Good talk, but disappointing that he doesn't mention the issue of node failures: With W=1 and a single node failure at the wrong time, you don't have eventual consistency any more... you have data loss.
Great presentation Mark! This is what I LOVE about Basho - use science and metrics to show how their Riak compares to other NoSQL stores rather than marketing.<p>TL;DR version: In all databases, you have to choose which 2 parts of CAP you want: Consistency, Availability, Partition tolerance. If you choose Availability- just HOW Consistent is my data? If I want my data to be more consistent, then how does this affect my availability? This gives an actual formula to calculate the best design that meets your applications goals.
I can't see the talk at the moment but here is my 5 cents on the subject:<p>The main problem of eventual consistency is not how often that happens, it is: What damage will it do WHEN it happens?<p>Imagine you're a bank, you handle big clients you lose track of a write of 50+ million dollars. Where did the money go ? How to differentiate that from a fraud attempt ?<p>If you have customers how will you tell them that you just lose, probabilistically speaking "one in 10 million packages ?"<p>But that's a very interesting question that also has philosophical repercussions: How come that we are in a society that did not build system that accept a certain degree of failure?