Using MD5 for this is undeniably a bad habit, of course. Just use the output of an SHA-2 function, even if you have to truncate it to 128 bits. Anything to get people to stop using MD5.
Why? Because a bunch of academics managed to create a collision after 2 weeks of trying on a beefed up machine? Really?
MD5 is much faster than SHA-2. As long as you're aware of the security implications, there's no need to stop using it.
WTF?! Are you out of your mind? MD5 was exploited to falsify a real-world RapidSSL CA certificate... in 2008. MD5's weaknesses haven't been "academic" in years.
People reach for what they know, and most do not have the capacity to determine the security of cryptographic algorithms in any particular context. Getting them to stop using it at all is the surest way to get them to stop using it where it's critical.
> MD5 is much faster than SHA-2.
Largely irrelevant in today's world. Single-threaded, my two year old laptop runs over 125MB/sec through SHA256, and around 100MB/sec through SHA512. SHA-2 is not a major bottleneck.
Why not just use a random 64 or 128 bit number (or UUID)? This would be faster and would not require that the input (that you are hashing) already be unique.
The only legitimate use of MD5 I can think of is to verify legacy MD5 checksums.