On Wed, Nov 16, 2011 at 3:15 AM, Owen Davis owen@wikia-inc.com wrote:
John Du Hart <compwhizii <at> gmail.com> writes:
It's just another style I've encountered on other projects and I personally like.
The syntax itself is fine, but at Wikia we have found (after a recent post mortem) that out of 23 "Fatal Error" code defects found in production, 7 of them were due to method calls on null object references. If any method in that chain returns null then the request fails. It most cases core MediaWiki objects do return a valid stub object of some kind, but not all of them do (and in some cases they return null. intermittently so the code works "most of the time", which is in many ways the worse scenario). Introducing a pattern like this in a code base this large is therefore problematic. The "clean looking code" benefit is perhaps outweighed by the fact that you need to add extra "if" conditions or try/catch blocks everywhere to handle local null object exceptions. In our case, we are trying to ensure during code reviews that we just check for null objects in all "if" conditions, which does also lead to more highly nested and less readable code.
Any individual interface should generally have one of three patterns:
A) cannot fail, ever -- always returns a live object B) always returns a live object -OR- throws an exception C) returns a live object -OR- null/false as error
The most annoying problems seem to come from treating C code as if it's A: everything works fine until you come across some invalid input, and then somewhere else in your code you die, utterly surprised.
The thing I like about B) is that failing to handle the error case gives more immediate feedback: the error output & backtrace tell you where the *thing that didn't find its data* happened, instead of where some other code tried to use that previously-fetched data.
My ideal unicorn-magic world tends to fall mostly in the A-B spectrum (though exceptions do not always make sense in every case.).
Throwing chaining onto things like OutputPage could be done but seem unlikely to be worthwhile; the operations are generally independent of each other.
Chaining is good for things that work conceptually like a filter chain. For instance in jQuery code we'll often do something like:
$(element).find('a') .click(magicClickHandler) .attr('title', 'Click for magic!');
The real magic actually came from the progress from $(element) to .find('a') and then to performing operations on the resulting collection.
There's a much smaller difference when there's a single variable as a common prefix:
$foo .click(magicClickHandler) .attr('title', 'Click for magic!');
vs
$foo.click(magicClickHandler) $foo.attr('title', 'Click for magic!');
is only slightly "cleaner".
Everything is a tradeoff, but IMHO you do need to check for null in chains like that, which means you can't really get away with it (at least not for long).
jQuery's chaining API is explicitly built to make sure that you don't -- it's style A in the layout above. Each step in the chain returns a collection, even if it's an empty collection... acting on an empty collection may not do anything, but it doesn't cause a fatal error!
Any chaining APIs we do add should similarly be built explicitly this way, such as the Message class which we have both PHP and JavaScript versions of.
-- brion