Colin's Journal: A place for thoughts about politics, software, and daily life.
Tim Bray’s assertion that IT professionals are implementing Enterprise Systems wrongly is a fairly seductive argument at first. The one sentence summary of his argument is:
The community of developers whose work you see on the Web, who probably don’t know what ADO or UML or JPA even stand for, deploy better systems at less cost in less time at lower risk than we see in the Enterprise.
He goes on to list companies such as Facebook, Google, Twitter, etc as examples of doing it the right way.
I don’t agree with his analysis. As commenters to the post were quick to point out, comparing a list of failed enterprise projects (including ones such as the NHS National Programme for IT that stretches beyond software solutions) to those few internet firms that have thrived is not instructive. For every Twitter (rumoured to be barely making a profit) there are hundreds of failed internet start-ups. IT department do not have the luxury to fail projects as frequently as start-ups fold. While governance of enterprise projects is often slow and cumbersome, it’s designed to improve the odds of success above this base level.
The assertion that UML and other enterprise IT tools and techniques are not used by internet firms is unsubstantiated. A quick search reveals that Google, Amazon and eBay have the enterprise role of “architect” among their job descriptions. I suspect all of the mature internet firms have similar roles, and use many of the same enterprise techniques as more traditional institutions.
As Erik Engbrecht points out, comparing the legacy-system ridden architecture of a typical enterprise to the “internet app is our company” environment of Twitter is also not useful. In many cases the quick dirty approach used to develop internet applications is the technique that creates the spaghetti architecture found within many enterprises.
There are undoubtedly lessons from internet companies that could be learnt by enterprise system professionals. Many of the technologies used in the enterprise are overly complex in comparison to internet based solutions. For example the poor use of SOA while integrating systems brings additional cost with limited benefits. However such learning is not a revolution, and will not reduce the complexity of implementing enterprise systems substantially.
I’ve had a Google Wave account for a short while now, but (like many users I’m sure) the slow invite system resulted in me having very few people to test it with. Last night I had my first real-time three way conversation, and it really highlighted the short-comings of the current user experience.
The most irritating issue with Wave’s current UI is that it’s slow as molasses when lots of things are going on. Scrolling fails to work properly, clicks don’t register when they should, and you get false clicks as the screen re-draws under your cursor. This is in an up-to-date version of Firefox (3.5.5), on a fast laptop and with a fast broadband connection. I think a native client rather than a web app would be a significant improvement in this respect, though I doubt we’ll see one from Google.
The next challenge is fundamental to the way that Wave works. Because any user can start a new response to any part of the wave, or edit any existing part of the wave, it quickly becomes difficult to track what is happening. I admit that we were messing around, but the volume of text we had was relatively low, and it still became tedious tracking down which edits you’d read and which were new.
So my first impressions were not overly positive. Having said that there are a number of features that Wave has that really do set it apart from email & IM:
While I really like the principal behind Google Wave I’m yet to be convinced that the current implementation will be good enough to replace email and instance messaging.
Python is my language of choice, something that I’ve used for a number of websites, applications and tools. It’s speed of execution has never been an issue for me, even when using it to wrangle large XML files into something more useful.
Despite this, I’m happy to see the existence of the Unladen Swallow project. Whether this tiny team of two Google engineers will really be able to make Python 5 times faster remains to be seen, but the attempt will surely be worthwhile. Already improvements have been made, and brought back into the mainline of Python, and the approach being taken seems solid.
The project’s plan is both practical and well thought out. There have been several attempts to re-implement or speed up Python before (PyPy, Pysco, IronPython, Jython), none of which have come close to challenging CPython in terms of adoption. By accepting that CPython is the common implementation of choice, and concentrating on making it better, the project’s benefits should be widely felt. Similarly, planning to gain the performance improvements through application of the relevant academic research, rather than trying to discover something new, should ensure a measure of success in a short period of time.
I have finished porting SimpleTAL to Python 3. Release 5.0 of SimpleTAL is for Python 3.1 and provides similar functionality as SimpleTAL 4.2 does for Python 2.5. The differences between using 4.2 and 5.0 are documented on the SimpleTAL notes page.
At first the porting process was fairly easy. I started by getting all test cases to run cleanly under Python 2.6 with the -3 flag, and then ran 2to3 to convert the basic syntax. The next step was to run the test cases under Python 3 to highlight issues that required manual changes. Sgmllib has been removed from the standard library, so I had to remove HTMLStructureCleaner from simpleTALUtils (it was unused within the library itself). The Iterator protocol change from “next” to “__next__” meant my iterator detecting code had to be updated.
The changes to character set handling in Python 3 introduced slightly more complex changes for the template handling. In Python 2.x the SimpleTAL library would handle all encoding / decoding itself, but in Python 3 this is not always required as there is now a clean separation between bytes and strings.
One issue that I hit when porting to Python 3 was the use of regular expressions. In order for SimpleTAL to pass through singleton XML elements from the template (i.e. <tag /> rather than <tag></tag>) it needs to carry out a regex check against the raw XML that the SAX library provides. This is done by retrieving the xml.sax.handler.property_xml_string property, which is documented as returning a string. In practise however the Python 3 SAX implementation returns bytes, which at first I assumed the regex library would not work with. A little bit of research later, and I learned that the regex library can work on bytes as well.
One final surprise was the huge performance gain moving from Python 2.6 to 3.1. The SimpleTAL performance tests show a minimum speed increase of 60% (on the METAL test), with some tests clocking in 90% increases. Both HTML and XML basic template expansions are now hitting over 1600 pages/sec on a single 1.7GHz CPU.
Email: colin at owlfish.com