I think the most interesting work in programming languages right now is about the runtime, not syntax or even the languages themselves. Which places PyPy in an interesting position, as they have put a great deal of effort into abstracting out the concept of runtime from the language they are implementing (Python).

There are of course other runtime environments available to Python. The main environment has and continues to be CPython — the runtime developed in parallel with the language, and with continuous incremental feedback and improvement by the Python developer community. It is the runtime that informs and is informed by the language. It’s also the runtime that is most easy-going about integrating with C libraries, and by extension it is part of the vague but important runtime environment of "Unix". There’s also Jython and IronPython. I frankly find these completely uninteresting. They are runtimes controlled by companies, not communities, and the Python implementations are neither natural parts of their runtime environments, nor do the runtimes include many concessions to make themselves natural for Python.

PyPy is somewhere different. It still has a tremendous challenge because Python was not developed for PyPy. Even small changes to the language seem impossible — something as seemingly innocuous as making builtins static seems to be stuck in a conservative reluctance to change. But unlike Jython and IronPython they aren’t stuck between a rock and a hard place; they just have to deal with the rock, not the hard place.

So here is my unsolicited advice on what PyPy-the-runtime should consider. Simple improvements to performance and the runtime are fine, but being incrementally better than CPython only goes so far, and I personally doubt it will ever make a big impact on Python that way.

PyPy should push hard on concurrency and reliability. If it is fast enough then that’s fine; that’s done as far as I’m concerned. I say this because I’m a web programmer, and speed is uninteresting to me. Certainly opinions will differ. But to me speed (as it’s normally defined) is really really uninteresting. When or if I care about speed I’m probably more drawn to Cython. I do care about latency, memory efficiency, scalability/concurrency, resource efficiency, and most of all worst cases. I don’t think a JIT addresses any of these (and can even make things worse). I don’t know of benchmarks that measure these parameters either.

I want a runtime with new and novel features; something that isn’t just incrementally better than CPython. This itself might seem controversial, as the only point to such novel features would be for people to implement at least some code intended for only PyPy. But if the features are good enough then I’m okay with this — and if I’m not drawn to write something that will only work on PyPy, I probably won’t be drawn to use PyPy at all; natural conservatism and inertia will keep me (and most people) on CPython indefinitely.

What do I want?

• Microprocesses. Stackless and greenlets have given us micro-threads, but it’s just not the same. Which is not entirely a criticism — it shows that unportable features are interesting when they are good features. But I want the next step, which is processes that don’t share state. (And implicitly I don’t just want standard async techniques, which use explicit concurrency and shared state.)
• Shared objects across processes with copy-on-write; then you can efficiently share objects (like modules!) across concurrent processes without the danger of shared state, but without the overhead of copying everything you want to share. Lack of this is hurting PHP, as you can’t have a rich set of libraries and share-nothing without killing your performance.
• I’d rather see a break in compatibility for C extensions to support this new model, than to abandon what could be PyPy’s best feature to support CPython’s C extension ecosystem. Being a web programmer I honestly don’t need many C modules, so maybe I’m biased. But if the rest of the system is good enough then the C extensions will come.
• Make sure resource sharing that happens outside of the Python environment is really solid. C libraries are often going to be unfriendly towards microprocesses; make sure what is exposed to the Python environment is solid. That might even mean a dangerous process mode that can handle ctypes and FFI and where you carefully write Python code that has extra powers, so long as there’s a strong wall between that code and "general" code that makes use of those services.
• Cython — it’s doing a lot of good stuff, and has a much more conservative but also more predictable path to performance (through things like type annotation). I think it’s worth leaning on. I also have something of a hunch that it could be a good way to do FFI in a safe manner, as Cython already supports multiple targets (Python 2 and 3) from the same codebase. Could PyPy be another target?
• Runtime introspection of the runtime. We have great language introspection (probably much to the annoyance of PyPy developers who have to copy this) but currently runtime introspection is poor-to-nonexistant. What processes are running? How much memory is each using? Where? Are they holding on to resources? Are they blocking on some non-Python library? How much CPU have they been using? Then I want to be able to kill processes, send them signals, adjust priorities, etc.

And I guess it doesn’t have to be "PyPy", but a new backend for PyPy to target; it doesn’t have to be the only path PyPy pursues.

With a runtime like this PyPy could be an absolutely rocking platform for web development. Python could be as reliable as, oh… PHP? Sorry, I probably won’t win arguments that way ;) As good as Erlang! Maybe we could get the benefits of async without the pain of callbacks or Deferreds. And these are features people would use. Right now I’m perceiving a problem where there’s lots of people standing on the sidelines cheering you on but not actually using PyPy.

So: I wouldn’t tell anyone what to do, and if someone tries this out I’ll probably only be on the sidelines cheering you on… but I really think this could be awesome.

Update: there’s some interesting comments on Hacker News as well.

At PyCon there was an open space about deployment, and the idea of drop-in applications (Java-WAR-style).

I generally get pessimistic about 80% solutions, and dropping in a WAR file feels like an 80% solution to me. I’ve used the Hudson/Jenkins installer (which I think is specifically a project that got WARs on people’s minds), and in a lot of ways that installer is nice, but it’s also kind of wonky, it makes configuration unclear, it’s not always clear when it installs or configures itself through the web, and when you have to do this at the system level, nor is it clear where it puts files and data, etc. So a great initial experience doesn’t feel like a great ongoing experience to me — and it doesn’t have to be that way. If those were necessary compromises, sure, but they aren’t. And because we don’t have WAR files, if we’re proposing to make something new, then we have every opportunity to make things better.

So the question then is what we’re trying to make. To me: we want applications that are easy to install, that are self-describing, self-configuring (or at least guide you through configuration), reliable with respect to their environment (not dependent on system tweaking), upgradable, and respectful of persistence (the data that outlives the application install). A lot of this can be done by the "container" (to use Java parlance; or "environment") — if you just have the app packaged in a nice way, the container (server environment, hosting service, etc) can handle all the system-specific things to make the application actually work.

At which point I am of course reminded of my Silver Lining project, which defines something very much like this. Silver Lining isn’t just an application format, and things aren’t fully extracted along these lines, but it’s pretty close and it addresses a lot of important issues in the lifecycle of an application. To be clear: Silver Lining is an application packaging format, a server configuration library, a cloud server management tool, a persistence management tool, and a tool to manage the application with respect to all these services over time. It is a bunch of things, maybe too many things, so it is not unreasonable to pick out a smaller subset to focus on. Maybe an easy place to start (and good for Silver Lining itself) would be to separate at least the application format (and tools to manage applications in that state, e.g., installing new libraries) from the tools that make use of such applications (deploy, etc).

Some opinions I have on this format, exemplified in Silver Lining:

• It’s not zipped or a single file, unlike WARs. Uploading zip files is not a great API. Geez. I know there’s this desire to "just drop in a file"; but there’s no getting around the fact that "dropping a file" becomes a deployment protocol and it’s an incredibly impoverished protocol. The format is also not subtly git-based (ala Heroku) because git push is not a good deployment protocol.
• But of course there isn’t really any deployment protocol inferred by a format anyway, so maybe I’m getting ahead of myself ;) I’m saying a tool that deploys should take as an argument a directory, not a single file. (If the tool then zips it up and uploads it, fine!)
• Configuration "comes from the outside". That is, an application requests services, and the container tells the application where those services are. For Silver Lining I’ve used environmental variables. I think this one point is really important — the container tells the application. As a counter-example, an application that comes with a Puppet deployment recipe is essentially telling the server how to arrange itself to suit the application. This will never be reliable or simple!
• The application indicates what "services" it wants; for instance, it may want to have access to a MySQL database. The container then provides this to the application. In practice this means installing the actual packages, but also creating a database and setting up permissions appropriately. The alternative is never having any dependencies, meaning you have to use SQLite databases or ad hoc structures, etc. But in fact installing databases really isn’t that hard these days.
• All persistence has to use a service of some kind. If you want to be able to write to files, you need to use a file service. This means the container is fully aware of everything the application is leaving behind. All the various paths an application should use are given in different environmental variables (many of which don’t need to be invented anew, e.g., $TMPDIR).
• It uses vendor libraries exclusively for Python libraries. That means the application bundles all the libraries it requires. Nothing ever gets installed at deploy-time. This is in contrast to using a requirements.txt list of packages at deployment time. If you want to use those tools for development that’s fine, just not for deployment.
• There is also a way to indicate other libraries you might require; e.g., you might lxml, or even something that isn’t quite a library, like git (if you are making a github clone). You can’t do those as vendor libraries (they include non-portable binaries). Currently in Silver Lining the application description can contain a list of Ubuntu package names to install. Of course that would have to be abstracted some.
• You can ask for scripts or a request to be invoked for an application after an installation or deployment. It’s lame to try to test if is-this-app-installed on every request, which is the frequent alternative. Also, it gives the application the chance to signal that the installation failed.
• It has a very simple (possibly/probably too simple) sense of configuration. You don’t have to use this if you make your app self-configuring (i.e., build in a web-accessible settings screen), but in practice it felt like some simple sense of configuration would be helpful.

Things that could be improved:

• There are some places where you might be encouraged to use routines from the silversupport package. There are very few! But maybe an alternative could be provided for these cases.
• A little convention-over-configuration is probably suitable for the bundled libraries; silver includes tools to manage things, but it gets a little twisty. When creating a new project I find myself creating several .pth files, special customizing modules, etc. Managing vendor libraries is also not obvious.
• Services are IMHO quite important and useful, but also need to be carefully specified.
• There’s a bunch of runtime expectations that aren’t part of the format, but in practice would be part of how the application is written. For instance, I make sure each app has its own temporary directory, and that it is cleared on update. If you keep session files in that location, and you expect the environment to clean up old sessions — well, either all environments should do that, or none should.
• The process model is not entirely clear. I tried to simply define one process model (unthreaded, multiple processes), but I’m not sure that’s suitable — most notably, multiple processes have a significant memory impact compared to threads. An application should at least be able to indicate what process models it accepts and prefers.
• Static files are all convention over configuration — you put static files under static/ and then they are available. So static/style.css would be at /style.css. I think this is generally good, but putting all static files under one URL path (e.g., /media/) can be good for other reasons as well. Maybe there should be conventions for both.
• Cron jobs are important. Though maybe they could just be yet another kind of service? Many extra features could be new services.
• Logging is also important; Silver Lining attempts to handle that somewhat, but it could be specified much better.
• Silver Lining also supports PHP, which seemed to cause a bit of stress. But just ignore that. It’s really easy to ignore.

There is a description of the configuration file for apps. The environmental variables are also notably part of the application’s expectations. The file layout is explained (together with a bunch of Silver Lining-specific concepts) in Development Patterns. Besides all that there is admittedly some other stuff that is only really specified in code; but in Silver Lining’s defense, specified in code is better than unspecified ;) App Engine provides another example of an application format, and would be worth using as a point of discussion or contrast (I did that myself when writing Silver Lining).

Discussing WSGI stuff with Ben Bangert at PyCon he noted that he didn’t really feel like the WSGI pieces needed that much more work, or at least that’s not where the interesting work was — the interesting work is in the tooling. An application format could provide a great basis for building this tooling. And I honestly think that the tooling has been held back more by divergent patterns of development than by the difficulty of writing the tools themselves; and a good, general application format could fix that.

All the cool kids love Node.js. I’ve used it a little, and it’s fine; I was able to do what I wanted to do, and it wasn’t particularly painful. It’s fun to use something new, and it’s relatively straight-forward to get started so it’s an emotionally satisfying experience.

There are several reasons you might want to use Node.js, and I’ll ignore many of them, but I want to talk about one in particular:

Javascript on the client and the server!

Is this such a great feature? I think not…

Many years ago I wrote a fairly straight-forward port of Python’s doctest to Javascript. I thought it was cool, but I didn’t really talk about it that much. Especially because I knew it had one fatal flaw: it was very unfriendly towards programming with callbacks, and Javascript uses a lot of callbacks.

On a recent flight I decided to look at it again, and realized fixing that one flaw wasn’t actually a big deal. So now doctest.js really works. And I think it works well: doctest.js (reborn).

I have yet to really use doctest.js on more than a couple real cases, and as I do (or you do?) I expect to tweak it more to make it flow well. But having tried a couple of examples I am particularly liking how it can be used with callbacks.

Testing with callbacks is generally a tricky thing. You want to make assertions, but they happen entirely separately from the test runner’s own loop, and your callbacks may not run at all if there’s a failure.

I came upon some tests recently that used Jasmine, a BDD-style test framework. I’m not a big fan of BDD but I’m fairly new to serious Javascript development so I’m trying to withhold judgement. The flow of the tests is a bit peculiar until you realize that it’s for async reasons. I’ll try to show something that roughly approximates a real test of an XMLHttpRequest API call:

So, the basic pattern is it() creates a group of tests, and each call to run() is a set of items to call sequentially. Then between these run blocks you can have signals to the runner to wait for some result, either a timeout (which is fragile), or you can setup specific conditions.

Another popular test runner is QUnit; it’s popular particularly because it’s what jQuery uses, and my own impression is that QUnit is just very simple and so least likely to piss you off.

QUnit has its own style for async:

stop() confused me for a bit until I realized what they were really referring to stopping the test runner; of course the function continues on regardless. What will happen is that the function will return, but nothing will have really been tested — the success callback will not have been run, and cannot run until all Javascript execution stops and control is given back to the browser. So the test runner will use setTimeout to let time pass before the test continues. In this case it will continue once start() is called. And expect() also makes it fail if it didn’t get at least one assertion during that interval — it would otherwise be easy to simply miss an assertion (though in this example it would be okay because if the success callback isn’t invoked then start() will never be called, and the runner will timeout and signal that as a failure).

So… now for doctest.js. Note that doctest.js isn’t "plain" Javascript, it looks like what an interactive Javascript session might look like (I’ve used shell-style prompts instead of typical console prompts, because the consoles didn’t exist when first I wrote this, and because >>>/... kind of annoy me anyway).

With doctest.js you still get a fairly linear feel — it’s similar to how Jasmine works, except every $ prompt is potentially a place where the loop can be released so something async can happen. Each prompt is equivalent to run() (though unless you call wait, directly or indirectly, everything will run in sequence).

There’s also an implicit assertion for each stanza, which is anything that is written must be matched ({writes: true} makes the spy/mock object write out any invocations). This makes it much harder to miss something in your tests.

I’ve never actually found Python’s doctest to be a particularly good way to write docs, and I don’t expect any different from doctest.js, but I find it a very nice way to write and run tests… and while Python’s doctest is essentially abandoned and lacks many features to make it a more humane testing environment, maybe doctest.js can do better.

I find myself working in a Windows environment due to some temporary problems with my Linux installation. In terms of user experience Windows is not terrible. But more notable, things mostly just feel the same. My computing experience is not very dependent on the operating system… almost. Most of what I do is in a web browser — except programming itself. Probably a lot of you have the same experience: web browser, text editor, and terminal are pretty much all I need. I occasionally play with other tools, but none of them stick. Of course underlying the terminal and text editor UI is a whole host of important software — interpreters, version control tools, checkouts of all my projects, etc. So really there’s two things keeping us from a browser-only world: a few bits of UI, and a whole bunch of tools. Can we bridge this? I’m thinking (more speculatively than as an actual plan): could I stay on Windows without ever having to "use" Windows?

Browsers are clearly capable of implementing a capable UI for a terminal or editor; not a trivial endeavor, but not impossible. We need a way of handling the tools. The obvious answer in that case is a virtual machine. The virtual machine would certainly be using Linux, as there’s clear consensus that if you remove the UI and hardware considerations and just consider tools then Linux is by far the best choice — who uses Mac servers? And Windows is barely worth mentioning. I worked in a Linux VM for a while but found it really unsatisfying — but that was using the Linux UI through a VMWare interface.

So instead imagine: you start up a headless VM (remembering the tools are not about UI, so there’s no reason to have a graphical user interface on the VM), you point your browser at this VM, and you use a browser-based developer environment that mediates all the tools (the lightest kind of mediation is just simulating a terminal and using existing console-based interfaces). Look at your existing setup and just imagine a browser window in place of each not-browser-window app you are using.

I’m intrigued then by the idea of adding more to these interfaces, incrementally. Like HTML in the console, or applications lightly wrapping individual tools. IDEs never stick for me, maybe in part because I can’t commit, and also there’s collaboration issues with these tools (I’m never in a team where we would be able to agree on a single environment). But incremental decentralized improvements seem genuinely workable — improvement more in the style of the web, the browser providing the central metaphor.

I call this a Browser Desktop because it’s a fairly incremental change at this point and other terms (Web OS, Cloud OS) are always presented with unnecessarily hyperbole. What "operating system" you are using in this imagined system is a somewhat uninteresting semantic question; the OS hasn’t disappeared, it’s just boring. "The Cloud" is fine, but too easy to overthink, and there are many technical reasons to use a hybrid of local and remote pieces. "Internet Operating System" is more a framing concept than a thing-that-can-be-built. Chromium OS is essentially the same idea… I’m not really sure how they categorize themselves.

What would be painful right now? Good Javascript terminals exist. Bespin is hard at work on an editor worthy of being used by programmers. The browser needs to be an extremely solid platform. Google Chrome has done a lot in this direction, and Firefox is moving the same direction with the Electrolysis project. It’s okay to punt for now on all the "consumer" issues like music and media handling… and anyway, other people are hard at work on those things. Web sockets will help with some kinds of services that ideally will connect directly to a port; it’s not the same as a raw socket, but I feel like there’s potential for small intermediaries (e.g., imagine a Javascript app that connects to a locally-hosted server-side app that proxies to ssh). Also AddOns can be used when necessary (e.g., ChatZilla <https://addons.mozilla.org/en-US/firefox/addon/16>).

I’d like much better management of all these "apps" aka pages aka windows or tabs — things like split screens and workspaces. Generally I think using such a system heavily will create all sorts of interesting UI tensions. Which might be annoying for the user, but if it’s a constructive annoyance…

On the whole… this seems doable. It’s navel gazing in a sense — programmers thinking about programming — but one good thing about navel gazing is that programmers have traditionally been quite good at navel gazing, and while some results aren’t generally applicable (e.g., VM management) the exercise will certainly create many generally applicable side products. It would encourage interesting itch-scratching. There’s lots of other "web OS" efforts out there, but I’ve never really understood them… they copy desktop metaphors, or have weird filesystem metaphors, or create an unnecessarily cohesive experience. The web is not cohesive, and I’m pretty okay with that; I don’t expect my experiences in this context to be any more cohesive than my tasks are cohesive. In fact it’s exactly the lack of cohesiveness that interests me in this exercise — the browser mostly gives me the level of cohesiveness I want, and I’m open to experimentation on the rest. And maybe the biggest interest for me is that I am entirely convinced that traditional GUI applications are a dead end; they rise and fall (mobile apps being a current rise) but I can’t seriously imagine long term (10 year) viability for any current or upcoming GUI system. I’m certain the browser is going to be along for the long haul. Doing this would let us Live The Future ;)

I’ve yet to see another testing system for local web testing that I like as much as WebTest… which is perhaps personal bias for something I wrote, but then I don’t have that same bias towards everything I’ve written. Many frameworks build in their own testing systems but I don’t like the abstractions — they touch lots of internal things, or skip important steps of the request, or mock out things that don’t need to be mocked out. WSGI can make this testing easy.

There’s also a hidden feature here: because WSGI is basically just describing HTTP, it can be a means of representing not just incoming HTTP requests, but also outgoing HTTP requests. If you are running local tests against your application using WebTest, with just a little tweaking you can turn those tests into HTTP tests (i.e., actually connect to a socket). But doing this is admittedly not obvious; hence this post!

Here’s what a basic WebTest test looks like:

The app object is a wrapper around the WSGI application, and each of those methods runs a request and gets the response. The response object is a WebOb response with several additional helpers for testing (things like .click() which finds a link in HTML and follows it, or .json which loads the body as JSON).

You don’t have to be using a WSGI-centric framework like Pylons to use WebTest, it works fine with anything with a WSGI frontend, which is just about everything. But the point of my post: you don’t have to use it with a WSGI application at all. Using WSGIProxy:

It’s a little crude to control this with an environmental variable ($TEST_REMOTE), but it’s an easy way to pass an option in when there’s no better way (and many test runners don’t make options easy). The extra_environ option puts in the host and scheme information into each request (the default host WebTest puts in is http://localhost). WSGIProxy lets you send a request to any host, kind of bypassing DNS, so SERVER_NAME is actually the server the request goes to, while HTTP_HOST is the value of the Host header.

Going over HTTP there are a couple features that won’t work. For instance, you can pass information about your application back to the test code by putting values in environ['paste.testing_variables'] (which is how you’d make resp.templatevars work in the first example). It’s also possible to use extra_environ to pass information into your application, for example to get your application to mock out user authentication; this is fairly safe because in production no request can put those same special keys into the environment (using custom HTTP headers means you must carefully filter requests in production). But custom environ values won’t work over HTTP.

The thing that got me thinking about this is the work I’m doing on Silver Lining, where I am taking apps and rearranging the code and modifying the database configuration ad setup to fit this deployment system. It would be really nice having done that to be able to run some functional tests, and I really want to run them over HTTP. If an application has tests using something like Selenium or Windmill that would also work great, but those tools can be a bit more challenging to work with and applications still need smaller tests anyway, so being able to reuse tests like these would be most useful.

I’ve been casually perusing Paradigms of Artificial Intelligence Programming: Case Studies in Common Lisp. One of the things I am noticing is that Lisp traditionally has a terrible lack of sentinels: special objects denoting some kind of meaning. Specifically in Common Lisp the empty list and false and nil are all the same thing. As a result there’s all these cases where you want to distinguish false from empty, especially when false represents a failure of some sort. In these AI examples, usually a failure to find something, while in many cases the empty list could mean "the thing is already found, no need to look". But there’s also lots of other examples when this causes problems.

More modern languages usually distinguish between these objects. Python for instance has [], False and None. They might all test as "falsish", but if you care to tell the difference it is easy to do; especially common is a test for x is None. Modern Lisps also stopped folding together all these notions (Scheme for example has #f for false as a completely distinct object, though null and the empty list are still the same). XML-RPC is an example of a language missing null… and though JSON is almost the same data model, it is a great deal superior for having null. In comparison no one seems to care much one way or the other about making a strong distinction between True/False and 1/0.

These are all examples of sentinels: special objects that represent some state. None doesn’t mean anything in particular, but it means lots of things specifically. Maybe it means "not found" in one place, or "give me anything I don’t care" in another. But sometimes you need more than one of these in the same place, or None isn’t entirely clear.

One thing I noticed while reading some Perl 6 examples is that they’ve added a number of new sentinels. One is *. So you could write something like item(*) to mean "give me any item, your choice". While the Perl tendency to use punctuation is legend, words work too.

I wonder if we need a few more sentinel conventions? If so what?

Of course any object can become a sentinel if you use it like that, None isn’t more unique than any other object. (None is conveniently available everywhere.)

Any seems useful, ala Perl’s *. But… there’s already an any available everywhere as well. It happens to be a function, but it’s also a unique named object… would it be entirely too weird to do obj is any? And there’s very few cases where the actual function any would be an appropriate input, making it a good sentinel.

Another WSGI web server benchmark was published. It’s a decent benchmark, despite some criticisms. But it benchmarks what everyone benchmarks: serving up a trivial app really really quickly. This is not very useful to me. Also, performance is not to me the most important differentiation of servers.

In Silver Lining we’re using mod_wsgi. Silver Lining isn’t tied to mod_wsgi (applications can’t really tell), and we may revisit that decision (mostly because of memory concerns), but it is a deliberate choice. mod_wsgi is one of the few multiprocess WSGI servers, and it manages its children (the same way Apache manages all its children). So if a child stops responding, it gets taken out of the pool and killed (brutal efficiency! Or at least brutal terminology). Child processes are also recycled, guarding against memory leaks or other peculiarities. Sometimes these kinds of things are dismissed for covering up bugs, but (a) production is a lousy time to learn about bugs, (b) it’s like a third tier of garbage collection, and (c) the bugs you are avoiding are often bugs you can’t fix anyway (for instance, if your mysql driver leaks memory, is that the application developer’s fault?)

I wish there was competition among servers not to see who can tweak their performance for entirely unrealistic situations, but to see who can implement the most fail-safe server. We’re missing good benchmarks. Unfortunately benchmarks are a pain in the butt to write and manage.

But I hope someone writes a benchmark like that. Here’s some things I’d like to see benchmarked:

• A "realistic" CPU-bound application. for i in xrange(10000000): pass is a reasonable start.
• An application that generates big responses, e.g., "x"*100000.
• An I/O bound application. E.g., one that reads a big file.
• A simply slow application (time.sleep(1)).
• Applications that wedge. while 1: pass perhaps? Or lock = threading.Lock(); lock.acquire(); lock.acquire(). Wedging in C and wedging in Python are different, so a bunch of different kinds of wedging.
• Applications that segfault. ctypes is specially designed for this.
• Applications that leak memory like a sieve, e.g., global_var.extend(['x']*10000).
• Large uploads.
• Slow uploads, like a client that takes 30 seconds to upload 1Mb.
• Also slow downloads.
• In each case it is interesting what happens when something bad happens to just a portion of requests. E.g., if 1% of requests wedge hard. A good container will serve the other 99% of requests properly. A bad container will have its worker pool exhausted and completely stop.
• Mixing and matching these could be interesting. For instance Dave Beazley found some bad GIL results mixing I/O and CPU-bound code.
• Add ideas in the comments and I’ll copy them into this list.

The hardest part of writing this is not the applications (they are simple). One annoyance is wiring up the applications, but handily Nicholas covers that well in his benchmark. You also have to make sure to clean up, as many servers will not exit cleanly from some of the tests. Another nuisance is that some of these require funny clients. These aren’t too hard to write, but you can’t just use ab. Then you have to report.

Anyway: I would love it if someone did this, and packaged it as repeatable/runnable code/scripts. I’ll help some, but I can’t lead. I’d both really like to see the results, and in my ideal world people writing servers would start using these benchmarks to make their servers more robust.

Lately I’ve been writing code using WebOb and just a few other small libraries. It’s not entirely obvious what this looks like, so I thought I’d give a simple example.

I make each application a class. Instances of the class are "configured applications". So it looks a little like this (for an application that takes one configuration parameter, file_path):

Then the app needs to be a WSGI app, because that’s how I roll. I use webob.dec:

Somewhere separate from the application you actually instantiate Application. You can use Paste Deploy for that, configure it yourself, or just do something ad hoc (a lot of mod_wsgi .wsgi files are like this, basically).

I use webob.exc for things like exc.HTTPNotFound(). You can raise that as an exception, but I mostly just return the object (to the same effect).

Now you have Hello World. I then sometimes do something terrible, I start handling URLs like this:

This is lazy and a very bad idea. So you want a dispatcher. There are several (e.g., selector). I’ll use Routes here… the latest release makes it a bit easier (though it could still be streamlined a bit). Here’s a pattern I think makes sense:

Another way you might do it is to skip the class, which means skipping a clear place for configuration. I don’t like that, but if you don’t care about that, then it looks like this:

Then application is pretty much boilerplate. You could put configuration in the request if you wanted, or use some other technique (like Contextual).

I talked some with Ben Bangert about what he’s trying with these patterns, and he’s doing something reminiscent of Pylons controllers (but without the rest of Pylons) and it looks more like this (with my own adaptations):

That’s a lot of code blocks, but they all really say the same thing ;) I think writing apps with almost-no-framework like this is pretty doable, so if you have something small you should give it a go. I think it’s especially appropriate for applications that are an API (not a "web site").

Since I’ve been thinking about deployment I’ve been thinking a lot more about what "configuration management" means, how it should work, what it should do.

I guess my quick summary of configuration management is that it is setting up a server correctly. "Correct" is an ambiguous term, but given that there are so many to configuration management the solutions are also ambiguous.

Silver Lining includes configuration management of a sort. It is very simple. Right now it is simply a bunch of files to rsync over, and one shell script (you can see the files here and the script here — at least until I move them and those links start 404ing). Also each "service" (e.g., a database) has a simple setup script. I’m sure this system will become more complicated over time, but it’s really simple right now, and I like that.

The other system I’ve been asked about the most about lately is Puppet. Puppet is a real configuration management system. The driving forces are very different: I’m just trying to get a system set up that is in all ways acceptable for web application deployment. I want one system set up for one kind of task; I am completely focused on that end, and I care about means only insofar as I don’t want to get distracted by those means. Puppet is for people who care about the means, not just the ends. People who want things to work in a particular way; I only care that they work.

That’s the big difference between Puppet and Silver Lining. The smaller difference (that I want to talk about) is "push" vs. "pull". Grig wrote up some notes on two approaches. Silver Lining uses a "push" system (though calling it a "system" is kind of overselling what it does) while Puppet is "pull". Basically Silver Lining runs these commands (from your personal computer):

This is what happens when you run silver setup-node server: it pushes a bunch of files over to the server, and runs a shell script. If you update either of the files or the shell script you run silver setup-node again to update the server. This is "push" because everything is initiated by the "master" (in this case, the developer’s personal computer).

Puppet uses a pull model. In this model there is a daemon running on every machine, and these machines call in to the master to see if there’s any new instructions for them. If there are, the daemon applies those instructions to the machine it is running on.

Grig identifies two big advantages to this pull model:

1. When a new server comes up it can get instructions from the master and start doing things. You can’t push instructions to a server that isn’t there, and the server itself is most aware of when it is ready to do stuff.
2. If a lot of servers come up, they can all do the setup work on their own, they only have to ask the master what to do.

But… I don’t buy either justification.

First: servers don’t just do things when they start up. To get this to work you have to create custom images with Puppet installed, and configured to know where the master is, and either the image or the master needs some indication of what kind of server you intended to create. All this is to avoid polling a server to see when it comes online. Polling a server is lame (and is the best Silver Lining can do right now), but avoiding polling can be done with something a lot simpler than a complete change from push to pull.

Second: there’s nothing unscalable about push. Look at those commands: one rsync and one ssh. The first is pretty darn cheap, and the second is cheap on the master and expensive on the remote machine (since it is doing things like installing stuff). You need to do it on lots of machines? Then fork a bunch of processes to run those two commands. This is not complicated stuff.

It is possible to write a push system that is hard to scale, if the master is doing lots of work. But just don’t do that. Upload your setup code to the remote server/slave and run it there. Problem fixed!

What are the advantages of push?

1. Easy to bootstrap. A bare server can be setup with push, no customization needed. Any customization is another kind of configuration, and configuration should be automated, and… well, this is why it’s a bootstrap problem.
2. Errors are synchronous: if your setup code doesn’t work, your push system will get the error back, you don’t need some fancy monitor and you don’t need to check any logs. Weird behavior is also synchronous; can’t tell why servers are doing something? Run the commands and watch the output.
3. Development is sensible: if you have a change to your setup scripts, you can try it out from your machine. You don’t need to do anything exceptional, your machine doesn’t have to accept connections from the slave, you don’t need special instructions to keep the slave from setting itself up as a production machine, there’s no daemon that might need modifications… none of that. You change the code, you run it, it works.
4. It’s just so damn simple. If you don’t start thinking about push and pull and other design choices, it simply becomes: do the obvious and easy thing.

In conclusion: push is sensible, pull is needless complexity.