Friday 7 August 2009

Faking startup and shutdown with ActiveSupport::TestCase

Time-was we didn't just have setup and teardown in our Tests (which are run for every single test case)... we could also have a startup/shutdown pair that ran once per proper Test-case class (ie once at the beginning of all tests in a set).

This was a perfect place to put run-once-only things - especially useful if, say, you wanted to run ruby-prof over your tests. You could put the RubyProf.start in startup and have it print out a callgraph in the shutdown...

ActiveSupport::TestCase doesn't let us have our startup/shutdown anymore... so what to do?

Fake it with at_exit

  # quick-and-dirty class var to tell us if we've already registered the
  # at_exit RubyProf function.
  @@register_done = false
  # register the stop-ruby-prof stuff
  # We want it to spit out a RubyProf callgraph after it's run this test
  # case.
  # You can either: call this function directly in a single test case or
  # test-suite eg: "register_rubyprof"
  # If you pass a value to "RUN_RUBY_PROF" on the command-line - it will
  # also run this register function for you. eg:
  # RUN_RUBY_PROF=true rake test:units TEST=test/unit/widget_test.rb
  def register_rubyprof
    require 'ruby-prof'
    return if @@register_done
    p "starting RubyProf"
    RubyProf.start

    path_base = "#{RAILS_ROOT}/tmp/performance/ruby-prof"

    p "started! Now registering exit-function"
    at_exit do
      p "All done - stopping RubyProf" 
      result = RubyProf.stop

      p "Stopped! now printing call graph"
      timestamp = Time.now.strftime('%Y-%m-%d-%H-%M-%S')
      # try and make the callgraph filename meaningful to the test-run
      filename_base = "rubyprof_#{self.class.name}_#{timestamp}"

      # Print a call tree profile to text for kcachegrind to use
      printer = RubyProf::CallTreePrinter.new(result)
      printer.print(File.open("#{path_base}/#{filename_base}_calltree.txt", 'w'), 0)
      p "All Done. Good luck profiling!"
      end
    @@register_done = true
  end
  # will setup the reguster-rubyprof function if we have set the
  # RUN_RUBYPROF constant to true
  setup :register_rubyprof if RUN_RUBY_PROF

Thursday 6 August 2009

HyRes : ActiveResource that actually works!

It's three months on from my original post whinging about the lack of Validations in Active Resource.

At that time I put my money where my mouth was and forked a copy of the HyperactiveResource plugin, which provided a very crude, basic improvement over vanilla Active Resource.

So What have I achieved?

I've actually done a lot since then. I have implemented a lot of the TODOs that I wrote down as essential for us to get a good, solid basis for a workable system (see the done list below).

I'd class this plugin as working and functionally useful. Right now it's almost as useful as Active Record was a few years back... which is a vast improvement on the basic Active Resource interface.

I think there's still a lot of room for improvement, but the basics are there, and it *feels* like Active Record now - which before it definitely did not. Before it felt like you could play around with it for toy systems, or implement minor bits of functionality - but now you can really implement a fully-functional system on top of it. I know this - because we have done[1].

What's missing?

HyRes still has some niggling issues, and some stuff that would make it a much smoother migration for Active Record. I'm working on these and there is a perpetually-renewed TODO list on the HyRes README file

The funkier aspects of Active Record

I haven't yet re-implemented proper AR reflections - so the associations need some work... you can't do: widget.wodgets.find(...) and you can't use named_scope - which means that many basic plugins (eg restful_authentication) won't work out-of-the-box. But there's enough functionality there that pulling together a system that is totally non-Active Record is feasible.

API assumptions

Currently HyRes makes some assumptions about your API - it assumes a RESTful interface as the optimal configuration. If your API does not match the assumptions, there are some workaround available, but it might not be as useful.

An example is the validates_uniqueness_of method. This currently assumes that your API takes an array of conditions on the query-string, and that this will filter your returned set of findable objects.

If your API doesn't do this... the method currently defaults to fetching *all* objects and filtering on the rails-end... which is likely to be extremely slow (and may lead to timeouts). But it's there in case it's necessary. You may have to simply re-write that method with your own API-requirements in mind. I welcome alternative solutions to the problem...

testing...

Probably one of the nastier downsides atm is that HyRes doesn't have its *own* set of tests... We currently test it through the full suite of tests on our own system. This is mainly due to the fact that it's really quite hard to set up tests that don't rely on an existing API that's up and running. In our system we use a combination of mocks in our functional tests, backed by a fully-functional, running mock API for our integration tests... this is near-impossible for the HyRes plugin itself... so I'm still thinking about how I'll test it independently.

So what *does* it do?

The current list of implemented features (along with how to set up your API to use them) is also available on the HyRes README file but a quick snapshot of the list as of today is below. Note - in some cases, I still find it hard to believe that Active Resource didn't already implement these...

Columns

Because Active Resource isn't backed by a db... you can't use the table columns to determine the known set of attributes of a resource. ARes currently works by accepting any incoming data as an attribute, and using MethodMissing as an accessor for any known attribtues. This is fine for situations where you don't know what attributes will be returned by the remote system.

The downside is that if no value is returned for an attribute and you try to access it... it throws a MethodMissing exception (uuugly!).

If, however, you know what attributes to expect (because it's an agreed API-interface ), it'd be nice to be able to tell Active Resource which attributes we expect - and have it return any missing attributes an nil (rather than explosion).

Thus was born the columns= method.

Currently this method also works a little like attr_accessible - any attributes listed like this will be passed through on create/update... but nothing else will. This allows you to set temporary accessors (eg password and password_confirmation without them being passed over the wire.

Validations

directly pulls them in from ActiveRecord::Validations. For Rails 3.0 - we can use the ActiveModel::Validations component.

Note: validates_uniqueness_of is still experimental - as it requires a network-hit to actually determine whether you have any existing objects with the given field-value. You can't rely on uniqueness-of as there is no way of locking the remote-system's db - and thus somebody could have added a resource with that value in the meantime.

Callbacks

There are now callback hooks for before_validate, before_save etc - all the standard Active Record callbacks we've come to know and love. You can use them for callback chaining a la Active Record (still experimental - may have missed some, but you can definitely use validate :my_validation_method)

Finders

conditional finders

eg Widget.find(:all, :conditions => {:colour => 'blue'}, :limit => 5). This functionality working depends on your API accepting the above filter fields and returning something sensible. There's a lot of doco on how to set up your API (if you have that luxury) in the HyRes README

Dynamic finders and instantiators

eg find_by_X, find_all_by_X, find_by_X, find_last_by_X These rely on your API accepting filter fields as they are really just a bit of a convenient alias for find(:conditions => X). Dynamic finders take any number of arguments using _and_: find_by_X_and_Y_and_Z

We also have dynamic instantiators: find_or_create_by_X OR find_and_instantiate_by_X - both of which also take any number of args, just like the dynamic finders.

Dynamic finders/instantiators also take ! eg: find_or_create_by_X! will throw a ResourceNotValid exception if create fails

Other ActiveRecord::Base-like functions

  • update_attribute / update_attributes - actually exist now!
  • save! / update_attributes! / update_attribute! / create! - actually exist, now that we can do validation. They raise HyperactiveResource::ResourceNotSaved on failure to save/validate
  • ModelName.count (still experimental) - with optional finder-args. This works by first trying a /count request on your model object - and if the route fails it just pulls out all the objects and returns the length (ie - a nasty - but functional fallback).
  • updated collection_path that allows suffix_options as well as prefix_options = allows us to generate Rails-style named routes for collection-actions
  • no explosion for find/delete_all/destroy_all when none are to be found. (Active Record just returns nil - so should we)
  • ActiveRecord-like attributes= (updates rather than replaces)
  • ActiveRecord-like load that doesn't dup attributes (stores direct reference)
  • reload that does a full clear/fetch to reload (also clears associations cache, now that we have one)

Associations

A lot of the basic associations were in HyRes before I arrived - I've polished a few up (eg adding the belongs_to/has_many functions) and bugfixed

  • Resources can be associated with records
  • Records can be associated with records
  • Awareness of associations between resources: belongs_to, has_many, has_one (note - no HABTM yet)
    • Patient.new.name returns nil instead of MethodMissing
    • Patient.new.races returns [] instead of MethodMissing
    • pat = Patient.new; pat.address_id = 1; pat.address # returns the address object
  • Can fetch associations even with a nested route by using the ":nested" option on the nested resource's class. This command automatically adds a prefix-path, and will pre-populate the parent's id when you do an association collection_fetch.
  • Supports saving resources that :include other resources via:
    • Nested resource saving (creating a patient will create their associated addresses)
    • Mapping associations ([:address].id will serialize as :address_id)

What's next?

Well, one big thing that's next is that I'm planning on starting to work this stuff back into Rails itself. The brilliant component-architecture and upcoming Active Model changes for Rails 3 are just made for this. What I've been doing on HyRes can feed into that process to make Active Resource what we've always wanted - a real replacement-candidate for Active Model. For this purpose I have a fork of rails - but work will progress slowly (as I'm working on it in my spare time).

Until then, I'll keep on upgrading HyRes as needed for our own corporate requirements. Below, I've listed some features that Active Record provides and that I'd *personally* like to see implemented next... but this is certainly not an exhaustive list. (and I'm very open to other ideas...)

One thing I'd love to see is other people using HyRes for their own projects. I'd love to hear feedback on how it works (or not) as that will feed the project and feed into the work that eventually goes back into Rails. Better-still, if you're willing to work on implementing some of the still-missing features... you would be most welcome. Feel free to check out the Hyperactiveresource project on github and have a go.

Still TODO

  1. Testing - inside the plugin! (currently we test via our actual web-app's set of exhaustive tests)
  2. MyModel.with_scope/named_scope/default_scope
  3. MyModel.find(:include => ...)
  4. attr_protected/attr_accessible
  5. MyModel.calculate/average/minimum/maximum etc
  6. Reflections. There should be no reason why we can't re-use ActiveRecord-style reflections for our associations. They are not SQL-specific. This will also allow a lot more code to automatically Just Work (eg an Active Record could use has_many :through a HyRes)
  7. has_and_belongs_to_many
  8. Split HyRes into Base and other grouped functions as per AR
  9. validates_associated
  10. write_attribute that actually hits the remote API ???
  11. a default format for when it doesn't understand how to deal with given mime-formats? One which will just pass back the raw data and let you play with it?
  12. cache the raw (un-decoded) data onto the object so we don't have to do a second fetch? Or at least allow a universal attribute to be set that turns on caching

Notes

[1] The subject of an upcoming-blogpost will be on the work we've done: "Rewriting monolithic legacy systems in Rails"