3 ms·
For me the most interesting part was: > ...we can prove that certain levels of consistency cannot be achieved without making operation latency proportional to
by jamii 11y ago
For me the most interesting part was:
> ...we can prove that certain levels of consistency cannot be achieved without making operation latency proportional to network delay.
'Consistency requires waiting' is a pretty well known rule of thumb for distributed systems but this is the first quantitative proof that I've seen. It's really useful see exactly what kinds of consistency impose latency and how that latency varies with respect to network conditions.
- seiji 11y agoWell, guaranteed consistency requires coordination and if your coordination mechanism is over a network, then the speed of your writes will be multiple factors of your network latency. "network" doesn't necessarily have to mean high latency ethernet though. You can have a network running on top of an embedded backplane in a blade system. There are ways to minimize latency, but latency is as latency does.
- brudgers 11y agoThere's no way for node A to distinguish between latency and partitioning in regards to node B until node A receives a response from node B. Partitioning is just another name for timing out.