Author of pycapnp here. For those too lazy to click through, pycapnp is tremendously slower than C++/Rust. The equivalent benchmarks take about 20-50x as long.
What's more interesting to me is comparing it to other Python serialization libraries. I've been meaning to update the benchmarks with some popular ones (msgpack, thrift, and the built-in json module are at the top of my list). If there's some interest, I'll go ahead and post it to HN when I'm done.
Does anyone know if Amazon ec2 instances would actually be vulnerable to this attack? For instances with dedicated cpu being promised (i.e. everything except micro), it seems reasonable that they'd be running Xen with processor affinity turned on, thereby preventing any attack exploiting L1 cache.