Over the last year, I’ve been helping (part-time) Freeset build their ecommerce website. David Hussman introduced me to folks from Freeset.
Following is a list of random topics (most of them are Agile/XP practices) about this project:
Project Inception: We started off with a couple of meetings with folks from Freeset to understand their needs. David quickly created an initial vision document with User Personas and their use cases (about 2 page long on Google Docs). Naomi and John from Freeset, quickly created some screen mock-ups in Photoshop to show user interaction. I don’t think we spent more than a week on all of this. This helped us get started.
Technology Choice: When we started we had to decide what platform are we going to use to build the site. We had to choose between customer site using Rails v/s using CMS. I think David was leaning towards RoR. I talked to folks at Directi (Sandeep, Jinesh, Latesh, etc) and we thought instead of building a custom website from scratch, we should use a CMS. After a bit of research, we settled on CMS Made Simple, for the following reasons
We needed different templates for different pages on the site.
PHP: Easiest to set up a PHP site with MySQL on any Shared Host Service Provider
Planning: We started off with an hour long, bi-weekly planning meetings (conf calls on Skype) on every Saturday morning (India time). We had a massively distributed team. John was in New Zealand. David and Deborah (from BestBuy) were in US. Kerry was in UK for a short while. Naomi, Kelsea and other were in Kolkatta and I was based out of Mumbai. Because of the time zone difference and because we’re all working on this part time, the whole bi-weekly planning meeting felt awkward and heavy weight. So after about 3 such meetings we abandoned it. We created a spreadsheet on Google Docs, added all the items that had high priority and started signing up for tasks. Whenever anyone updated an item on the sheet, everyone would be notified about the change.
User Stories: We started off with User Persona and Stories, but soon we just fell back to simple tasks on a shared spreadsheet. We had quite a few user related tasks, but just one liner in the spread sheet was more than sufficient. We used this spreadsheet as a sudo-backlog. (by no means we had the rigor to try and build a proper backlog).
Short Releases: We (were) only working on production environment. Every change made by a developer was immediately live. Only recently we created a development environment (replica of production), on which we do all our development. (I asked John from Freeset, if this change helped him, he had mixed feelings. Recently he did a large website restructuring (added some new section and moved some pages around), and he found the development environment useful for that. But for other things, when he wants to make some small changes, he finds it an over kill to make changes to dev and then sync it up with production. There are also things like news, which makes sense to do on the production server. Now he has to do in both places). So I’m thinking may be, we move back to just production environment and then create a prod on demand if we are plan to make big changes.
Testing: Original we had plans of at least recording or scripting some Selenium tests to make sure the site is behaving the way we expected it to. This kind of took a back seat and never really became an issue. Recently we had a slight set back when we moved a whole bunch of pages around and their link from other parts of the site were broken. Other than that, so far, its just been fine.
Evolutionary Design: Always believed in and continue to believe in “Do the Simplest, Dumbest, thing that could Possibly work“. Since we started, the project had taken interesting turns, we used quite a lot of different JavaScript libraries, hacked a bit of PHP code here and there. All of this is evolving and is working fine.
Usability: We still have lots of usability and optimization issues on our site. Since we don’t have an expert with us and we can’t afford one, we are doing the best we can with what we have on hand. We are hoping we’ll find a volunteer some day soon to help us on this front.
Versioning: We explored various options for versioning, but as of today we don’t have any repository under which we version our site (content and code). This is a drawback of using an online CMS. Having said that so far (been over a year), we did not really find the need for versioning. As of now we have 4 people working on this site and it just seems to work fine. Reminds me of YAGNI. (May be in future when we have more collaborators, we might need this).
Continuous Integration: With out Versioning and Testing, CI is out of question.
Automated Deployment: Until recently we only had one server (production) so there was no need for deployment. Since now we have a dev and a prod environment, Devdas and I quickly hacked a simple shell scrip (with mysqldump & rsync) that does automated deployment. It can’t get simpler than this.
Hosting: We talked about hosting the site on its own slice v/s using an existing shared host account. We could always move the site to another location when our existing, cheap hosting option will not suit our needs. So as of today, I’m hosting the site under one of my shared host account.
Rich Media Content: We questioned serving & hosting rich media content like videos from our site or using YouTube to host them. We went with YouTube for the following reasons
We wanted to redirect any possible traffic to other sites which are more tuned to catering high bandwidth content
We wanted to use YouTube’s existing customer base to attract traffic to our site
Since we knew we’ll be moving to another hosting service, we did not want to keep all those videos on the server which then will have to be moved to the new server
Customer Feedback: So far we have received great feedback from users of this site. We’ve also seen a huge growth in traffic to our site. Currently hovering around 1500 hits per day. Other than getting feedback from users. We also look at Google Analytics to see how users are responding to changes we’ve made and so on.
We don’t really have/need a System Metaphor and we are not paying as much attention to refactoring. We have some light conventions but we don’t really have any coding standards. Nor do we have the luxury to pair program.
Distributed/Virtual Team: Since all of us are distributed and traveling, we don’t really have the concept of site. Forget on-site customer or product owner.
Since all of this is voluntary work, Sustainable pace takes a very different meaning. Sometimes what we do is not sustainable, but that’s the need of the hour. However all of us really like and want to work on this project. We have a sense of ownership. (collective ownership)
We’ve never really sat down and done a retrospective. May be once in a while we ask a couple of questions regarding how something were going.
Overall, I’ve been extremely happy with the choices we’ve made. I’m not suggesting every project should be run this way. I’m trying to highlight an example of what being agile really means.
I don’t like the I<something> naming convention for interfaces for various reasons.
Let me explain with an example. Lets say we have IAccount interface. Why is this bad?
All I care is, I’m talking to an Account, it really does not matter whether its an interface, abstract class or a concrete class. So the I in the name is noise for me.
It might turn out that, I might only ever have one type of Account. Why do I need to create an interface then? Its the speculative generality code smell and violating the YAGNI principle. If someday I’ll need multiple types of Accounts, I’ll create it then. Its probably going to be as much effort to create it then v/s doing it now. Minus all the maintenance overhead.
Let’s say we’ve multiple types of Accounts. Instead of calling it IAccount and the child classes as AccountImpl1 or SavingAccountImpl, I would rather refer to it as Account and the different types of account as SavingAccount or CreditAccount. It might also turn out that, there is common behavior between the two types of Account. At that point instead of having IAccount and creating another Abstract class called AbstractAccount I would just change the Account interface to be an abstract class. I don’t want to lock myself into an Interface.
Personally I think the whole I<something> is a hang-over from C++ and its precursors.
Some people also argue that using the I<something> convention is easy for them to know whether its an interface or not (esp. coz they don’t want to spend time typing newAccount() and then realizing that its an interface).
The way I look at it is good IDEs will show the appropriate icon and that can help you avoid this issue to some extent. But even if it did not, to me its not a big deal. The number of times the code is read is far more than its written. Maintaining one extra poorly named Interface is far more expensive than the minuscule time saved in typing.
P.S: I’m certainly not discouraging people from creating interfaces. What I don’t like is having only one class inheriting from an interface. May be if you are exposing an API, you might still be able to convenience me. But in most cases people use this convention through out their code base and not just at the boundaries of their system.
In lot of cases I find myself starting off with an Interface, because when I’m developing some class, I don’t want to start building its collaborator. But then when I start TDDing that class, I’ll change the interface to a concrete class. Later my tests might drive different types of subclass and I might introduce an Interface or an Abstract class as suitable. And may be sometime in the future I might break the hierarchy and use composition instead. Guess what, we are dealing with evolutionary design.
On the Protest project when I was building the “integration with Ant” feature, I adopted the same thin slice principle. Following are the thin slices I came up with:
Create an Ant task which can call Protest, it simply returns the tests in the same order as given to it. (Essentially was a copy of the JUnit Ant task)
Add support for a voter (happened to be Dependency Voter), so that we can actually prioritize the tests based on the dependency algorithm. At this point we went ahead and released this task
Add support for multiple voters. By now we had created 3 different voters and we wanted to use all the voters
Provide a way to specify weightage for each voter. Some voters should be able to influence the prioritized list of test more than the others
Once we have a prioritized list of tests, provide a way to specify what top percentage of that list should be executed. This provides the user tighter control over how much feedback they need depending on the type of change they have just made
And so on…
Now we could have sat and first designed how the Ant task should look and later wrote the task and then integrated with Protest. But the problem with this approach is:
We won’t have anything functional and usable until we finish all the tasks. Too scary from a feedback standpoint
We won’t be able to test anything for real until we finish all the tasks. There is a huge risk involved in this approach
Essentially we are building an inventory with each of those tasks that are not used immediately. Turn around time for a feature is high
It requires a lot of upfront thinking, which I’m not generally good at. At least we’ll have to think through all the input each voter would need and so on. Right now we don’t even have all the voters in place and this forces us to think about them now or introduces an unwanted dependency now.
Lots of people argue that the evolutionary approach will be less efficient (more expensive and time consuming) because it gives an impression of thrashing and rework. In my experience, the big upfront design leads to more rework generally. It creates an illusion of streamlined process but in reality, it is actually lot more work and also leads to a rigid and over-engineered design
We can also add all the disadvantages of big upfront design here
I hope this example demonstrates the technique of thin slicing and its advantages over its alternatives.
For a given feature, we can come up with multiple thin-slices which can be incrementally (sequentially) built.
Thin Slice is the simplest possible functional, usable, end to end slice of functionality.
Thin Slicing a feature is not a new concept. Generally development teams consider the simple happy path scenario as the first thin slice. The alternative flows are considered as subsequent slices. Today when I think of a thin slice, its slightly more sophisticated than just considering happy path and alternative paths.
Let me explain with an example: Lets consider we are building a web app and we need a feature which allows the user to upload their profile picture.
We can come up with the following thin slices:
User can upload any photo (predefined set of image format) - won’t look good coz the image size can vary and hence where ever its displayed won’t align well. Might not have AJAX support (depends on what is simple and quick to do). All the bells and whistles are pushed for later.
Build an image scaling facility so that we can reduce the image resolution and hence its size
Provide an image cropping facility so that users can crop their profile images
Instead of uploading an image from my disk, provide a facility to pull it from the web. The user provides the URL
and so on….
Each slice is functional (end to end, we are not just doing the UI or backend bit). This is great for internal feedback. Might not be good enough for a public release. Esp. with just the first slice. In some sense we are incrementally add more power to the feature with each thin slice. From another perspective it feels like we are iterating over the feature and refining the feature as we go along. In either case, We build the feature by adding one by one the thin slice, till the point we feel we can release this feature. Post release we can continue to add more slices.
It is also possible to come up with thin slices which are perfectly releasable. Lets take the same example and see how we could come up with thin slices which are releasable.
User can upload a photo (predefined set of image format + clear message showing the expected size and resolution). Any image which does not qualify this criteria is rejected.
Knock off the resolution constraint and build image scaling facility so that we can reduce the image resolution and hence its size
Knock off the image size constraint and provide an image cropping facility so that users can crop their images to fit the right size
Please note that the concept of thin slice is also applicable at a much broader scope than just at a feature level. For Ex: On a given project we use the same concept to plan our small & frequent releases. At a portfolio level we apply the same principle to break projects into smaller loosely coupled projects and then prioritize them.
