The formula given at the end of the article implies that to have a 5% chance of a collision you would need 3.87e23 objects. (SHA-1 is 160 bits.) Even storing 1 trillion objects gives a collision probability of 3e-19.

So this is unlikely to be a concern, ever.

Hmmm, I wonder how many objects are currently stored in S3? And what rate that's rowing at? Or how many documents Google has stored in bigtable?

But yeah, 160 bits is probably "enough" :-)

Doesn't even need to be in filenames, but the content of files aswell.

I believe bigiain is referring to the fact that git objects are stored by their SHA-1 digest (so a hash-collision will cause a collision in these filenames) not the filenames of the contents. (The objects are stored under the .git/objects directory, in the format xx/y...y, where the directory xx is the first two digits of the digest, and the 38 y's are the rest of the digest.)