That guy was reporting a 2% error rate from one S3 server, but there has been additional discussion on that thread which suggests that error rate is an aberration. Personally, I have no hard data, just the anecdote that when I started developing with S3 I wasn't using retries and that became a problem after approximately a few thousand queries. Unfortunately I've never seen Amazon say what error rate to expect, just that you should be sure to retry on error. I think I'll start logging how often I need to retry so I can get some hard data.

Keep in mind that when you're hosting a site on S3, you have to multiply the error rate by the number of resources you're serving from S3 to get the probability an individual user will have problems. www.ismytwitterpasswordsecure.com serves 4 resources from S3. Assuming a 0.1% error rate, 400 out of the 100,000 users would have had problems.

As another commenter has suggested, you can serve from Cloudfront instead of directly from S3. That will significantly reduce the number of hits which are made to S3.