It's a blog

Month: August 2015

Misled by PHPUnit at() method

So it turns out the at() method doesn’t quite do what I had initially thought….

I have recently been working on some tests for the new Newsletter extension for Mediawiki, specifically to test the NewslettterTablePager class. This thing extends the TablePager class in Mediawiki which is designed to make displaying information from a database table on a special page on a mediawiki site easy, and also easily enable things such as sorting.

The code interacts with the database and gets a ResultWrapper object, and the Pager uses the numRows(), seek() and fetchObject() methods, all of which I thought would be incredibly simple to mock.

Attempt 1

My first attempt where I first notice I have been thinking about the at() method all wrong can be seen below:

This methods returns a mock Database that the Pager will use. As you can see the only parameter is an array of objects to be returned by fetchObject() and I am using the at() method provided by phpunit to return each object at the index that it is stored in the array. This is when I discovered that at() in phpunit does not work in the way I first thought…

at() refers to the index of calls made to the mocked object as a whole. This means that in the code sample above, all of the calles to numRows() and seek() are increasing the current call counter index for the object and thus my mocked fetchObject() method is never returning the correct value or returning null.

Attempt 2

In my second attempt I made a guess that phpunit might allow multiple method mocks to stack and thus the return values of those methods be returned in the order that they were created. Thus I changed my loop to simply use any():

But of course this also does not work and this result in the same $resultObject being returned for all calls.

Final version

I ended up having to to do something a little bit nasty (in my opinion) and use returnCallback() and use a private member of the testcase within the callback as a call counter / per method index:


It would be great if phpunit would have some form of per method index expectation!

Some of the code examples here are slightly cut down, the full change can be seen at in the NewsletterTablePagerTest file.

If only I had Googled the issue sooner!

Removing use of Mediawiki’s API ‘rawmode’ in Wikibase

Rawmode was a boolean value used to determine if an API result formatter in Mediawiki needed extra metadata in order to correctly format the result output. The main use of said metadata was in the XML output of the Mediawiki API. How hard can removing it be? This is the story of the struggle to remove the use of this single boolean value from the Wikibase codebase.


The first commit for this task was made on the 6th July 2015 and the final commit was about to be merged on the 27th August. So the whole removal took just under 2 months.

During this two months roughly 60 commits were made and merged working towards removal.

Overall 9290 lines were removed and 5080 lines were added.

I’m glad that is all done. (This analysis can be found on Google sheets). Sorry there are not more pictures in this post…..

Reason for removal

Well, rawmode is being remove from Mediawiki to remove API complexity. Instead of having to check what the API formatters need they will instead just accept all metadata and simply use what they need and discard the rest.

The change to “Finish killing ‘raw mode'” can be seen on Gerrit and has been around since April of this year. The relevant task can be found on Phabricator.

Process overview

The first step on the path was to remove the old serialization code from Wikibase (otherwise known as the lib serialization code) and replace all usages with the new WikibaseDataModelSerialization component. This component was already used in multiple other places in the code but not in the API due to its reliance on the way the lib serialization code handled the rawmode requirement of the API at the time.

Removal of the lib serialization code was the the first of the two major parts of the process and after around 50 commits I managed to remove it all! Hooray for removing 6000 lines with no additions in a commit…

The next and final step was to make the ResultBuilder class in Wikibase always provide metadata for the API and to remove any dirty hacks that I had to introduce in order to kill the lib code. Again this was done over the course of multiple commits, mainly adding tests for the XML output which at the time was barely tested. Finally a breaking change had to be made to remove lots of the hacks that I had added and the final uses of raw mode.

The final two commits can be seen at and

Final notes

Look! You can even see this on the GitHub code frequency graph (just….)

You can also find my draft API break announcement post here.

— edit —

It looks like I did break 1 thing incorrectly: , thought a fix is on it’s way our to the beta site! :)

© 2017 Addshore

Theme by Anders NorenUp ↑