care about writing application code, Django and Web2py and to some extent Pyramid and CherryPy hide the existence of the WSGI callable, while other frameworks require that the programmer instantiate it or invoke it. If I were to write a framework, I’d want to make it easy on the developer as the former projects do, but wouldn’t keep it completely out of sight.

Python Web Frameworks – Development Server

One of the most frequent tasks of a web developer is running the app while it’s being written. So we’ll begin our examination of Python web frameworks by looking at the development servers provided and how does one run the app.

Django developers are used to invoking

python manage.py runserver

The runserver option runs Django’s development server which is invoked via the run() function in django.core.servers.basehttp. This in turn runs a WSGIServer which is derived from Python’s wsgiref’s simple_server.

Twisted Web has its own server based on its core reactor object. Here’s a simple usage example:

from twisted.web import server, resource
from twisted.internet import reactor

class HelloResource(resource.Resource):
    ...

reactor.listenTCP(8080, server.Site(HelloResource()))
reactor.run()

Pyramid uses Paste’s paste.httpserver which in turn is based on Python’s BaseHTTPServer. Sample usage:

from paste.httpserver import serve
from pyramid.config import Configurator

def hello_world(request):
   ... 

if __name__ == '__main__':
   config = Configurator()
   ...
   app = config.make_wsgi_app()
   serve(app, host='0.0.0.0')

In CherryPy, the normal server is cherrypy.engine which is invoked as:

cherrypy.engine.start()
cherrypy.engine.block()

cherrypy.engine is actually an instance of cherrypy.process.wspbus.Bus. A simpler alternative to the engine calls is to use:

cherrypy.quickstart(app())

Under the covers, CherryPy also defines a global server variable that is an instance of Server which implements the HTTP server, with help from cherrypy._cpwsgi_server.CPWSGIServer.

Werkzeug has a built-in server, used as follows:

from werkzeug.serving import run_simple
from myproject import make_app

app = make_app(...)
run_simple('localhost', 8080, app, use_reloader=True)

run_simple() invokes its own (werkzeug.serving) make_server which creates one of three types of XxxWSGIServer’s, which are derived from Python’s BaseHTTPServer.HTTPServer.

When Web2py is invoked, a GUI dialog is shown. The dialog is implemented in gluon.widget and its web2pyDialog.start method creates a main.HttpServer. gluon.main’s HttpServer class uses the Rocket server whose code is in gluon.rocket.

In web.py, to run an application you instantiate a web.application and invoke its run() method, e.g.:

app = web.application(urls, globals())
app.run()

This eventually runs a BaseHTTPServer.HTTPServer from the Python standard library.

Flask‘s WSGI application class (Flask) has a run() method:

run(host='127.0.0.1', port=5000, debug=None, **options)

This invokes the Werkzeug serving.run_simple function (see above).

WebOb doesn’t offer any server per se, but suggests using Paste’s httpserver.serve or wsgiref.simple_server.

Bottle has a standalone run() function:

run(app=None, server='wsgiref', host='127.0.0.1', port=8080,
    interval=1, reloader=False, quiet=False, plugins=None, **kargs)

This runs one of several web servers supported, by default, wsgiref, which uses the wsgiref.simple_server.

Pesto suggests using the Python wsgiref directly, e.g.:

import pesto
dispatcher = pesto.dispatcher_app()
...
httpd = make_server('', 8080, dispatcher)
httpd.serve_forever()

Diva uses Python’s WSGIServer from the standard library to implement its main() and serve() standalone functions. The former can be used as follows:

from diva.core import Application
from diva.server import main

class MyApp(Application): pass

main(MyApp())

Summary: With the exception of Twisted and CherryPy, all frameworks base their development servers on the standard library’s wsgiref or BaseHTTPServer, Paste or Rocket (which claims to be CherryPy-compatible). Lesson for prospective framework creators: Don’t write your own server. However, you may want to enhance your server with auto-reloading, threading, debugging. etc.

There are two approaches to running the app or server. Django and Web2py don’t want the developer to bother with writing app.run() or similar, the other frameworks expect the programmer to plug the various pieces together.

Best wishes for 2012 to all readers!

Python Web Frameworks – Candidates

I’ve been researching web frameworks, partly with a view to writing my own, and thought I’d share my findings. This first post will present the candidates briefly and subsequent posts will delve into particular features.

The candidates are presented based on the number of users as found on Ohloh, as a rough measure of popularity.  Each includes the number of downloads for the latest release on PyPI (sometimes misleading if a package had a recent release) and the salient parts of its advertised description (caveat emptor!). All of them are open source and written in Python, but some of them are WSGI libraries rather than full frameworks.

Django (843 users, 126,426 downloads): A “high-level … Web framework that encourages rapid development and clean, pragmatic design.”

Twisted Web (136 users, 14,413 downloads [both for Twisted]):  An “HTPP server that can be used as a library or run … stand-alone …, an HTML templating engine [and] an HTTP client library.”

Pyramid (aka Pylons on Ohloh, 125 users, 606 + 640 downloads): A “very general … web framework [designed] to make it easier for a developer to create an arbitrary web application.”

CherryPy (75 users, PyPI download info not available): A “minimalist … pythonic, object-oriented web framework.”

Werkzeug (31 users, 4225 downloads): A “WSGI utility library, … [it] is Simple … And Powerful.”

Web2py (21 users, 1157 downloads): A “full-stack framework for rapid development of fast, scalable, secure and portable database-driven web-based applications.”

Web.py (18 users, 11,708 downloads): A “web framework … that is as simple as it is powerful” (hmmm … see Werkzeug above).

Flask (16 users, 42,657 downloads): A “microframework … based on Werkzeug, Jinja 2 and good intentions.”

WebOb (16 users, 29,664 downloads): A “library that provides wrappers around the WSGI request environment, and an object to help create WSGI responses.”

Bottle (4 users, 163 downloads): A “fast, simple and lightweight WSGI micro web-framework.”

Pesto (no users, 421 downloads): Not a framework, but a library for “writing WSGI web applications.”

Diva (not on Ohloh or PyPI): This is Christopher Lenz’s  framework experimentation sandbox: “a lightweight web framework … built on top of WSGI and integrated with … Genshi.” My first, unpublished attempt at writing a framework was based on Diva.

Left Out: I had to draw the line somewhere. Compared to Richard Jones’ micro-framework battle, I included three of what he called “mega frameworks” and excluded three of his micro frameworks. For the record, here are the ones I left out: Aspen, Bobo, CubicWeb, Grok, Itty, Milla, Nagare, Nevow, Pump, Pyjamas, Pylons, Python Paste, Quixote, Spyce, Tornado, TurboGears, Webware, Zope 2 and Zope 3Bluebream. But that’s not all: see the Python.org wiki, WSGI.org and Wikipedia.

Unlike Richard, I will not attempt to develop an application with each framework. Instead, I’ll analyze specific features along the lines of Christopher’s Diva documentation.

A Multi-Layered Test Cake

Dobos layered cake*

A recurrent theme in software engineering are the multiple layers (or tiers) into which systems are subdivided. In the ideal case, a given layer only interacts with the immediate layers “above” and “below” it. In practice, that clean separation of responsibilities is not always possible or, some may argue, desirable.

Nevertheless, thinking in terms of layers is quite useful. For example, the database application I’ve been presenting in this series consists of a PostgreSQL database, psycopg2 (a Python DB-API adapter for PostgreSQL), a business logic module that calls on psycopg2, Werkzeug (a WSGI library), the application modules that call on the business logic and Werkzeug, the HTTP protocol, and the user’s web browser.

As application developers, we need not concern ourselves with the details of how PostgreSQL stores the data that we send to it, nor with testing its internal functions. We rely on PG hackers to cover that testing. Similarly, we rely on Psycopg2 programmers to test communications with PostgreSQL and the Werkzeug team to test its WSGI and HTTP-related functionality. We do have responsibility for testing the business logic, the application code, and the interface presented to the user via the browser.

Tag v0.4.1 of the Database UI Tutorial at GitHub adds a set of unit tests for exercising the business logic module (i.e., werkzeug/bl/film.py). The test module (werkzeug/tests/bl/test_film.py) is pretty straightforward: there is a test method for each method in the business logic class. In addition, there are methods for verifying failure conditions, such as duplicate inserts or deleting a missing record.

Although the unit tests were written after the business logic module (i.e., not in keeping with test-driven development), they helped to resolve an open design issue. Earlier, some methods in dblib.py (thin interface to psycopg2) took an “expected count” argument to check the number of rows affected by a database operation. That didn’t look right, so I had removed it.  While developing the unit tests, it became clear that the checks ought to live in the business logic layer.

Tag v0.4.2 adds unit tests for testing the WSGI or web interface. I was very pleased with how easy it was to use Werkzeug’s testing faciities, Client and BaseResponse, to create the tests: simply “drop” the database app instance into the Client and then call its get or post methods to get the response headers and data. I only had to add the standard library’s xml.dom.minidom to examine particular elements in the HTML body.

Both the business logic and web test modules depend on a utility module which is “reheated” from the Pyrseas version control testing. A nice capability is that the base DbAppTestCase, in its setUpClass method (a Python 2.7 feature), ensures the test database is in sync with the YAML spec.

Coming up: testing the user (browser) interface, via Selenium.


* Dobos cake photograph, courtesy of Bruce Tuten.

Dueling Frameworks, revisited

Werkzeug

In the closing remarks of “Dueling Frameworks”, I hinted that I may continue to explore Flask. That road first led me to take a closer look at the recent state of Werkzeug.

As I had done with Flask and CherryPy, the natural first step was to port the minimalist WSGI database user interface (but taking advantage of the robustness and templating (Jinja2) enhancements) to Werkzeug. This turned out to be very straightforward (see GitHub tag v0.3.2). Key changes:

  • replace my Request/Response objects by Werkzeug’s own
  • add a request argument to the various handler methods
  • simplify the dispatch methods, using request.path and request.form
  • replace the DatabaseApp.__call__ by a much simpler wsgi_app
  • eliminate serve_css (static files are served by Werkzeug’s SharedDataMiddleware)
  • use Werkzeug’s development server.

The second modification (see tag v0.3.3) was to use Werkzeug’s Map and Rule classes to refactor the URL routing. What I liked about this, compared to Flask’s, is that I could keep the delegation of routing within the FilmHandler class.

The final version —for now (see tag v0.3.4), improves the URLs to make them more RESTful. While they’re not completely in agreement with Oliver Charles comments, they’re a step in the right direction. This is something I could not achieve with CherryPy (without bringing in or writing an external module).

The bottom line is that I’ve decided to use Werkzeug for continuing with this tutorial. Perhaps what won me over was its debugger. Also, Werkzeug is a library, not a framework or even a micro-framework, so it feels like a much lighter dependency than CherryPy or Flask. If there is a concern, it is that Werkzeug lacks support for Python 3. This may make a difference —in terms of “traction”— now that WebOb has been ported, but I’m hoping Werkzeug/Flask fans will correct that in the not too distant future.