The report has stirred up a lot of commentary in the media, mostly partisan and focused on [...]]]> Today the Siena Research Institute (SRI) published a list of the best and worst U.S. Presidents, based on a survey of “238 presidential scholars, historians and political scientists that responded via mail or web to an invitation to participate.”

The report has stirred up a lot of commentary in the media, mostly partisan and focused on whether pundits agree with the rankings of recent Presidents.

What’s most interesting to me, though, is the detailed, dimensionalized data that SRI released alongside the report. The survey respondents were asked to rank each President in 20 different categories.

With this categorized data available, I wondered: What specific leadership qualities mattered most in these experts’ overall perception of a President?

To find out, I loaded the data into a spreadsheet and computed the correlation (R-squared for a linear, best-fit trendline) between each individual category and the overall Presidential rankings.

My own interpretation of what these correlations mean:

• First of all, when the survey respondents were asked to provide a single, overall ranking of each President (the “Present Overall View” category), the results were very highly correlated with the derived ranking computed from all the other categories (“Overall Rank”). This indicates to me that each respondent was self-consistent and didn’t, for example, give poor per-category rankings to some President whom they named best overall.
• Background (Family, Education, Experience) mattered relatively little. To me, that’s refreshing: in a true land of opportunity, someone with potential should be able to succeed regardless of background.
• Some personality characteristics widely regarded as virtues–Integrity, Intelligence, and Ability to Compromise–had surprisingly low correlations with Overall Rank. On the whole, it seems, results mattered more than style.
• Domestic Accomplishments were more strongly correlated with Overall Rank than Foreign Policy Accomplishments.
• Luck had one of the lower correlations, suggesting that the particular set of circumstances that fate presented to each President was less important than that President’s handling of the crises and opportunities.

Based on SRI’s survey data, the archetype of a historically well-regarded U.S. President seems to be a skilled organizational manager and good communicator who delivers results on domestic issues. I see parallels to how corporate CEOs are judged, too: imagination and leadership capabilities are valued highly, and results (Domestic Accomplishments in this survey, quarterly earnings in the corporate world) are paramount.

I’d love to hear some feedback from folks in other countries: do people judge leaders by the same criteria worldwide, or does it vary from place to place?

Other professions get to use a variety of big words. Physicians, for [...]]]> People from Computer Science backgrounds often use the word “orthogonal.” I think that’s partly because it concisely describes an often-important concept, but mostly because it’s one of the few impressive words in our vernacular. (“Idempotent” is my other favorite CS term.)

Other professions get to use a variety of big words. Physicians, for example, have a rich, domain-specific vocabulary. So do economists. Lawyers have perhaps the best lexicon; not only do they get to use big, fancy words, but half of those words are in Latin.

Thus it was a pleasant surprise to read that a lawyer had adopted our humble “orthogonal” and earned kudos from the Justices of the United States Supreme Court for it.

“Slow”? What does that mean?

It means that when I start up Firefox, even if it’s already in the filesystem cache, the app’s Mac dock icon completes two full bounces before the browser window appears. Safari’s doc icon completes approximately [...]]]> I’ve found myself using Safari more and more, instead of Firefox, because Firefox is slow.

“Slow”? What does that mean?

It means that when I start up Firefox, even if it’s already in the filesystem cache, the app’s Mac dock icon completes two full bounces before the browser window appears. Safari’s doc icon completes approximately half of one bounce before its window appears. The just-released Chrome for Mac also displays its window in half a bounce.

Each icon bounce takes approximately one second. Thus I’ve developed a preference not to use Firefox–an otherwise excellent product–because it takes 1.5 seconds longer to start than the competition.

I’ll typically keep my web browser running for hours or even days after starting it up, so Firefox’s extra startup time is immaterial when amortized over the time between restarts. That’s my quantitative assessment, anyway. In practice, though, I tend to click the Safari icon because, qualitatively, it’s the fast one.

In software, speed matters.

There was one technical challenge: the project was using nginx as a webserver, and none of [...]]]> I recently helped an old friend with some prototyping and strategy for a stealth-mode project he’s doing. He wanted to use MobWrite, a framework that allows multiple users to do real-time, collaborative editing of content in a web page.

There was one technical challenge: the project was using nginx as a webserver, and none of the existing web gateway implementations for MobWrite were well-suited for use with nginx. MobWrite has a back-end server that speaks a custom protocol, and it needs a bit of additional software to translate HTTP requests from web browsers into the server’s protocol. The MobWrite distribution contains PHP, JSP, and Python pages that do this translation.

It is possible to add support for any and all of those languages to nginx, but doing so adds substantial complexity. Specifically, the event-driven design of nginx works well only with plugins that do their work in small, event-driven steps and do not block for I/O or lengthy computation. PHP, mod_python, and JSP, in contrast, all are based on a programming model where the code that handles the request expects to run synchronously until completion. Running such code inside the single-threaded, asynchronous nginx could allow a long-running PHP/Python/JSP request to block all other requests. One common solution is to use an nginx gateway plugin like mod_scgi that dispatches requests for these synchronous languages to a separate pool of processes. This enables a single nginx server to handle both synchronous PHP/Python/JSP requests and asynchronous static-content or HTTP-proxy requests without either type of request starving out the other. However, adding a separate process pool for the synchronous requests means adding configuration and tuning complexity, thus diminishing one of the inherent advantages of using nginx.

Determined to have my cake and eat it too–i.e., to support MobWrite and other types of content in a single nginx server while still enjoying nginx’s single-process simplicity and efficiency–I made a new nginx plugin called nginx-mobwrite. This plugin translates incoming HTTP requests from MobWrite clients into the MobWrite internal format and proxies them asynchronously to the MobWrite server. The nginx architecture includes an “upstream” framework that manages requests to back end servers as part of nginx’s event loop. The nginx-mobwrite uses this framework to dispatch translated requests to the MobWrite server and send the responses back to clients in an asynchronous manner, without blocking any other requests the nginx server might be handling at the same time.

The nginx-mobwrite plugin is available as open source, under v2.0 of the Apache License, on GitHub.

I [...]]]> I decided recently to try installing Windows 7 on my iMac, in a Boot Camp dual-boot configuration alongside OS X. I ultimately got it working, but I ran into one perplexing problem along the way. Here’s a summary of the symptoms and the solution, in case the information is helpful to other people.

I used the installation process described here. After I clicked the “Start Installation” button in step 8, the computer rebooted and greeted me with a blank black screen with a blinking white underline cursor in the upper left.

A bit of web searching led me to a solution that worked: I had to reset the computer’s PRAM. The PRAM apparently is an area of nonvolatile memory containing information about USB devices such as keyboards and mice. The full process I had to use was:

• Power off the computer.
• Unplug the keyboard and plug it into a different USB port.
• Turn on the computer.
• Immediately press the Command, Option, P, and R keys and hold them down until the system restarts.

I grabbed the .tar.gz distribution of the latest MySQL 5.1 release and installed [...]]]> I wanted to try out Ruby on Rails with MySQL on OS X 10.6. I ran into a few common problems along the way, so I am documenting the process here in case it might be helpful to others.

Step 1: Install MySQL

I grabbed the .tar.gz distribution of the latest MySQL 5.1 release and installed it in a nonstandard location on my computer. “Nonstandard” in this case means “somewhere other than /usr/local/mysql.” This detail would be important in a later step.

Step 2: Install the Ruby mysql gem

I’d installed a 64-bit version of MySQL, so the installation process for the mysql gem was:

sudo env ARCHFLAGS="-arch x86_64" gem install -t mysql \
  -- --with-mysql-config=/path/to/my/mysql/installation/bin/mysql_config

For a 32-bit MySQL installation, replace the “x86_64″ with “x86″ in the ARCHFLAGS.

Step 3: Set the DYLD_LIBRARY_PATH

I omitted this step initially, and when I tried to run “rake db:migrate” for my Rails application, I got an error message:

uninitialized constant MysqlCompat::MysqlRes

In the environment of each user account where you plan to run Ruby apps that depend on MySQL, do this:

export DYLD_LIBRARY_PATH=$DYLD_LIBRARY_PATH:/path/to/my/mysql/installation/lib

Note that most of the command-line tools supplied with binary builds of MySQL appear to be statically linked with libmysqlclient. This means that these tools may work flawlessly even if you’ve neglected to set the DYLD_LIBRARY_PATH. The Ruby mysql gem, however, is dynamically linked to libmysqlclient, so it is essential that the MySQL lib directory be included in your DYLD_LIBRARY_PATH if MySQL is installed in a nonstandard location.

For most other Unix systems, the environment variable to set is LD_LIBRARY_PATH rather than DYLD_LIBRARY_PATH.

Twitter Data Analysis: An Investor’s Perspective

As the author notes, it was possible to deduce so many key metrics because “Twitter uses auto-incrementing ID numbers (1,2,3,4…) for both users and tweets.”

Auto-incremented ID numbers are a common web development practice. When you run an [...]]]> This post on TechCrunch contains quite a nice analysis of Twitter’s user behavior:

Twitter Data Analysis: An Investor’s Perspective

As the author notes, it was possible to deduce so many key metrics because “Twitter uses auto-incrementing ID numbers (1,2,3,4…) for both users and tweets.”

Auto-incremented ID numbers are a common web development practice. When you run an online business, you and/or your users create data–hopefully lots of data. Each record of data needs a unique identifier that you can use to go find that bit of data for later use. Automatically-incrementing integer IDs are a common implementation choice because they have some desirable properties:

• Compactness – Shorter URLs are generally a good thing. In addition, IDs that can fit into a 32- or 64-bit integer representation allow for efficient storage in a database and efficient manipulation within most programming languages.
• Uniqueness – It is trivially easy to generate unique IDs with an auto-incrementing approach: keep track of the last ID you generated, and add one.
• Efficiency – Generating the next sequential ID value doesn’t take very long. In more precise terms, it takes O(1) time; the speed with which you can generate the next integer is a constant, and this operation will not get slower as your data size grows.
• Convenience – Most databases have a built-in feature that can auto-generate auto-incrementing IDs.

Despite these virtues, auto-incrementing IDs can reveal information about the rate at which you are accumulating customers or content. If your user IDs are sequentially generated, for example, a competitor need only sign up for a new account on your website once a day to determine how fast your user base is growing. You may not mind that risk; the engineering team at Twitter probably discussed this issue when designing their data model and consciously decided not to obscure their ID values.

In cases where you do want to avoid using auto-incrementing ID values, some alternative ID generation strategies are:

• Random – Generate a random number every time you need a new ID. Guaranteeing uniqueness requires a bit more work with this approach, though.
• Random increment – When generating a new ID, instead of adding one to the previously generated ID, add a random number. This approach will exhaust the range of your chosen integer representation more quickly than the traditional auto-increment, though, so it is perhaps best used with a 64-bit representation.
• UUID – Use UUIDs instead of integers as IDs. UUIDs have the properties of efficiency and uniqueness (even with random UUID generation, there are enough bits to make duplicates very rare) but are less compact than a typical auto-incrementing numeric ID. A UUID occupies 36 bytes in human-readable form or 16 bytes (128 bits) in binary form.