Skip to main content

LevelTSDB

I've started splitting out useful time-series database functions from ETSDB into their own library as LevelTSDB. This is mostly so I don't have to test everything again for some things I would eventually like to make.

Nikola Generator

Starting to use a new static site generator now that there are bunch of good ones in Python. I find Python/pip more sane to use than Ruby/bundler/rbenv

Deploying Python Without Downtime

When you first start out deploying your application it can be easy to just run supervisor restart all or service my_app restart to get your current version into production. This is great when you are starting out but eventually you will try to connect while your application is starting up and see HTTP 503s while you application is booting up.

Eventually you might discover that Gunicorn and uWSGI can reload your application without closing the socket so your web requests will just be delayed a bit delayed as your application starts. This works fine as long as your application doesn't take too long to start. Unfortunately some applications at work can take a minute to start, too long to have connections waiting at the socket.

The Gunicorn reloading using kill -HUP $PID will stop all worker processes then start them again. The slow init for workers tends to cause problems. uWSGI has chain reloading which will restart workers one at a time. I need support for Tornado which doesn't fit well with uWSGI.

With a Load Balancer

A common technique is to remove a single server from the load balancer, upgrade/restart the application, then bring it back. We are using load balancers but it requires coordination while provisioning nodes using the HAProxy management socket in order to schedule this. Our deploys currently deploy to all nodes simultaneously, not one-by-one, an even larger change. It would also be possible to fool the healthcheck by 404'ing the status page then waiting for LBs to take the node out of the pool. That requires a bit more waiting than I want, 2 healthcheck failures with 5 second intervals, for each server, plus time to reintegrate the web process once the upgrade is finished.

Gunicorn Reload ++

Gunicorn will automatically restart failed web processes so it would be possible to just kill each process, sleeping in between, until you get through all the child processes. This works but if application start times change significantly we are either waiting too long for restarts or not long enough and risking some downtime.

Since Gunicorn includes Python hooks into the application it should be possible to write a snippet that will notify the restart process when the worker application is ready. Gunicorn didn't have the needed hook but it was simple to contribute the change. It requires master until a new release is made.

Now our restart process takes advantage of the fact that a single socket has multiple processes accepting connections. Restarting will slightly diminish our capacity (1/N) but we will continue to handle traffic without letting connections wait too long.

The general process for this is

  for child_pid of gunicorn-master:
    kill child_pid
    wait for app startup

My first version of this used shell and nc to listen on UDP for an application startup. This worked well although integrating our process manager into shell was a bit more then I would like to do.

The restart script should be called with the PID of the Gunicorn master restart.sh $PID

and works in tandem with a post_worker_init script that will notify the script when the app is running.

If we had this WSGI application for example:

We could even do things like check the /_status page to verify the application is working.

Be careful with trying to run too much of your application in this healthcheck, if for any reason your post_worker_init raises an error then the worker will exit, preventing your application from starting. This may be a problem when you are checking a DB connection that may go away, even if you application could work it won't be able to boot.

Now with our applications that take a minute to start we can do a rolling restart without taking the application down or dropping any connections!

Ordering of Rebar Dependencies

As I am starting out with Erlang I've just added dependencies to the end of my Rebar config and everything just kind of worked. I added each dependency one-by-one and didn't have a problem until I cleaned out the deps folder and tried to recompile. Then I ran into this error:

src/ranch_protocol.erl:none: undefined parse transform 'lager_transform'

I knew that it was working before and that the parse transform wasn't an issue. Turns out the dependency ordering matters! Shouldn't be too big a surprise but Rebar uses the list of dependencies as the ordering for compilation, not any kind of introspection. I just had to put the Lager dependency above Ranch and everything worked out.

SeatGeek RSS

I've setup an RSS feed for local concerts powered by SeatGeek. We (at SeatGeek) don't have one built-in but we do have an API. The page isn't pretty but I find it useful for finding any events I may want to go to. With tagging in NewsBlur I can filter events more easily.

I built this with Erlang as a way to test out the language. There isn't really a direct need for high concurrency but it was a good chance to give it a try. I've learned that I really like Erlang, it's rather terse and has constructs built into OTP that make writing software a joy. At some point I need to tackle using releases, but I'm not there yet.

When I spend more time on the RSS feed I'll eventually include affiliate links. It takes a lot of traffic to make money with affiliates especially at most concert prices. But maybe it will be an incentive for me to turn this into something even more useful.

More on RabbitMQ Priorities

With a single process consuming from multiple queues the prefetch count could be a good enough solution to balancing the work from each queue.

After you have set up priorites with multiple queues you still need to consume from them. You could setup separate processes for each queue or a single process that consumes from multiple queues.

I usually set consumers to a prefetch count of 10, it works well enough and latency isn't much of a concern. When consuming from multiple queues setting each queue to the same prefetch count will give you a fair distribution of work to that consumer.

What I finally took the time to try this week was changing the prefetch counts based on priority. In my case we had 2 queues, high and low priority. The higher priority was based on user actions and we wanted to happen quickly. There was only 1 set of processes consuming from both queues and had the same prefetch counts. Since the messages are sent to the consumer ahead of time there were 20 messages for each process. Adjusting the low priority queue to a prefetch of 2 meant that there would be 12 items sent to the consumer, still plenty of work. These 12 items are put into a single queue in the library, no work needed in your code, and will give a 5/1 distribution of work in the consumer.

With the adjusted prefetch counts we are able to control which portion of the work we wind up doing when queues start to backup. In this case you have to sacrifice latency to do it, the higher priority queue may give more work to a busy consumer when others could be empty. In practice for us this did not matter, we set the prefetch on the high priority queue to 10 anyway.

This has the nice property that low priority items are still processed while high priority items exist and will be consumed at the highest rate as soon as the high priority items are drained. With more than 2 queues this technique may be cause more latency than you would like but it has be working well and required no code changes. I was planning on making a locking mechanism, and if you didn't want any low priority work in progress while there was high priority work you would still need to, but I don't think one will be needed anytime soon.

Conference Going

This was the first Monitorama event, held in Boston, and was a great chance to meet the people behind a lot of the software/blogs I follow. The first day was a single-track set of talks regarding open source monitoring and the second day a hackathon to help improve the state of open source monitoring. I contributed a bit to correct a small pain point I had. I didn't enter it into the judging partially because I believed it to be a rather simple hack and partially because I didn't want to have to rush to get back to Amtrak for my train back to NY. I was pleasantly surprised to see food always available including plenty of healthy bits.

PhillyETE was definitely a big change from Monitorama, more people, more talks, not as great of food. Part of the benefit for me to go to PhillyETE is the trip to Philadelphia to see my family, especially as an Easter trip. The best part of the conference had to be seeing the push for Clojure as an enterprise language, (and slightly less interest to me, Scala). Given that they are JVM languages they fit in very well and can work side by side for evaluation. I tried Scala for a bit since I wanted to try out Akka but ran into JVM memory issues during the Play "Hello, World!" tutorial which really soured me towards Scala.

Basho's sponsorship of Monitorama also helped convince me (with a 25% discount) to attend Ricon East. That may be the end of my conference going this year, I haven't decided yet about Surge.

Graphite Pager - v0.0.6 - Links to Documentation

The change for this release was to add links to documentation for each alert. Currently the format of the URL is {docs_url}/{alert name}#{alert legend name} where the docs_url is specified in the YAML config and the rest is based on the alert that is triggering.

While people at work haven't jumped to create metrics and alerts for various things this will at least make it easier for them to know why this alert was created and how to fix the problem. Right now I have only documented a few alerts and will do so as existing alerts fire. If anyone needs alerts made I will make sure the wiki page exists ahead of time.

Prioritizing Emails with RabbitMQ

You should already have a worker that can send the email, just now you need setup RabbitMQ with priorities.

The main exchange you use, email, should be declared either topic or direct and will take all of the messages you intend to send but when declared you should include an alternate exchange of email-undeliverable that is declared as a fanout exchange. Now you just need a default queue bound to the default routing key for the email exchange and also bound to email-undeliverable. Now every email your try to send that doesn't have a specifically prioritized queue will be routed to the default queue.

All you need now is to start your workers consuming from each queue you create.

Provisioning AMQP

An alternative is to have consumers and producers take only the name of a queue or exchange and handle the rest outside of the application. This allows you to see and change in one place the configuration for all of your applications. When you need to provision a new broker it is done in a few seconds instead having to migrate some consumers, then all producers, then the rest of the consumers.

I've started writing and using Declare AMQP so that I can provision everything within Chef. It only supports the features I'm using but is very simple.

The migration is now much simpler as provisioning the server once is enough to make it ready for all applications. When I need to change exchanges or bindings I don't have to update any code. There is still the need to know which applications publish which routing key, but not a huge concern.

This has helped out as well configuring queues with specific priorities for the same type of tasks. Each application can be started with a queue to listen to and the configuration for both the broker and applications remains in one place.

Example Tornado AMQP Client with Pika

I wrote an client for internal use that handles the conditions I needed by default (including callbacks with the result of RabbitMQ publish confirmations) and after talking with a previous coworker put it into a gist.

It doesn't handle some things (like publish a content type with the encoded json) and could have some better names but it may be of use to more people.

MongoDB Lock Percentage in Graphite

Using Diamond you can easily get all of the MongoDB server status metrics into Graphite but the globalLock.ratio is a bit misleading in that it is based on the total uptime of Mongo, which could be a while, and not on recent usage patterns. And in 2.2 it disappears anyway!

The metrics that are included though that help are globalLock.totalTime and globalLock.lockTime which can be used to find the lock ratio/percentage over whatever sampling period you use.

The percentage winds up being scale(divideSeries(derivative(servers.MONGOHOSTNAME.MongoDBCollector.globalLock.lockTime),derivative(servers.MONGOHOSTNAME.MongoDBCollector.globalLock.totalTime)),100). You can remove the scale function to get the ratio. This doesn't work with globbing in Graphite though. You can scale the lockTime though to be able to get a globbable lock ratio for all of your Mongo servers, the exact value will depend on the sampling period.

SQL-to-Graphite

We use this and similar scripts (I'm going to move over to using this) at work in order to collect global metrics about our systems. I typically count any table that has a status column and the average/max age of any records that should be updated periodically.

I made this package once I hit the second repository where I would have to write a script to do this. It should be compatible with any database supported by SQLAlchemy.

After installing (pip install sql-to-graphite) you can run the sql-to-graphite command.

With a file like:

And start getting metrics into Graphite!

Autodetecting Your RSS Feed in the Browser

I noticed recently that my site didn't have the Syndication
Icon icon anymore. I'm not sure when I lost it but to add it back I just added the link field that lets browsers know where my Atom feed is. Simple enough to add <link type="application/atom+xml" rel="alternate" href="/atom.xml"/> to the head of the page.

Resque Metrics with StatsD

A recent task of mine was to add some metric collection to a Rails application at SeatGeek. One of the main components (and critical if there was a problem) is the set of Resque background workers. There is actually a Resque Plugin (abandoned, maintained that will collect stats. The gem sadly is not maintained so I forked the maintained repo in order to provide a stable source. I use the commit hash to make sure I get the version but if the repository we used disappears that would cause problems, so a fork solves that.My fork doesn't change much except for some of the paths used for the metrics. At some point I may clean up the README and package my first gem.

Tagging in Jekyll

Categories in Jekyll have annoyed me for a while because of the URLs generated. The path would be something like /tag1/tag2/year/month/day/title which works so long as you don't change the categories used. Since tags are also an option and don't have the same issue I've switched. I followed this post about tagging archive pages in Jekyll that made it rather painless.

Graphing Influence

This started with a few minutes after lunch at SeatGeek where we were checking various Klout scores. Since I tend to graph... everything... I quickly setup a cron script to start collecting the metrics for Graphite.

To run it:

Ideally this is run in cron, we use 30 minutes. Over the course of 2 weeks there is already a few rank changes and large jumps due to adding new social networks to Klout.

14 Days of Kout

Monitoring Service Health Check Duration

Check Duration

This service can usually hits the 50ms range for health checks although it started getting much worse. The service is actually written in Tornado although has a few blocking calls that are used. Non-blocking IO should allow the health checks to be very quick to respond as in this case it returns a static response.

The root cause for the problem is that calls to MongoDB in a particular handler were taking longer than before and will hold back other handlers as it is currently a blocking operation. If the HAProxy health checks pass a threshold it will remove the nodes from the pool, a good precaution, although in our case can cause flickering if MongoDB takes longer than expected.

I did receive alerts thanks to alerting of per-service health checks with Graphite Pager.

We are using Diamond at SeatGeek which easily collects metrics from HAProxy. Check duration is (by default) stored at servers.HAPROXY-SERVER.haproxy.BACKEND.HOST-SERVER.check_duration. The metric we alert on is the moving median for each server regardless of the HAProxy server aliasByNode(movingMedian(groupByNode(servers.*.haproxy.*.*.check_duration,3,"averageSeries"),10),0).

Graphite Pager - An Easy Way to Send Alerts From Graphite

Right now I'm testing it at SeatGeek running it on Heroku, the example of how to set up Graphite Pager on Heroku is small and straightforward. It has already helped detect a few problems before our other monitoring tools and (eventually) can alert on actual business metrics!

The alert format looks like:

Pretty Simple. It supports globbing with unique alerts for each metric. Graphite Pager can't determine a disappearing host from the glob, maybe in the future, but will set the alert for all metrics returned.

New Job

AWeber was a great opportunity that turned me from a programmer into a developer. I wish everyone there the best of luck.

Now that I've left I'm upset that most of what I worked on there was not open source so I will no longer be able to use the things that I've built. Towards the end of my time there I was trying to put more projects on Github that could have been useful to someone. Hopefully at SeatGeek I will have the change to make some larger contributions!

Kanban What

The official blurb for the Kanban talk

When faced with the challenges of managing a growing email marketing software and 40-person development team, AWeber turned to the project management system Kanban. In this presentation with Ethan McCreadie and Philip Cristiano, they shared AWeber's journey into Kanban, how it functions within AWeber's team structure, and the advantages and disadvantages other companies should take into consideration before implementing the system.

Free Shirts (Focus on Quality)

Companies that I've seen give away American Apparel shirts include:

Even when the company is a competitor to my employer (MailChimp), I'm likely to wear the shirt.

Companies that at least make a comfortable shirt without crappy graphics may still be worn after I'm through the American Apparel shirts.

Like in software, even if you are giving away something for free, you should focus on quality.

Brightcove's Diamond and Contributing to Projects

I have spent a significant amount of time recently looking into monitoring and metrics collection. At work we have Nagios and Cacti currently and are looking at other options. After setting up Ganglia we decided to give Graphite a try. There is a script to send data from Ganglia to Graphite although the whole system gets to be more complex than I'd like. The chain winds up being: monitored server -> Ganglia -> Ganglia parser -> Graphite.

Looking at the Graphite Tools page I learned about Diamond which can collect metrics using a Python agent and send them directly to Graphite. It has some very useful collectors already and the big benefit, makes it very easy to add a new collector!

In about an hour I was able to make a collector for MongoDB and submit a pull request. The changeset was accepted making it my first contribution to an existing open-source project! I plan on creating more collectors for software that we are using as work. Next up is likely to be RabbitMQ.

And if you haven't, try Graphite!

vimrc Updates

I've kept my VIM/dot files online for a while in my Github but I recently spent some time to update my .vimrc file.

One of the changes that bugged me on OSX is that it ships with VIM 7.2 which doesn't have ColorColumn support. I like highlighting the 80th column in Python. As I discovered, the code to do this is :

{% highlight vim %} if version >= 730 autocmd FileType python set cc=80 hi ColorColumn ctermbg=darkgrey guibg=darkgrey endif {% endhighlight %}

Another change was to add Gundo support. This adds a window to navigate your undo tree, an incredibly useful feature.

Branch Coverage with Nose

Since I heard about the addition of branch coverage tracking for Coverage I've wanted to give it a try. Originally it required a beta release which somehow I never got working.

Once it was in a normal release I somehow forgot about it. There is still no commandline argument to turn it on when using Nose. You can however use the .coveragerc file to enable it.

In .coveragerc simply put

{% highlight bash %} [run] branch = True {% endhighlight %}

And next time you run Nose with coverage you'll have branch coverage too! I was finally reminded about this when coming across the Test Coverage Analysis post by Kai Lautaportti.

Code Reviews with Git

A few weeks ago at work we improved our code review process by using Git more effectively. Previously a code review happened after the topic branch was merged into master. This obviously was not very effective as changes could have broken master without a proper review and there was less incentive to perform as careful of a review since the code was "working" already. This was carried over from when we were using SVN until we realized we were no longer forced to work in the dark ages.

Since we were already using Git we could easily change our workflow for a better review process. Once a topic branch was ready for review we could push a remote branch. Our remote branches take the form /review/{initials}/{topic}

To push a new remote branch:

{% highlight bash %} git push origin {branch name}:/review/{initials}/{branch name} {% endhighlight %}

And then we would move the Kanban card to the review column and find someone to review our changes.

When the code has been reviewed and any necessary changes made the reviewer will merge them into master.

{% highlight bash %} git checkout master git merge --no-ff --no-commit /review/{initials}/{branch name} git commit -s {% endhighlight %}

The merge command turns off fast-forward merging and commiting so when we commit with a -s we can sign-off on the changes. This shows who reviewed the code. We don't permit anyone to push their own changes without review although Git doesn't prevent you from changing the committer or the sign-off.

Then finally remove the remote branch by pushing an empty branch over it and delete your local copy of the branch.

{% highlight bash %} git push origin :/review/{initials}/{branch name} git branch -d /review/{initials}/{branch name} {% endhighlight %}

Plug v0.1.2 Bug Fixes

This release fixes a bug cause by linking all plug service instances to the installed plug. Runit uses a ./supervise in the service's directory to maintain state which would be clobbered when multiple services link to the same plug.

Now the virtualenv is copied to /srv/plug/plug_instances/ and linked into runit.

There is also a fix for uninstalling plugs leaving orphaned processes. Now Plug will stop the processes before removing them to prevent this.

Plug v0.1.1

Release v0.1.1 adds an uninstall command to Plug that takes a --plug= option and removes the virtualenv and all runit links.

You can get Plug on PyPi and try it out. As always, report any issues.

TestHTTPServer v0.1.2 - Beta Status!

Release v0.1.2 includes the ability to set custom response headers and the __version__ package attribute! This release should be "complete" for my own needs and for common uses. If you see anything you need open an issue and I can probably get it updated and packaged in a day or so.

After I use the package a bit more I will bump the version and switch it to production status.

TestHTTPServer v0.1.1

I've push v0.1.1 of TestHTTPServer. This release adds the ability to handle any method as well as storing request headers and content for all methods.

For v0.1.2 I should be adding more to the server reponse options.

TestHTTPServer v0.1.0

I've pushed a new PyPi package TestHTTPServer based on some work at AWeber where I needed to test processes that make web requests.

This probably shouldn't be used right away as it was from memory that I created the package. It will be expanded to record more of the request data and handle more requests.

Plug v0.1.0

I've pushed the first version of Plug. You can download it with pip

pip install plug

This is mostly to help me manually test Plug by using it in a few of my own projects and getting together a list of glaring issues.

Google Music

I've been using Google Music quite a bit lately. It was rather painless to setup although the upload process can take a while. The service works out well with my Macbook Air since I only have a 64GB SSD and don't want to fill my drive with music I may or may not be listening to very frequently.

The only issue for me is that the Flash performance can frequently cause studdering if I am doing too much in Chrome at the same time.

Worker Process

The main design is splitting out how worker's should run from what they are supposed to be doing. Extending the BaseWorker class will give a main classmethod used to start new workers.

The BaseWorker class is split from the WorkerRunner class giving a common interface for all workers and Unixy interaction.

You can take a look at what I have so far on the projects page.

Packaging

Roughly 2 weeks ago I started Plug which aims to create a package format for Python daemons. The project started after seeing how Supervisor handles 150+ processes.

A current project at work can easily have many daemon processes with differing number of running instances that may need to be adjusted frequently. Deploying with Supervisor can be a problem given the amount of time Supervisor would take to start/stop processes.

Plug installs each package into a virtualenv then uses runit to manage each daemon instance.

I have a prototype version working now with a packaged version to come in the next few weeks after giving it more testing.

The biggest issue is after watching To Package or Not to Package I am falling in more of the "to package" crowd and despite Plug being a packaging solution smells a bit too much of NIH.

Python TDD with Dingus - A Markdown Function

It should take 2 paths, the source directory and the output directory.

Acceptance Test

In acceptance tests you will want to test as large a feature as possible. The test in this case will assert that a file in the src directory is converted to HTML in the output directory.

Since we want this to create the files on disk we will need to import os.path for some helper functions.

{% highlight python %} import os.path {% endhighlight %}

Then we import the function we plan on testing.

{% highlight python %} from markdown_processor import process_markdown {% endhighlight %}

From there we can begin by creating the test class that will handle all of our setup and our assertion.

{% highlight python %} class WhenRunningProcessor(object):

    @classmethod
    def setup_class(cls):
        cls.src_dir = 'src_example'
        cls.target_dir = 'target_dir'

        process_markdown(cls.src_dir, cls.target_dir)

{% endhighlight %}

This creates a class WhenRunningProcessor That inherits from object. Before each test case Nose allows us to run code to setup the test. In this case we use the @classmethod decorator and setup_class. This function will be run once before all tests in this class. Acceptance tests will take longer to run then unit tests and usually do not require the same level of isolation.

Then we define the src_dir and target_dir since we will be using them a few times.

Finally we run the function we plan on testing passing in the src_dir and target_dir.

Now we can write the acceptance test for this function

{% highlight python %} def should_have_html_hello_world(self): file_path = os.path.join(self.target_dir, 'hello_world.html') content = open(file_path, 'r').read() assert '

Hello World!

' in content {% endhighlight %}

Our test is checking that the text '<p>Hello World!</p>' is in a file in the output directory. This requires some fixture data in the source directory which is only

Hello World!

The file_path is the our target_dir folder and the hello_world.html file. The file is read and then an assertion checking that the test exists.

The whole test file will look like this:

{% highlight python %} import os.path

from markdown_processor import process_markdown

class WhenRunningProcessor(object):

    @classmethod
    def setup_class(cls):
        cls.src_dir = 'src_example'
        cls.target_dir = 'target_dir'

        process_markdown(cls.src_dir, cls.target_dir)

    def should_have_html_hello_world(self):
        file_path = os.path.join(self.target_dir, 'hello_world.html')
        content = open(file_path, 'r').read()
        assert '<p>Hello World!</p>' in content

{% endhighlight %}

Unit Tests

In our unit tests we will test how we plan to implement this functionality. The mocking library Dingus will allow us to isolate the our function from the OS and our other libraries. After the function is run we can test to make sure the code works how we expected it to.

First will will import everything we need for the test

{% highlight python %} from dingus import Dingus, DingusTestCase

from markdown_processor import process_markdown
import markdown_processor as mod

{% endhighlight %}

The dingus library provides us with a Dingus class which we will use to assert what our function is doing and the DingusTestCase will automatically isolate our function.

As well as the function we want to test we also import the module to help us make assertions about what goes on outside the function.

Now we can setup a base class to use for our tests. This will hold the common elements we use for our tests.

{% highlight python %} class BaseProcessing(DingusTestCase(process_markdown)):

    def setup(self):
        super(BaseProcessing, self).setup()
        self.src_dir = Dingus('src_dir')
        self.target_dir = Dingus('target_dir')

        mod.os.listdir.return_value = ['hello_world.markdown']
        mod.os.path.splitext.return_value = ('hello_world', 'markdown')

        self.md = mod.markdown.Markdown()

{% endhighlight %}

Our BaseProcessing class inherits from DingusTestCase while passing in our function. This will isolate our function and replace everything around it with Dingus objects. The setup method will be run before each test, unlike the acceptance tests where setup_class was only run once. We once again setup our src_dir and target_dir but this time use Dingus objects while passing in a helpful name. In this case the strings for our directories shouldn't be modified but this allows you to verify any operations performed on the arguments.

Next we set the return values of some function we plan on using. Remember everything but our function is a Dingus which only return a new Dingus when called. The listdir function will return a list of entries in the directory. In this case we just want to return a list with 1 file. We will need to get the filename without an extension so we will set the return_value to be a tuple of the name and extension.

The last part is setting a shorter name for the Markdown instance.

Test Cases

Of first test is to assert that we see if the target directory exists.

{% highlight python %} def should_check_existance_of_target_dir(self): assert mod.os.path.calls('exists', self.target_dir) {% endhighlight %}

We will use os.path.exists to check. The assert uses the .calls method to see if os.path.exists was called with our target directory.

Next will want to assert that a markdown instance is created

{% highlight python %} def should_create_markdown_instance(self): assert mod.markdown.calls( 'Markdown', extensions=['codehilite'] ).once() {% endhighlight %}

This time we check that the object is called with the extension 'codehilite' and that it is only called once.

Now we need to get the files we plan on converting so we will use the os.listdir that uses our mocked return value.

{% highlight python %} def should_find_markdown_files(self): assert mod.os.calls('listdir', self.src_dir) {% endhighlight %}

Finally we can test that our file is actually converted by markdown

{% highlight python %} def should_join_source(self): assert mod.os.path.calls('join', self.src_dir, 'hello_world.markdown')

def should_join_target(self):
    assert mod.os.path.calls('join', self.target_dir, 'hello_world.html')

def should_convert_files(self):
    in_file = mod.os.path.join()
    out_file = mod.os.path.join()
    assert self.md.calls('convertFile', in_file, out_file)

{% endhighlight %}

We first assert that our function properly creates the paths to use for the input and output files then calls markdown's convertFile function. The in_file and out_file are actually the same dingus in this case, the join() will return an object we can use to test with.

Now that the tests are written we can add the classes to run them.

{% highlight python %} class WhenProcessingMarkdown(BaseProcessing):

    def setup(self):
        BaseProcessing.setup(self)
        mod.os.path.exists.return_value = False
        process_markdown(self.src_dir, self.target_dir)

    def should_create_target_directory(self):
        assert mod.os.calls('mkdir', self.target_dir)

{% endhighlight %}

WhenProcessingMarkdown inherits from our BaseProcessing class including all of the test cases. In our setup this time we want to make sure we run BaseProcessing.setup with our instance and set the return value of os.path.exists. After that we run our function.

Since our output directory doesn't exist in this test it will want to make sure that we try to create it.

Because we need to test the other case where the directory already exists we create a new test class where that is true.

{% highlight python %} class WhenProcessingMarkdownAndDirectoryExists(BaseProcessing):

    def setup(self):
        BaseProcessing.setup(self)
        mod.os.path.exists.return_value = True
        process_markdown(self.src_dir, self.target_dir)

    def should_not_create_target_directory(self):
        assert not mod.os.calls('mkdir')

{% endhighlight %}

This is very similar except that the exists() return value is True so we will want to assert that we do not call mkdir at all.

Implementation

Our tests are complete and we can move on to our implementation.

{% highlight python %} import os

import markdown

def process_markdown(src_dir, target_dir):
    "Converts files from :param src_dir: to html in the :param target_dir:"
    if not os.path.exists(target_dir):
        os.mkdir(target_dir)

    md = markdown.Markdown(extensions=['codehilite'])

    for file in os.listdir(src_dir):
        name, ext = os.path.splitext(file)

        in_file = os.path.join(src_dir, file)
        out_file = os.path.join(target_dir, name + '.html')
        md.convertFile(in_file, out_file)

{% endhighlight %}

The implementation shouldn't be surprising after planning it out. Our function takes 2 arguments, possibly creates the target directory, then converts each file from our source directory.

You can download the full source

Test Sizes

Testing isn't covered much in school. If your projects are only 2 weeks long and disappear after you've handed them in, testing doesn't prove it's worth. When tests are mentioned it's never anything practical, never how to test certain features, what parts are even worth testing?

My current project has 3 sets of tests, unit, integration, and acceptance tests. Google has small, medium, and large tests which roughly correspond to what we write.

Acceptance Tests

These are written first to tell us what we want from a feature. Usually these are structure as interaction stories of the Cucumber variety.

{% highlight bash %}
When a customer has a list of subscribers
And sends a message to the list
The subscribers receive the message
{% endhighlight %}

A common example, definitely a feature you want working. The tests are a little more complicated than that, including assertions to make sure certain steps execute as anticipated.

Integration tests

Integration tests started as a need to test various states of our database and queuing systems. They generally involve testing components to make sure they behave when everything isn't perfect.

What happens when a worker dies in the middle of a job?

Can another worker pick up the job immediately or does it require human intervention?

These stories enumerate the states our models can occupy. When an event comes in from the queue it may be the first, a duplicate, or a re-fired event from a failed worker. A simplified case for sending an email that can be either :

{% highlight bash %}
New
Trying to send
Sending
Sent
{% endhighlight %}

Our tests classes would cover the cases:

{% highlight bash %}
When sending a new email
When sending an email that is trying to send
When sending an email that is sending
When sending a sent email
{% endhighlight %}

Once we have the tests we can figure out which states are possible to recover from. These cases are much more difficult to cover in acceptance tests. They would require excessive test setup and extensive knowledge of our system.

Unit tests

We write unit tests to verify our implementation. This means mocking out as much as possible and isolating individual objects and functions. We use Dingus as our mocking library. At this point the feature is well defined and important test cases are covered, we can focus on how we want to solve them.

{% highlight bash %}
When sending an email
If email is New
Generate and send an email

When sending an email
If email is Trying to send, Sending, or Sent
Do not try and send the email
{% endhighlight %}

With tests of each size we gain greater confidence that our code is working and you can use them in a month when you forget how or why things were done this way.