Lix: A step forward in dependency management for Haxe projects

Part 1: Haxelib gets painful

I’ve been using Haxe for a long while, and for about 2-3 years I was using Haxe full time, building web applications in Haxe, so I know how important managing your dependencies is, and I know how painful it was with Haxelib, especially if you had a lot of dependencies, a lot of projects, or needed to collaborate with people on different computers.

Haxelib is okay when you’re just installing one or two libraries, and they’re libraries with stable releases where you don’t change versions often, and if you don’t need to come back to your code after long gaps in time. Basically, haxelib is fine if you’re doing weekend hackathons or contests like Ludum Dare where your projects probably aren’t too complex, you’re not collaborating with too many other people, you’re using existing frameworks, and you don’t have to worry about if it will still work fine in 4 months time.  Otherwise, it can be quite painful.

I tried to help with Haxelib at one point in time (I still am in the top 4 contributors on Github, most of that was back in 2013 though), but it proved pretty unruley – even skilled developers were afraid of changing too much or refactoring in a way that might break things for thousands of developers. And some changes were impossible to make without first changing the Haxe compiler. So it’s largely sat in the “too hard” basket and has not had many meaningful improvement since it first became it’s own project in 2013.

(No offense to anyone who has been working on it – you are a braver soul than I! But I think we all agree it’s not as good as it needs to be.)

Since mid 2016, I have been working in other jobs where I don’t use Haxe full time, instead spending more time with JS: using tools like NPM, Yarn, Webpack. And they’re certainly not perfect when it comes to dependency management, but there are a few things that they do right (Yarn especially).

Part 2: What the JS ecosystem gets right.

In Node JS land (and eventually normal JS land), there was a package manager called NPM – Node Package Manager. It had a registry of packages you could install. It would also let you install a package from Github or somewhere else. The basic things.

Here’s what I think it did right:

  • Used a standard format (package.json) to describe which packages a project uses.
  • Put all of the libraries in a standard location (node_modules/${my_cool_lib}/)
  • NodeJS didn’t care if you used NPM or not. As long as your stuff was in node_modules, it would be happy.

Why was this a good move? Because it allowed some talented people to build a competitor to NPM, called Yarn. By having simple expectations, you can have two competing package managers, and innovation can happen. Woo!

Yarn is what I use at work on a big project with 119 dependencies (and about 1000 sub-depedencies). Here’s what yarn did right:

  • Reproducible builds. While package.json has information about which version I want (say, React 16.* or above), Yarn would keep information in a file called yarn.lock which says exactly which version I ended up using (say, React 16.0.1). This was when my friend joins the project and tries to install things she won’t accidentally end up on a newer or older version than me – Yarn makes sure we’re all using exactly the same version, and all of our dependencies and sub-dependencies are also exactly the same.
  • A global cache. When Yarn came out, it was several times faster than NPM on our project because it kept a cache of dependencies and was able to resolve them quickly when switching between projects and branches. NPM has caught up now – but that’s the benefit of competition!

Part 3: Introducing lix (and its friends: switchx and haxeshim)

In 2015 I remember chatting to my friend Juraj Kirchheim (also one of the key contributors who just kind of gave up) about what an alternative might be, and he described something that sounded great, a futuristic utopian alternative to haxelib.

2 years later, and it turns out, it’s been built! And it’s called “lix”.

(What’s with the name? I’m guessing it is short for “LIbraries in haXe”, a leftover from when every Haxe project needed an X in it for cool-ness, and Haxe was spelt as haXe. That, and the lix.pm domain name must have been available).

Lix also depends on two other projects: haxeshim and switchx. The names aren’t super obvious, so here is my understanding of how it all works:

  • Haxe Shim intercepts calls to Haxe and does some magic. The Haxe compiler on it’s own explicitly calls haxelib, so you literally can’t replace haxelib without intercepting all calls to the compiler and getting rid of -lib arguments. So haxeshim is a shim that intercepts Haxe calls and sorts out -lib arguments so that haxelib is never needed.

    As a bonus, it also supports switching to the right version of Haxe for the current project. But for that, we also need “switchx”.

  • SwitchX lets you pick the Haxe version you need for your project, and automatically switches Haxe versions for whatever project you’re in. If you change between project A, on Haxe 3.4.3, and project B, in a different folder and running Haxe 4, it will always use the correct one.

    How?

    When you start a project you run switchx scope create. This makes a .haxerc file which says that this folder is a specific project, or “scope”, and should use the Haxe version defined in the .haxerc file.

    How do you change the version?

    You run switchx use latestor switchx use stable or switchx use nightly or switchx use 3.4.3 etc. It lets you instantly switch between different versions, and for the correct version to always be used while you’re in your project folder.

    Nice!

  • Lix is a package manager that you use to install packages. It is made to work with Haxe Shim, and creates a “haxe_libraries” folder, with a new hxml file for each dependency you install. It’s super fast because it uses a global cache (like yarn) and it makes sure you always have the correct version installed (like yarn). It supports installing dependencies from Haxelib, Github, Gitlab or HTTP (zip file). Anytime you update or change a dependency, one of the haxe_libraries/*.hxml files will be updated, you commit this change to Git, and it will update for all of your coworkers as well. Magic.

These tools are (for now) built on top of NodeJS, so you can install them with NPM or Yarn.

If you want to install each of these, you basically run these commands (warning: these will replace your current Haxe installation):

# Install all 3 tools and make their commands available.
yarn global add haxeshim switchx lix.pm

# Create a ".haxerc" in the current directory, informing haxeshim that
# this project should use a specific version of Haxe and specific
# `haxe_libraries` dependencies
switchx scope create

# Use the latest stable version of Haxe in this project.
switchx install stable

Part 4: What lix can do that haxelib cannot do (well).

With this setup, here’s what I can do that I couldn’t do before:

Be certain that I always have the exact right version installed, even if the project is being set up on someone else’s machine. Even if I pulled from a custom branch, using something like lix install github:haxetink/tink_web#pure (install the latest version of tink_web from the “pure” branch), when I run this on a different machine, it will use not only the same branch, but the exact same commit that it used on my machine, so we will be compiling the exact same code.

Easily get up and running on a machine where they don’t even have haxe installed. I tried this today – took a project on Linux, and set up its dependencies in Lix. It used a combination of Haxelib, Github, Gitlab, and custom branches. It was a nightmare with Haxelib. I also added haxeshim, switchx and lix.pm as “devDependencies” so they would be installed locally when I ran yarn intall. I opened a Windows machine that had Git installed, but not haxe, cloned the repo, and yarn install. It installed all of the yarn dependencies, including haxeshim, switchx, and lix, and then running lix download installed all of the correct “haxe_libraries”, and then everything compiled. Amazing!

Know if I’ve changed a dependency Today I was working on a change for haxe-react. In the past I would have used haxelib dev react /my/path/to/react-fork/. Now I edit haxe_libraries/react.hxml and change the class path to point to the folder my fork lives in. The great thing about doing this is, Git notices that I’ve changed it. And so when I go to commit the work on my project, git lets me know I’ve got a change to “react.hxml”, I’ve changed that dependency. In this case, I knew what to do: push my fork to Github, and then run lix install gh:jasononeil/haxe-react#react16 to get Lix to properly register my fork in a way that will work with my project going forward. I then commit the change, and people who use my project will get the up-to-date fork.

Start a competing package manager The great thing about all of this, is that “lix” has some great features, but if I want to write better ones, I can. Because of the way “haxeshim” just expects dependencies to haxe a “haxe_libraries/*.hxml” file, I could write my own package manager, that does things in my own way, and just places the right hxml file in the right place, and I’m good to go. This makes it possible to have multiple, competing package managers. Or even multiple, co-operating package managers.

Part 5: Vote on the future

So, I think Lix has learnt from a lot of what has gone “right” in the NodeJS ecosystem, and built a great tool for the Haxe ecosystem. I love it, and will definitely be using it in my Haxe projects going forward.

The question is, do we really need “haxeshim” and “switchx” and other such tools just in order to have a competing package manager? For now sadly, because of the way haxe and haxelibare tied at the hip, you do need a hack like this. But there’s a discussion to change that. (See here and here).

If you care about Haxe projects having maintainable dependency management, you can help by voting up comments in a discussion that’s happening right now. Here are the comments that I think will help Haxe support something like Lix, and more competing package managers, as first class citizen going forward. Feel free to upvote with a thumbs up emoji:

https://github.com/HaxeFoundation/haxe-evolution/issues/30#issuecomment-333298948https://github.com/HaxeFoundation/haxe-evolution/issues/30#issuecomment-333299543https://github.com/HaxeFoundation/haxe-evolution/issues/30#issuecomment-333302625https://github.com/HaxeFoundation/haxe-evolution/issues/30#issuecomment-333311522https://github.com/HaxeFoundation/haxe-evolution/issues/30#issuecomment-333323976

And two of my own comments

https://github.com/HaxeFoundation/haxe-evolution/issues/30#issuecomment-333367950https://github.com/HaxeFoundation/haxe-evolution/issues/30#issuecomment-333368097

Feel free to have a look and contribute to the discussion. For now though – if you don’t mind installing haxeshim and switchx, there is a very good solution for managing your haxelibs and dependencies in a reliable, consistent, but still flexible way. And it’s called Lix.

Juggling haxelibs between multiple projects

Once you have more than one project you’re building in Haxe, you tend to run into situations where you use different versions of dependencies.  Often you can get away with using the latest version on every project, but sometimes there are compatibility breaks, and you need different projects to use different versions.

There is a work-in-progress issue for Haxelib to get support for per-project repositories.  Until that is finished, here is what I do:

cd ~/workspace/project1/

mkdir haxelibs

haxelib setup haxelibs

haxelib install all

And then when I switch between projects:

cd ~/workspace/project2/

haxelib setup haxelibs

What this does:

  • Switch to your current project
  • Create a folder to store all of your haxelibs for this project in
  • Set haxelib to use that folder (and when I switch to a different project, I’ll use a different local folder).
  • Install all the dependencies for this project.

Doing this means that each project can have it’s own dependencies, and upgrading a library in one project doesn’t break the compile on another project.

Hopefully that helps someone else, and hopefully the built in support comes soon!

Accept either String or Int, without resorting to Dynamic

A quick code sample:

`AcceptEither`, a way to accept either one type or another, without resorting to “Dynamic”, and still have the compiler type-check everything and make sure you correctly handle every situation.

Try it on try.haxe.org

The next step would be having Accept2Types, Accept3Types etc, and have a macro automatically build the code for whichever option you ask for.

Super Quick Haxe Abstracts Example

I just posted a quick gist to the Haxe mailing list showing how one way that abstracts work.

They’re a great way to wrap a native API object (in this case, js.html.AnchorElement) without having to create a new wrapper object every single time. Which means they’re great for performance, the end result code looks clean, and thanks to some of the other abstract magic (implicit casts, operator overloading etc) there is a lot of cool things you can do.

Have a look at the sample, read the Haxe manual, and let me know what you think or if you have questions :)

My Haxe Log: Week #2

This week:

  • My token one day a week I continued on my Node-Webkit project.  This time I made externs for Kue (externs, which appear to be working) and FFMpeg (externs, not functional just yet).  Still enjoying working with Node-Webkit, and with the Node-API library especially.  Sad I didn’t get to make more progress on it this week.
  •  Ufront:
      • Make tracing / logging work reliably between multiple requests. After enabling neko.Web.cacheModule(), I began to find areas where Ufront was not very multiple-request-friendly. These would have surfaced later with a port to Client JS or Node JS, but it’s good to find them now.One problem was that our tracing and logging modules were behaving as if there was only one request at a time. This could result in a trace message for one request ending up being output to somebody else’s request, which is obviously bad.

        The problem is a tricky one, as trace() always translates to haxe.Log.trace(), and you with Ufront’s multiple-request-at-a-time design, you can’t know which request is the current one from a static method. If I think of a clever way to do it, possibly involving cookies and sessions, then I might include a HttpContext.getCurrentContext() static method. This would probably have to be implemented separately for each supported platform.

        The solution for now, however, was to not keep track of log messages in the TraceModule, but in the HttpContext. Then on the onLogRequest event, the trace modules get access to the log messages for the current context, and can output them to the browser, to a file, or whichever they choose.

        The downside is that you have to use httpContext.ufTrace() rather than trace(). I added a shortcut for this in both ufront.web.Controller and ufront.remoting.RemotingApiClass, so that in your controllers or APIs you can call uftrace() and it will be associated with the current request. There is also ufLog, ufWarn and ufError.

        I also made RemotingModule work similarly with tracing and logging – so logs go to both the log file and the remoting call to the browser.

      • Fix logging in ErrorModule. One of the things that made debugging the new ufront really hard was that when there was an Error, the ErrorModule displayed, but the trace messages did not get sent to the browser or the log file. I did a bit of code cleanup and got this working now.
      • Fixed File Sessions / EasyAuth. Once able to get my traces and logs working more consistently, I was able to debug and fix the remaining issues in FileSession, so now EasyAuth is working reliably, which is great.
      • Added Login / Logout for UF-Admin. With UF-Admin, I added a login screen and logout, that works with EasyAuth for now. I guess I will make it pluggable later… For now though it means you can set up a simple website, not worry about auth for the front end, but have the backend password protected. If you use EasyAuth for your website / app, the same session will work on the ufadmin page.
      • Created uf-content for app-generated files. I moved all app-generated files (sessions, logs, temp files etc) into a folder called “uf-content”. Then I made this configurable, and relative to httpContext.request.scriptDirectory. You can configure it by changing the contentDirectory option in your UfrontConfiguration. This will make it easier when deploying, we can have instructions to make that single directory writeable but not accessible via the web, and then everything that requires FileSystem access can work reliably from there.
      • Pushed new versions of the libraries. Now that the basics are working, I pushed new versions of the libraries to Haxelib. They are marked as ufront-* with version 1.0.0-beta.1. From here it will be easy to update them individually and move towards a final release.
      • Demo Blog App. To demonstrate the basics of how it works, and a kind of “best practices” for project structure, I created a demo app, and thought I would start with a blog. I started, and the basic setup is there, including the config structure and each of the controller actions, and the “ufadmin” integration. But it’s not working just yet, needs more work.
      • Identified Hair website. I have a website for a friend’s small business that I’ve been procrastinating working on for a long time. On Saturday I finally got started on it, and set up the basic project and routes in Ufront. In about 4 hours I managed to get the project set up, all the controllers / routes working, all the content in place and a basic responsive design with CSS positioning working. All the data is either HTML, Markdown or Database Models (which get inserted into views). Once I’ve got their branding/graphics included, I’ll use ufront to provide a basic way to change data in their database. And then if they’re lucky, I might look at doing some Facebook integration to show their photo galleries on the site.

My Haxe Log: Week #1

Every week as part of my work and as part of my free time I get to work on Haxe code, and a lot of that is contributing to libraries, code, blog posts etc.  Yesterday was one of those frustrating days where I told someone I’d finish a demo ufront app and show them how it works, but I just ran into problem after problem and didn’t get it done, and was feeling pretty crap about it.

After chatting it out I looked back at my week and realised: I have done alot.  So I thought I should start keeping a log of what I’ve been working on – mostly for my own sake, so I can be encouraged by the progress I have made, even if I haven’t finished stuff yet.  But also in case anything I’m working on sparks interest or discussion – it’s cool to have people know what I’m up to.

So I’d like to start a weekly log.  It may be one of those things I do once and never again, or it may be something I make regular: but there’s no harm in doing it once.

So here we go, my first log.  In this case, it’s not just this week, some of it requires me to go back further to include things I’ve been working on, so it’s a pretty massive list:

  • Node Webkit: On Monday’s I work at Vose Seminary, a tertiary college, and I help them get their online / distance education stuff going.  Editing videos, setting up online Learning Management Systems etc.  I have a bunch of command line utils that make the video editing / exporting / transcoding / uploading process easier, but I want to get these into graphics so other staff can use it.  Originally I was thinking of using OpenFL / StablexUI.  I’m far more comfortable with the JS / Browser API than the Flash API however, and so Node-Webkit looked appealing.  On Monday I made my first Haxe-NodeJS project in over a year, using Clement’s new Node-API repo.  It’s beautiful to work with, and within an hour and a half I had written some externs and had my first “hello-world” Node-Webkit app.  I’ll be working on it again this coming Monday.
  • neko.Web.cacheModule: I discovered a way to get a significant speed-up in your web-apps.  I wrote a blog post about it.
  • Ufront: I’ve done a lot of work on Ufront this week.  After my talk at WWXthis year, I had some good chats with people and basically decided I was going to undertake a major refactor of ufront.  I’m almost done!  Things I’ve been working on this week (and the last several weeks, since it all ties in together):
    • Extending haxe.web.Dispatch (which itself required a pull request) to be subclassed, and allowing you to 1) execute the ‘dispatch’ and ‘executeAction’ steps separately and 2) allow returning a result, so that you can get the result of your dispatch methods.  This works much nicer with Ufront’s event based processing, and allows for better unit testing / module integration etc.  The next step is allowing dispatch to have asynchronous handlers (for Browser JS and Node JS).  I began thinking through how to implement this also.
    • After discovering neko.Web.cacheModule, I realised that it had many implications for Ufront.  Basically: You can use static properties for anything that is generic to the whole application, but you cannot use it for anything specific to a request.  This led to several things breaking – but also the opportunity for a much better (and very optimised) design.
    • IHttpSessionState, FileSession: the first thing that was broken was the FileSession module.  The neko version was implemented entirely using static methods, which led to some pretty broken behaviour once caching between requests was introduced.  In the end I re-worked the interface “IHttpSessionState” to be fairly minimal, and was extended by the “IHttpSessionStateSync” and “IHttpSessionStateAsync” interfaces, so that we can begin to cater for Async platforms.  I then wrote a fresh FileSession implementation that uses cookies and flat-files, and should work across both PHP and Neko (and in Future, Java/C#).  The JS target would need a FileSessionAsync implementation.
    • IAuthHandler / EasyAuth: At the conference I talked about how I had an EasyAuth library that implemented a basic User – Group – Permission model.  At the time, this also was implemented with Static methods.  Now I have created a generic interface (IAuthHandler) so that if someone comes up with an auth system other than EasyAuth, it can be compatible.  I also reworked EasyAuth to be able to work with different a) IHttpSessionState implementations and b) different IAuthAdapter’s – basically, this is an interface that just has a single method: `authenticate()`.  And it tells you if the user is logged in or not.  EasyAuth by default uses EasyAuthDBAuthAdapter, which compares a username and password against those in the database.  You could also implement something that uses OpenID, or a social media logon, or LDAP, or anything.  All this work trying to make it generic enough that different implementations can co-exist I think will definitely pay off, but for now it helps to have a well thought out API for EasyAuth :)
    • YesBoss: Sometimes you don’t want to worry about authentication.  Ufront has the ability to create a “tasks.n” command line file, which runs tasks through a Command Line Interface, rather than over the web.  When doing this, you kind of want to assume that if someone has access to run arbitrary shell commands, they’re allowed to do what they want with your app.  So now that I have a generic interface for checking authentication, I created the “YesBossAuthHandler” – a simple class that can be used wherever an authentication system is needed, but any permission check it always lets you pass.  You’re the boss, after all.
    • Dependency Injection: A while ago, I was having trouble understanding the need for Dependency Injection.  Ufront has now helped me see the need for it.  In the first app I started making with the “new” ufront, I wanted to write unit tests.  I needed to be able to jump to a piece of code – say, a method on a controller – and test it as if it was in a real request, but using a fake request.  Dependency injection was the answer, and so in that project I started using Minject.  This week, realising I had to stop using statics and singletons in things like sessions and auth handling, I needed a way to get hold of the right objects, and dependency injection was the answer.  I’ve now added it as standard in Ufront.  There is an `appInjector`, which defines things that should be injected everywhere (modules, controllers, APIs etc).  For example, injecting app configuration or a caching module or an analytics API.  Then there is the dispatchInjector, which is used to inject things into controllers, and the remotingInjector, which is used to inject things into APIs during remoting calls.  You can define things you want to make available at your app entry point (or your unit test entry point, or your standalone task runner entry point), and they will be available when you need them.  (As a side note, I now also have some great tools for mocking requests and HttpContexts using Mockatoo).
    • Tracing: Ufront uses Trace Modules.  By default it comes with two: TraceToBrowser and TraceToFile.  Both are useful, but I hadn’t anticipated some problems with the way they were designed.  In the ufront stack, modules exist at the HttpApplication level, not at the HttpRequest level.  On PHP (or uncached neko), there is little difference.  Once you introduce caching, or move to a platform like NodeJS – this becomes a dangerous assumption.  Your traces could end up displaying on somebody else’s request.  In light of this, I have implemented a way of keeping track of trace messages in the HttpContext.  My idea was to then have the Controller and RemotingApiClass have a trace() method, which would use the HttpContext’s queue.  Sadly, an instance `trace()` method currently does not override the global `haxe.Log.trace()`, so unless we can get that fixed (I’m chatting with Simon about it on IRC), it might be better to use a different name, like `uftrace()`.  For now, I’ve also made a way for TraceToBrowser to try guess the current HttpContext, but if multiple requests are executing simultaneously this might break.  I’m still not sure what the best solution is here.
    • Error Handling: I tried to improve the error handling in HttpApplication.  It was quite confusing and sometimes resulted in recursive calls through the error stack.  I also tried to improve the visual appearance of the error page.
    • Configuration / Defaults:The UfrontApplication constructor was getting absurd, with something like 6 optional parameters.  I’ve moved instead to having a `UfrontConfiguration` typedef, with all of the parameters, and you can supply, all, some or none of the parameters, and fall-backs will be used if needed.  This also improves the appearance of the code from:new UfrontApplication( true, “log.txt”, Dispatch.make(new Routes()) );

      to

      new UfrontApplication({
      urlRewrite: true,
      dispatchConf: Dispatch.make( new Routes() ),
      logFile: “log.txt”
      });

    • More ideas: last night I had trouble getting to sleep.  Too many ideas.  I sent myself 6 emails (yes 6) all containing new ideas for Ufront.  I’ll put them on the Ufront Trello Board soon to keep track of them.  The ideas were about Templating (similar abstractions and interfaces I have here, as well as ways of optimising them using caching / macros), an analytics module, a request caching module and setting up EasyAuth to work not only for global permissions (CanAccessAdminArea), but also for item-specific permissions: do you have permission to edit this blog post?
    • NodeJS / ClientJS: after using NodeJS earlier in the week, I had some email conversations with both Clement and Eric about using Ufront on NodeJS.  After this week it’s becoming a lot more obvious how this would work, and I’m getting close.  The main remaining task is to support asynchronous calls in these 3 things: Dispatch action execution calls, HttpRemotingConnection calls, and database calls – bringing some of the DB Macros magic to async connections.  But it’s looking possibly now, where as it looked very difficult only 3 months ago.
  • CompileTime: I added a simple CompileTime.interpolateFile() macro.  It basically reads the contents of the file at macro time, and inserts it directly into the code, but it inserts it using String Interpolation, as if you had used single quotes.  This means you can insert basic variables or function calls, and they will all go in.  It’s like a super-quick and super-basic poor-man’s templating system.  I’m already using it for Ufront’s default error page and default controller page.
  • Detox: this one wasn’t this week, but a couple of weeks ago.  I am working on refactoring my Detox (DOM / Xml Manipulation) library to use Abstracts.  It will make for a much more consistent API, better performance, and some cool things, like auto-casting strings to DOM elements:”div.content”.find().append( “<h1>My Content</h1>” );
  • My Work Project: Over the last two weeks I’ve updated SMS (my School Management System project, the main app I’ve been working on) to use the new Ufront.  This is the reason I’ve been finding so much that needs to be updated, especially trying to get my app to work with “ufront.Web.cacheModule”.

neko.Web.cacheModule()

Until now I haven’t had to worry much about the speed of sites made using Haxe / Ufront,  – none of the sites or apps I’ve made have anywhere near the volume for it to be a problem, and the general performance was fast enough that no one asked questions. But I’m going to soon be a part of building the new Haxe website, which will have significant volume.

So I ran some benchmarks using ab (Apache’s benchmarking tool), and wasn’t initially happy with the results. They were okay, but not significantly faster than your average PHP framework. Maybe I would have to look at mod_tora or NodeJS for deployment.

Then I remembered something: a single line of code you can add that vastly increases the speed: neko.Web.cacheModule(main).

Benchmarks

Here is some super dumb sample code:

class Server {
  static var staticInt = 0;
  static function main() {
    #if neko
      neko.Web.cacheModule(main); // comment out to test the difference
    #end 
    var localInt = 0; 
    trace ( ++staticInt ); 
    trace ( ++localInt ); 
  } 
} 

And I am testing with this command:

ab -n 1000 -c 20 http://localhost/ 

Here are my results (in requests/second on my laptop):

  • Apache/mod_php (no cache): 161.89
  • NekoTools server: 687.49
  • Apache/mod_neko (no cache): 1054.70
  • Apache/mod_tora (no cache): 745.94
  • Apache/mod_neko (cacheModule): 3516.04
  • Apache/mod_tora (cacheModule): 2185.30

First up: I assume mod_tora has advantages on sites that use more memory, but a dummy sample like this is more overhead than it’s worth.

Second, and related: I know these tests are almost worthless, we really need to be testing a real app, with file access and template processing and database calls.

Let’s do that, same command, same benchmark parameters:

  • Apache/mod_php (no cache): 3.6 (ouch!)
  • NekoTools server: 20.11
  • Apache/mod_neko (no cache): 48.74
  • Apache/mod_tora (no cache): 33.29
  • Apache/mod_neko (cacheModule): 351.42
  • Apache/mod_tora (cacheModule): 402.76

(Note: PHP has similar caching, using modules like PHP-APC. I’m not experienced setting these up however, and am happy with the neko performances I’m seeing so I won’t investigate further)

Conclusions:

  • the biggest speed up (in my case) seems to come from cacheModule(), not mod_tora. I believe once memory usage increases significantly, tora brings advantages in that arena, and so will be faster due to less garbage collection.
  • this could be made faster, my app currently has very little optimisation:
    • the template system uses Xml, which I assume isn’t very fast.
    • a database connection is required for every request
    • there is no caching (memcached, redis etc)
    • I think I have some terribly ineffecient database queries that I’m sure I could optimise
  • Ufront targeting Haxe/PHP is not very fast out-of-the-box. I’m sure you could optimise it, but it’s not there yet.
  • This is running on my laptop, not a fast server. Then again, my laptop may be faster than a low end server, not sure.

Usage

So, how does it work?

#if neko neko.Web.cacheModule( main ); #end 

The conditional compilation (#if neko and #end) is just there so that you can still compile to other targets without getting errors. The cacheModule function has the following documentation:

Set the main entry point function used to handle requests.
Setting it back to null will disable code caching.

By entry point, it is usually going to mean the main() function that is called when your code first runs. So when the docs ask for a function to use as the entry point, I just use main, meaning, the static function main() that I am currently in.

I’m unsure of the impact of having multiple “.n” files or a different entry point.

The cache is reset whenever the file timestamp changes: so when you re-compile, or when you load a new “.n” file in place.

If you wanted to manually disable the cache for some reason, you use cacheModule(null). I’m not sure what the use case is for this though… why disable the cache?

Gotchas (Static variable traps with cacheModule)

The biggest gotcha is that static variables persist in your module. They are initialized just once, which is a big part of the speed increase. Let’s look at the example code I posted before:

class Server {
  static var staticInt = 0;
  static function main() {
    #if neko
      neko.Web.cacheModule(main); // comment out to test the difference
    #end 
    var localInt = 0; 
    trace ( ++staticInt ); 
    trace ( ++localInt ); 
  } 
} 

With caching disabled, both trace statements will print “0” every time. With caching enabled, the staticInt variable does not get reset – it initializes at 0, and then every single page load will continue to increment it, it will go up and up and up.

What does this mean practically:

  • If you want to cache stuff, put it in a static variable. For example:
    • Database connections: store them in a static variable and the connection will persist.
    • Templates: read from disk once, store them in a static variable
    • App Config, especially if you’re passing JSON or Xml, put it in a static and it stays cached.
  • Things which should be unique to a request, don’t store in a static variable. For example:
    • Ufront has a class called NekoSession, which was entirely static methods and variables, and made assumptions that the statics would be reset between requests. Wrong! Having the session cached between different requests (by different users) was a disaster – everytime you click a link you would find yourself logged in as a different user. Needless to say we needed to refactor this and not use statics :) To approach it differently, you could use a static var sessions:StringMap<SessionID, SessionData> and actually have it work appropriately as long as the cache stayed alive.
    • Avoid singletons like Server.currentUser, or even User.current – these static variables are most likely going to be cached between requests leading to unusual results.

Language and API Design

Paul Graham has written some great essays over time, and most of them hold up surprisingly well as they age.  Take this article, “Five Questions About Language Design“.  Written in May 2001.  That’s over 12 years ago from today as I read it, and many of his predictions have come to pass (the dominance of web applications, the rise in popularity of niche programming languages), and most of his logic is still sound, and still insightful.

Two particular points I found interesting, and would be interesting to the Haxe community I spend most of my time in:

4. A Language Has to Be Good for Writing Throwaway Programs.

You know what a throwaway program is: something you write quickly for some limited task. I think if you looked around you’d find that a lot of big, serious programs started as throwaway programs. I would not be surprised if most programs started as throwaway programs. And so if you want to make a language that’s good for writing software in general, it has to be good for writing throwaway programs, because that is the larval stage of most software.

I think Haxe is beginning to show it’s strength here in gaming.  People are using it for Ludlam Dare (write a game in a single weekend), and then later the game scales up to be a full-featured, multi-platform, often commercial game.  Haxe is good for a quick win (especially when targetting flash), but then you have the options to target so many other platforms as your vision for your game/app expands.

I’d love to see ufront get this good.  I have so many things I would like to use as an app, but which I don’t feel justify an ongoing subscription to an online company.  Many of them could get an MVP (minimal viable product) out in a weekend, if the framework was good enough.  And it’s getting good enough, but still a little way to go :)

And then this:

2. Speed Comes from Profilers.

Language designers, or at least language implementors, like to write compilers that generate fast code. But I don’t think this is what makes languages fast for users. Knuth pointed out long ago that speed only matters in a few critical bottlenecks. And anyone who’s tried it knows that you can’t guess where these bottlenecks are. Profilers are the answer.

Language designers are solving the wrong problem. Users don’t need benchmarks to run fast. What they need is a language that can show them what parts of their own programs need to be rewritten. That’s where speed comes from in practice. So maybe it would be a net win if language implementors took half the time they would have spent doing compiler optimizations and spent it writing a good profiler instead.

Nice insight.  I sometimes worry that the code I’m writing will be slow.  It’s hard to know if optimisations are worth it, especially if it may compromise code readability etc.  A better solution would be to focus on making profiling easy.  On the JS target you can easily use the browser tools to show profiling information.  Another option is to bake it into your Haxe code, something that the “massivecover” lets you do.  Wiring that support from massivecover into ufront or other frameworks, and generating pretty reports, would be a very cool way to encourage people to write fast apps.

You cannot use @:build inside a macro : make sure that your enum is not used in macro

If you ever come across this error, it can be a bit cryptic to understand.

Here’s some sample code that produces the error:

import haxe.macro.Expr;
import haxe.macro.Context;
import neko.Lib;

class BuildInMacroTests 
{
    public static function main() {
        var s:MyModel = new MyModel();
        MyMacro.doSomething( s );
    }
}

class MyModel extends sys.db.Object // has a build macro
{
    var id:Int;
    var name:String;
}

private class MyMacro 
{
    public static macro function doSomething(output : Expr) : Expr {
        return output;
    }
}

I’m still not sure I have a full grasp on it, but here’s a few pointers:

  • It’s not necessarily an enum – it can happen with a class or interface that has a “@:build” macro also.
  • The basic sequence happens in this order:
    • BuildInMacrosTest.main() is the entry point, so the compiler starts there and starts processing.
    • When it hits “var s:MyModel”, it does a lookup of that type, realizes the class needs a build macro to run, and runs the build macro.
    • When it hits “MyMacro.doSomething()”, it types the expression, and realizes that it is a macro call.  To run the macro, it must find the macro class, load it into the macro context, and execute it.
    • It finds the macro class, it’s in this file.
    • It tries to load this module (hx file) into the macro context, so it goes through the whole process of typing it again.
    • As it’s loading it into the macro context, it hits the “MyModel” build macro again, which it can’t do at macro time, so it spews out the error.
  • The basic solutions:
    • Wrap your build macro declarations in conditionals:
      #if !macro @:build(MyMacro.build()) #end class Object { ... }
    • Wrap anything that is not a macro in conditionals:
      #if !macro 
        class BuildInMacroTests {}
        class MyModel {}
      #else
        class MyMacro {}
      #end
    • Keep your macros in seperate files:
      BuildInMacroTests.hx:
          class BuildInMacroTests {
              public static function main() {
                  var s:MyModel = new MyModel();
                  MyMacro.doSomething( s );
              }
          }
      
          class MyModel extends sys.db.Object {
              var id:Int;
              var name:String;
          }
      
      MyMacro.hx:
          import haxe.macro.Expr;
          import haxe.macro.Context;
          class MyMacro {
              public static macro function doSomething(output : Expr) : Expr {
                  return output;
              }
          }

A combination of the 1st and 3rd solutions is probably usually the cleanest.

Good luck!

Creating Complex URL Routing Schemes With haxe.web.Dispatch

Complex Routing Schemes with haxe.web.Dispatch

I’ve looked at haxe.web.Dispatch before, and I thought it looked really simple – in both a good and a bad way. It was easy to set up and fast. But at first it looked like you couldn’t do complex URL schemes with it.

Well, I was wrong.

At the WWX conference I presented some of the work I’ve done on Ufront, a web application framework built on top of Haxe. And I mentioned that I wanted to write a new Routing system for Ufront. Now, Ufront until now has had a pretty powerful routing system, but it was incredibly obese – using dozens of classes, thousands of lines of code and a lot of runtime type checking, which is usually pretty slow.

Dispatch on the other hand uses macros for a lot of the type checking, so it’s much faster, and it also weighs in at less than 500 lines of code (including the macros!), so it’s much easier to comprehend and maintain. To wrap my head around the ufront framework took me most of the day, following the code down the rabbit hole, but I was able to understand the entire Dispatch class in probably less than an hour. So that is good!

But there is still the question, is it versatile enough to handle complex URL routing schemes? After spending more time with it, and with the documentation, I found that it is a lot more flexible than I originally thought, and should work for the vast majority of use cases.

Learning by example

Take a look at the documentation to get an idea of general usage. There’s some stuff in there that I won’t cover here, so it’s well worht a read. Once you’ve got that understood, I will show how your API / Routing class might be structured to get various URL schemes:

If we want a default homepage:

function doDefault() trace ( "Welcome!" ); 

If we want to do a single page /about/:

function doAbout() trace ( "About us" ); 

If you would like to have an alias route, so a different URL that does the same thing:

inline function doAboutus() doAbout();

If we want a page with an argument in the route /help/{topic}/:

function doHelp( topic:String ) { 
 trace ( 'Info about $topic' ); 
} 

It’s worth noting here that if topic is not supplied, an empty string is used. So you can check for this:

function doHelp( topic:String ) { 
 if ( topic == "" ) { 
 trace ( 'Please select a help topic' ); 
 } 
 else { 
 trace ( 'Info about $topic' ); 
 } 
} 

Now, most projects get big enough that you might want more than one API project, and more than one level of routes. Let’s say we’re working on the haxelib website (a task I might take on soon), there are several pages to do with Projects (or haxelibs), so we might put them altogether in ProjectController, and we might want the routing to be available in /projects/{something}/.

If we want a sub-controller we use it like so:

/*
 /projects/ => projectController.doDefault("")
 /projects/{name}/ => projectController.doDefault(name)
 /projects/popular/ => projectController.doPopular
 /projects/bytag/{tag}/ => projectController.doByTag(tag) 
*/ 
function doProjects( d:Dispatch ) { 
 d.dispatch( new ProjectController() ); 
} 

As you can see, that gives a fairly easy way to organise both your code and your URLs: all the code goes in ProjectController and we access it from /projects/. Simple.

With that example however, all the variable capturing is at the end. Sometimes your URL routing scheme would make more sense if you were to capture a variable in the middle. Perhaps you would like:

/users/{username}/ 
/users/{username}/contact/ 
/users/{username}/{project}/ 

Even this can still be done with Dispatch (I told you it was flexible):

function doUsers( d:Dispatch, username:String ) { 
 if ( username == "" ) { 
 println("List of users"); 
 } 
 else { 
 d.dispatch( new UserController(username) ); 
 } 
} 

And then in your username class:

class UserController { 
 var username:String; 
 public function new( username:String ) { 
 this.username = username; 
 } 
 function doDefault( project:String ) { 
 if ( project == "") { 
 println('$username\'s projects'); 
 } 
 else { 
 println('$username\'s $project project'); 
 } 
 } 
 function doContact() { 
 println('Contact $username'); 
 } 
} 

So the username is captured in doUsers(), and then is passed on to the UserController, where it is available for all the possible controller actions. Nice!

As you can see, this better represents the heirarchy of your site – both your URL and your code are well structured.

Sometimes, it’s nice to give users a top-level directory. Github does this:

http://github.com/jasononeil/

We can too. The trick is to put it in your doDefault action:

function doDefault( d:Dispatch, username:String ) {
  if ( username == "" ) { 
 println("Welcome"); 
 } 
 else { 
 d.dispatch( new UserController(username) ); 
 } 
} 

Now we can do things like:

/ => doDefault() 
/jason/ => UserController("jason").doDefault("") 
/jason/detox => UserController("jason").doDefault("detox") 
/jason/contact => UserController("jason").doContact() 
/projects/ => ProjectController.doDefault("") 
/about/ => doAbout() 

You can see here that Dispatch is clever enough to know what is a special route, like “about” or “projects/something”, and what is a username to be passed on to the UserController.

Finally, it might be the case that you want to keep your code in a separate controller, but you want to have it stay in the top level routing namespace.

To use a sub-controller add methods to your Routes class:

inline function doSignUp() (new AuthController()).doSignUp(); 
inline function doSignIn() (new AuthController()).doSignIn(); 
inline function doSignOut() (new AuthController()).doSignOut(); 

This will let the route be defined at the top level, so you can use /signup/ rather than /auth/signup/. But you can still keep your code clean and separate. Winning. The inline will also help minimise any performance penalty for doing things this way.

Comparison to Ufront/MVC style routing

Dispatch is great. It’s light weight so it’s really fast, and will be easier to extend or modify the code base, because it’s easier to understand. And it’s flexible enough to cover all the common use cases I could think of.

I’m sure if I twisted my logic far enough, I could come up with a situation that doesn’t fit, but for me this is a pretty good start.

There are three things the old Ufront routing could do that this can’t:

  1. Filtering based on whether the request is Post / Get etc. It was possible to do some other filtering also, such as checking authentication. I think all of this can be better achieved with macros rather than the runtime solution in the existing ufront routing framework.
  2. LocalizedRoutes, but I’m sure with some more macro love we could make even that work. Besides, I’m not sure that localized routes ever functioned properly in Ufront anyway ;)
  3. Being able to capture “controller” and “action” as variables, so that you can set up an automatic /{controller}/{action}/{...} routing system. While this does make for a deceptively simple setup, it is in fact quite complex behind the scenes, and not very type-safe. I think the “Haxe way” is to make sure the compiler knows what’s going on, so any potential errors are captured at compile time, not left for a user to discover.

Future

I’ll definitely begin using this for my projects, and if I can convince Franco, I will make it the default for new Ufront projects. The old framework can be left in there in case people still want it, and so that we don’t break legacy code.

I’d like to implement a few macros to make things easier too.

So

var doUsers:UserController; 

Would automatically become:

function doUsers( ?d:Dispatch, username:String ) { 
 d.dispatch( new UserController(username) ); 
} 

The arguments for “doUsers()” would match the arguments in the constructor of “UserController”. If “username” was blank, it would still be passed to UserController.default, which could test for a blank username and take the appropriate action.

I’d also like:

@:forward(doSignup, doSignin, doSignout) var doAuth:AuthController; 

To become:

function doAuth( ?d:Dispatch ) { 
 d.dispatch( new AuthController() ); 
} 
inline function doSignup() (new AuthController()).doSignup(); 
inline function doSignin() (new AuthController()).doSignin(); 
inline function doSignout() (new AuthController()).doSignout(); 

Finally, it would be nice to have the ability to do different actions depending on the method. So something like:

@:method(GET) function doLogin() trace ("Please log in");

@:method(POST) function doLogin( args:{ user:String, pass:String } ) { 
 if ( attemptLogin(user,pass) ) { 
 trace ( 'Hello $user' ); 
 } 
 else { 
 trace ( 'Wrong password' ); 
 } 
}
 
function doLogin() { 
 trace ( "Why are you using a weird HTTP method?" ); 
} 

This would compile to code:

function get_doLogin() { 
 trace ("Please log in") 
} 

function post_doLogin( args:{ user:String, pass:String } ) { 
 if ( attemptLogin(user,pass) ) { 
 trace ( 'Hello $user' ); 
 } 
 else { 
 trace ( 'Wrong password' ); 
 } 
} 

function doLogin() { 
 trace ( "Why are you using a weird HTTP method?" ) 
} 

Then, if you do a GET request, it first attempts “get_doLogin”. If that fails, it tries “doLogin”.

The aim then, would be to write code as simple as:

class Routes { 
 function doDefault() trace ("Welcome!"); 
 var users:UserController; 
 var projects:ProjectController; 
 @:forward(doSignup, doSignin, doSignout) var doAuth:AuthController; 
} 

So that you have a very simple, readable structure that shows how the routing flows through your apps.

When I get these macros done I will include them in ufront for sure.

Example Code

I have a project which works with all of these examples, in case you have trouble seeing how it fits together.

It is on this gist.

In it, I don’t use real HTTP paths, I just loop through a bunch of test paths to check that everything is working properly. Which it is… Hooray for Haxe!

A new project: OurSchoolDiary

So I’m working on a new project that has been brewing in my mind for a while. I’m going to aim to have a working prototype by the end of August, have users beta testing through till December and hopefully have my first paying clients by next year.

A screenshot of some of the first code for the project.

What is it?

The App is called “OurSchoolDiary”. It lets each school create their own School Diary app, where students can sign up and keep track of their homework, assignments and tests. The big win over competing solutions is that here, when the teacher adds an assessment or task, it automatically shows up in the students diary. We’re helping even unorganised students stay on top of their work.

On top of that, we’ll let parents receive emails informing them of their child’s progress. This way parents and teachers can work together to make sure they support their child’s education as much as possible.

And the final advantage, the school gets to show off how cutting-edge they are because they get their own “app”.  Logo and everything!

A Blog

As I go, I hope to keep track of my progress on this blog. (I’ll post to this category specifically, so you can subscribe by RSS).

Why keep a blog?

  • For the history, in case it is ever significant
  • For my own reflection and understanding, regardless of if it is a success.
  • To show off some cool technologies I’m using:
    • Web Apps – I’m not aiming for distribution through the App Store, I’m going a full HTML stack. Not even a hybrid using PhoneGap. This is a page you can view in a browser, and use offline, and the experience will be identical.
    • Haxe – this is a sweet programming language that lets you target Flash, Javascript, PHP, Java, C#, C++ and a platform called Neko. Up until now it’s main success has been in the indie-games market, but I believe it is mature enough to work for a full web-based productivity app.  Client and server. We’ll find out how that goes!
  • To look back on my decision making processes after we’ve launched, and see if my thinking was accurate.
  • To raise interest in the project
  • To explore some of the topics I find fascinating:
    • Usability
    • Design
    • New Marketing
    • Business Strategy
    • Contribution to the community
    • And a few more…

That’s all for now. I’ll try submit a couple of posts a week (at least) until I get this thing off the ground.