Comment Re:You would think. (Score 1) 348
In fact, there are quite a few people out there using Anycast for TCP-sessions. It's really a matter on what timescale you're looking at. The networking guys see TCP as something to use for long-living connections - e.g. a BGP session running for days, weeks or even months. A flapping route in this setup will result in a broken session. But: what does this really mean to you? If your CDN distributes downloads which are "done" within a few minutes, such a rarely flapping route will result in a few broken sessions once a day out of millions of downloads successfully served. Compared to issues like non-working DNS, overloaded servers and filled lines, that's nothing and can actually enhance the overall CDN service.
A nice paper to read is this one from Matt Levine. He's working for a CDN provider using TCP-Anycast for years now and sums up the most important issues on TCP-Anycast.
Basically the most important one is that your anycasted servers really have to be spread far enough so that flapping routes at some peering point won't matter. As a rule of thumb, put one CDN loadbalancer on the US east coast, one to the US west coast, another one to western europe, one to Australia and one in Hong Kong. If you'd like to put multiple CDN loadbalancers to one continent, leave space between them, e.g. one box for each country/state.