raft issue: break linear consistency because of two leaders and raft lease read.

[tangyuanzhang]

leader_ and isBlindFollower_ act the same， I think isBlindFollower_ == True eq leader_ == HostAddr("", 0)

We should increase the limit when voting. The follwer host leader_ != HostAddr("", 0) and leader lease valid recorded by the current host should refuse to vote to prevent split-brain

[tangyuanzhang]

isBlindFollower_ maybe cause split brain， A（leader）B（follower） C(followe)， restart B and C， B start election
，C restart ok，dont start election，C vote for B。 B send request to A is lost， B is leader and A is leader

[liwenhui-soul]

why does B start election?

[tangyuanzhang]

start to election has gap

[critical27]

isBlindFollower_ maybe cause split brain， A（leader）B（follower） C(followe)， restart B and C， B start election ，C restart ok，dont start election，C vote for B。 B send request to A is lost， B is leader and A is leader

Both leader is possible, but they are not same term. This is a very common case for raft.

isBlindFollower_ is only used for speed up first election.

[critical27]

I don't think this is a issue, reopen it if necessary

[tangyuanzhang]

But if there is data written to B now, it can be submitted successfully, but the read data is routed to A, then the data cannot be read or old. Does this cause inconsistency?

[critical27]

I got your point.
It won't cause inconsistency. Because at least B or C has same log as A, let's say it is B, only B could be elected as leader at higher term.

Read could be routed to A, but write to A won't be success anymore. So again, no inconsistency happened.

[qiwei9743]

It indeed break the linear consistency. nebula implement pure lease read but didn't implement checkquorum and read index as ETCD does. The old leader may read the stale data from old leader after new write over B and C(in shixiang's case). Although it's a small gap, nebula can't declare it implement a strong consistency.

[critical27]

It indeed break the linear consistency. nebula implement pure lease read but didn't implement checkquorum and read index as ETCD does. The old leader may read the stale data from old leader after new write over B and C(in shixiang's case). Although it's a small gap, nebula can't declare it implement a strong consistency.

Perhaps you need to describe in more details how inconsistency happen. Liked I said in #5352 (comment), two leader doesn't mean split brain. Only when two leader of same term is. New leader of B or C can only trigger election after an election timeout passed. If B or C triggers election, which means, A's lease is valid as well. It won't serve read any more.

[critical27]

Although I agree that blind follower that may break the precondition that election could only after an election timeout passed.

I think we could remove the blind follower? WDYT