Sloppy Management of Structured P2P Services

While most structured peer-to-peer services are conceptually very simple, their implementation is not. Even though similar observations can be made about any type of software, it can be shown that large parts of implementation complexity derives from the way management tasks such as error condition handling are dealt with. The traditional way to manage distributed systems software is to list all possible error conditions such as churn and partial node or network failures, and come up with repair algorithms that take care of maintaining the desired structure despite adversary conditions. However, implementing repair algorithms is cumbersome, and any error can potentially lead to complex liveness bugs [6], not to speak about unwanted cross-interactions of repair schemes. We observe that most — if not all — structured peer-topeer systems deploy strategies to sustain temporary structure inconsistencies that derive from error conditions. In this context, our position is that explicit repair algorithms can and should be avoided in the implementation of structured peer-to-peer services. Instead, we should use continuous lazy background algorithms to handle non-functional management tasks such as routing table maintenance, while relying on the original structured algorithms for the functional tasks such as routing messages through a DHT. We call this form of probabilistic overlay maintenance “sloppy management”. We already demonstrated in two simple situations that such a dual approach based on a structured deterministic functional plane and an unstructured probabilistic nonfunctional can actually work [2, 8]. As an example we briefly discuss in Section 2 how gossip-based protocols can be used to maintain the routing tables of a Chord DHT. However, if we are to take the sloppy management approach seriously, many other non-trivial management tasks should be handled by the probabilistic plane. We discuss them in Section 3, and finally conclude in Section 4.