Get the latest tech news

Constraining writers in distributed systems

This is a note to describe a pattern for avoiding failures in distributed systems, variants of which has been written about multiple ways with multiple names (“copysets”, “read/write quorum systems”). Say we have a distributed storage system with N nodes that can store data, but some subset of them might fail (though the failure probability for any individual node is low).

This can be used for consensus systems like Paxos (as described below), but let’s use a simpler example: We want to store a bunch of files in a content-addressed database keyed by a cryptographic hash of their contents. (Side note: You can also describe Fast Flexible Paxos in these terms, which I think is much more natural than the read/write quorum system description. What’s the first mention of these ideas – either quorum systems that require overlap between two kinds of participants, or constraining write placement to reduce failure probability?

Get the Android app