Category: Uncategorized

Inigo talked about testing with Selenium in our developer meeting.

The detail is in the slides – but the high-level overview is that Selenium is a useful tool for producing automated tests for web application behaviour, but it can be hard to write solid tests that aren’t flakey and dependent on timing. It helps a lot to use an API approach – writing a class for each page in the web application and the operations it exposes, and then separately writing the tests to exercise that API. This leads to much cleaner and easier to maintain code than if you just write tests directly against the pages.

Dev meeting Selenium slides

Chris was recently at the Agile on the Beach 2015 conference, and this Friday he reported back to the rest of the dev team on some of the talks that he went to there.

The death of continuous integration

Steve Smith talked about “the death of continuous integration” – CI is a cultural thing, not just Jenkins. The key questions are: whether everyone commits something to trunk every day, and whether problems in the build get fixed quickly. He described two build models – synchronous and asynchronous – the former is waiting for the build to complete before continuing, and the latter is allowing the build to continue, and swapping back if the build breaks.

He discussed three branching models:

* “Long lived feature branches” – these are generally a nightmare.
* “Short-lived feature branches” – fine if everything works as it should, but reviewers aren’t always available immediately, and you might need to switch context back to an old branch if reviewing takes a while. There’s also less of an incentive to fix the build, because it doesn’t immediately effect you. It’s also more painful to change something that affects many, many files.
* “Trunk based development” – apparently Google do this. Everything is committed to trunk, all the time, there are no separate code reviews, and half-developed features are disabled via feature toggles. Large scale changes can be handled by putting the change behind an interface, and then switching code behind the interface.

We discussed these models. Several people expressed scepticism about trunk-based development – in their experience, committing directly to trunk has generally led to long pauses between commits, rather than shorter development ccycles. We use short-lived feature branches – and while we do sometimes have problems with huge changes, we don’t typically have a problem with people not dealing with broken builds (because our culture is that the build should be fixed quickly).

Testing in Production

In a talk about Testing in Production by Unruly, they described that their process has teams responsible for specific features from end-to-end. This encourages developers to make the code more robust, because the dev team are the people responsible for it in production and will be woken up by pagers if it goes wrong.

They test their code in production, because it is too much work to maintain an exact mirror of the live environment. They have no QA environments. Half-developed features are kept out of production via “feature toggles” and “data toggles”. They have lots of monitoring on their live servers – they monitor things that are of value to the company – like “are we making money” rather than “have the servers gone down”. “Monitoring driven development” is about first writing a monitor to check your new feature, and then developing the code for it – similar to test-driven development but much more so. They also use a chaos-monkey style approach for testing – with badly behaved client code, load testing with extreme events – because if someone is going to break the live environment, it might as well be them. They also use mob programming.

We discussed this – many of these approaches seem interesting and appropriate for a company that has very granular income, from a very large number of clients, that is coming in quickly and very responsively. They seem less relevant to more traditional companies.

Management 3.0

Pia-Maria Thorens spoke about delegation poker. Chris showed a set of “delegation poker” cards, that showed the continuum of delegation decisions between “the manager makes the decision on his own” and “the team makes the decisions entirely without the manager”.

We discussed this as an interesting way of thinking about management and delegation.

I’ve just resolved a very confusing bug in some of our MarkLogic XQuery code. It ultimately turned out to be very simple to fix – just adding an ‘indent=no’ parameter to the XSLT.

However, there were three unexpected bits of behaviour that I encountered along the way, that made fixing it much harder than it should have been.

First unexpected thing – MarkLogic was defaulting to ‘indent=yes’ in its XSLT processing.

Second unexpected thing – MarkLogic is applying its indent setting to the results of <xsl:copy-of select=”.”/>. I expected <textarea><xsl:copy-of select=”.”/></textarea> to give me an exact copy of the input, but it doesn’t, because indent is applied to it. Saxon doesn’t apply indent here, even if indent is set to yes.

Third unexpected thing – MarkLogic’s QConsole can’t be trusted to display whitespace accurately. I displayed two almost identical documents in exactly the same way, via doc(‘blah.xml’), in the console, and one of them was indented, and the other one wasn’t. These were the before-and-after documents for our ‘splitDocument’ function, so I naturally thought that splitDocument was applying indentation. It was only when I wrote Scala code to retrieve the XML that I saw what was happening.

Lessons:

1) Always turn on indent=’no’ in stylesheets used in MarkLogic
2) Don’t trust QConsole to debug whitespace issues. Instead, use code.

At last week’s dev meeting we swapped war stories about problems encountered in production, how we tracked down the root cause and lessons we learnt. Here are some highlights:

• Environment is very important. When trying to reproduce a bug, you need to try to replicate the environment as closely as possible. Even very minor changes to libraries, frameworks, subsystems or operating system patches can make a major difference to whether the bug is seen or not.
• “We haven’t touched anything”. The person telling you that nothing has changed may well believe this is the case. However, if a system mysteriously stops working it’s best to actually verify that this is the case. There was a bit of moaning about Windows at this point. At least with text file based configuration you can easily do a diff. Not so easy when someone has unchecked a vital property in Windows config settings.
• Test your systems like a real user. We discussed a problem on a website where the functionality in question had been tested thoroughly by people within the organisation. Unfortunately, when real external users came to use the site, it didn’t work as expected. This was due to a private address being used that could only be seen within the corporate network. So it worked for internal users, but not for real customers. If it had been tested via an external network this would have come to light before real users hit problems.
• The bug may be in old code. It’s possible that the bug you are seeing has been lying dormant for years and is being triggered by some innocuous change. We talked about a situation where a new release would cause a site to have severe performance problems. Much time was spent looking at all the changes going into that release to see if there was some change to database access or the like causing the problem. In the end it transpired that a small change to the cookies used was triggering a latent bug in a script used for load balancing. This script had been running fine in production for years until this seemingly minor change caused it fork processes like crazy and bring down the site.

In our dev meeting this week, we discussed the Reactive Functional Programming Scala course from Martin Odersky, that Chris has been doing. Several others of us have signed up for it, but haven’t been very good at actually doing the course. Chris told us what he had been doing, and what he had learned (aside from being impressed by Eric Meijer’s shirts).

Reactive Programming is:

• Event-driven – asynchronous with no blocking
• Scalable – across multiple cores and machine
• Resilient – against failure
• Responsive – it is responsive to the user

The RFP course has been about techniques for doing reactive programming that wrap these concepts neatly so the complexity of them is hidden and they can easily be composed.

Values can be separated along two axes – single/multiple and synchronous/asynch. Hence, in terms of Scala types:

• a single synchronous value – a Try[T]
• a single asynch value – a Future[T]
• multiple synchronous values – Iterable[T]
• multiple asynch values – Observable[T]

The “multiple asynch” area is where RFP fits.  An “Observable” is something to which you can give an Observer, and then it will send callbacks to your Observer. For example, a set of full pint glasses (i.e. a Round) on a table might be an Observable – and then the Round might call back to your Observer to inform it about each pint. You could combine this with another Observable – for example each pint can itself be an Observable that informs you when it is empty. Hence, you can identify when a new round needs to be bought by composing the Round observables with each Pint observable, and receving all the messages from each.

Another part of the course is using Akka. Akka is an actors library for Scala (used in Play). Actors are autonomous pieces of code that each individually run synchronously, and communicate via message passing. The core of an Akka actor is a receive method that takes a message and carries out some activity. It’s also possible to use “become” on an Akka actor to transform the actor to have different behaviour – this is a way of better managing the mutable state within the actor (you can “unbecome” as well…). Error handling for actors is interesting, because actors are responsible for their children – if a child actor dies, then the parent actor’s supervision strategy determines what should happen (e.g. throwing the child away and recreating a new one, or entering a suicide pact to have the parent actor die if the child actor dies, or sending poison pills to destroy other actors, and so on). It all gets quite macabre.

One of the course projects was to build a binary tree in which each node is an actor. If a node is removed, then it instead marks itself as removed, and a separate garbage collection process actually takes nodes out. Akka manages the ordering of messages, so once you get the basic logic right, it works very quickly. But – it’s hard to debug.

A surprising feature of Akka is that you can create messages of any type – this is in contrast to Scala’s normal philosophy of making everything as strongly typed as possible. It’s recommended to use case classes as messages to provide better typing.

I noticed yesterday that the output from some integration tests running on our CI server were producing large amounts of output. It turned out that 41000 lines of the 43000 lines were coming from a single test. The reason for this is the way Specs2 handles the contain() matcher with lists of Strings. It means that the following:

listOfIds must not(contain(badId))

is effectively the following, checking each string to see if it contains the given one:

listOfIds(0) must not(contain(badId))
listOfIds(1) must not(contain(badId))
listOfIds(2) must not(contain(badId))
....

So if you are looking at a long list, this yields some very verbose output. With types other than String, this seems to work as expected.

We use Selenium for testing front-end web pages. My colleague Chris Rimmer just added code to take advantage of a Selenium feature I hadn’t previously been aware of – now, the tests will automatically take a screenshot of the browser if they fail, which should help us in tracking down issues.

A feature of Scala that I hadn’t used before was Stream.cons. This allows you to create a stream (essentially, a lazily evaluated list) from a function. So, for doing some work on files based on their position in the directory hierarchy, we can create a list of their parents:

def fileParents(file: File) : Stream[File] = {
  val parent = file.getParentFile
  if (parent == null) Stream.empty else 
    Stream.cons(parent, fileParents(parent))
}

and then use standard Scala functionality for filtering and finding things in lists, rather than having to write code that iterates through the parent files manually.

In our developer meeting this week, we discussed parsing, and particularly parser combinators.

We’ve used the Scala parser combinator library in the past for parsing search query syntax – for example, to support a custom search syntax used by a legacy system and convert it into an XQuery for searching XML. We’ve also used Parboiled, a Java/Scala parser library, for parsing geographic latitude and longitude values from within scientific journal articles about geology. We’ve done simpler parsing with regular expressions in C# to identify citations within text like “(Brown et al, 2012)” and “(Brown and Smith, 2010; Jones, 2009)”.

The parser combinator approaches are typically better than using a traditional parsing method like Lex and YACC or JavaCC, because they’re written in the host language (e.g. Java or Scala), and so it’s much easier to write unit tests for them and to update them easily. They’re particularly approachable in Scala, because Scala’s support for domain-specific languages means that you can write code that looks like:

  “{” ~ ( comment | directive ) ~ “}”

where the symbols like ~ and | are Scala method invocations – which means that you can focus on the parsing, rather than the parser library syntax.

We briefly discussed where it makes sense to use regular expressions for parsing, and where it makes sense to use a more powerful parsing approach. We agreed that there was a danger of creating overly complex regular expressions by incremental “boiling a frog” extensions to an initially simple regex, rather than stopping to rewrite using a parser library.

For further processing of the content once it’s been parsed, we discussed using the Visitor pattern. For example, having created an abstract syntax tree from a search query, it’s useful to use a visitor approach to turn that tree into a pretty printed form, or into an HTML form for display, or into a query language form suitable for the underlying datastore.

We use Git Flow as our VCS process, which means that we develop on feature branches and merge those branches back into the master branch as part of our code review. It’s useful to delete these branches as they’re merged, because then anyone can see what is being worked on or needs code review by listing remote branches without being distracted by old branches. However, it’s easy for a reviewer to forget to delete the branches when they do the merge. These Git commands delete old branches that have already been merged with the “master” branch:

Delete local branches:

git branch –merged master | grep -v master | xargs -n1 git branch -d

Remove local tracking branches of remote branches that have already gone:

git prune

Remove remote branches that have been merged:

git branch -r –merged master | sed “s#origin/##” | grep -v master | xargs -n1 git push origin –delete