Otherwise the primary renews the lease and continues

Info iconThis preview shows page 1. Sign up to view the full content.

View Full Document Right Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: node needs to be added to the set performance optimization: we don’t need to save prepare/accept state on disk if we never allow a node to recover, but only allow replacement conceptually simple: just put an event in the sequence to change the set of voters; applies to all following events this “view change” event must be chosen by a majority of the PREVIOUS set of acceptors. This event is NOT commutative with respect to other instances, so you can’t run parallel instances Thus, if two of k nodes fail, you can potentially vote among the remainder that you will no longer accept proposals from the missing nodes. Then for new votes, you will only need to get a majority of the remainder. The failed nodes might not really be failed of course – but they won’t be able to form a majority (of the original set of nodes), since that would overlap at least with one of the nodes that approved the view change. 7) intermittent network outages Sometimes a particular connection may go up and down quickly – so that some nodes may think a node is down while it isn’t. If this happens to a regular node, it falls behind and has to catch up. (It can keep accepting proposals, but if enough nodes fail to keep up, you could have to redo a bunch of work to re- learn chosen proposals. If this happens to the leader, then a new leader will be chosen, but what happens when the old leader comes back online. Most systems try to avoid leader swaps, e.g.,...
View Full Document

Ask a homework question - tutors are online