mirror of
https://github.com/scylladb/scylladb.git
synced 2026-05-01 13:45:53 +00:00
bf823e34a4
The Raft PhD presents the following scenario.
When we remove a server from the cluster configuration, it does not
receive the configuration entry which removes it (because the leader
appending this entry uses that entry's configuration to decide to which
servers to send the entry to, and the entry does not contain the removed
server). Therefore the server keeps believing it is a member but does
not receive heartbeats from leaders in the new configuration. Therefore
it will keep becoming a candidate, causing existing leaders to step
down, harming availability. With many such candidates the cluster may
even stop being able to proceed at all. We call such servers
"disruptive".
More concretely, consider the following example, adapted from the PhD for
joint configuration changes (the original PhD considered a different
algorithm which can only add/remove one server at once):
Let C_old = {A, B, C, D}, C_new = {B, C, D}, and C_joint be the joint
configuration (C_old, C_new). D is the leader. D managed to append
C_joint to every server and commit it. D appends C_new. At this point, D
stops sending heartbeats to A because C_new does not contain A, but A's
last entry is still C_joint, so it still has the ability to become a
candidate. A can now become a candidate and cause D, or any other leader
in C_new, to step down. Even if D manages to commit C_new, A can keep
disrupting the cluster until it is shut down.
Prevoting changes the situation, which the authors admit. The "even if"
above no longer applies: if D manages to commit C_new, or just append it
to a majority of C_new, then A won't be able to succeed in the prevote
phase because a majority of servers in C_new has a longer log than A
(and A must obtain a prevote from a majority of servers in C_new because
A is in C_joint which contains C_new). But the authors continue to argue
that disruptions can still occur during the small period where C_new is
only appended on D but not yet on a majority of C_new. As they say:
"we also did not want to assume that a leader will reliably replicate
entries fast enough to move past the scenario (...) quickly; that might
have worked in practice, but it depends on stronger assumptions that we
prefer to avoid about the performance (...) of replicating log entries".
One could probably try debunking this by saying that if entries take
longer to replicate than the election timeout we're in much bigger
trouble, but nevermind.
In any case, the authors propose a solution which we call "sticky
leadership". A server will not grant a vote to a candidate if it has
recently received a heartbeat from the currently known leader, even if
the candidate's term is higher. In the above example, servers in C_new
would not grant votes to A as long as D keeps sending them heartbeats,
thus A is no longer disruptive.
In our case the situation is a bit
different: in original Raft, "heartbeats" have a very specific meaning
- they are append_entries requests (possibly empty) sent by leaders.
Thus if a node stops being a leader it stops sending heartbeats;
similarly, if a node leaves the configuration, it stops receiving
heartbeats from others still in the configuration. We instead use a
"shared failure detector" interface, where nodes may still consider
other nodes alive regardless of their configuration/leadership
situation, as part of the general "MultiRaft" framework.
This pretty much invalidates the original argument, as seen on
the above example: A will still consider D alive, thus it won't become
a candidate.
Shared failure detector combined with sticky leadership actually makes
the situation worse - it may cause cluster unavailability in certain
scenarios (fortunately not a permanent one, it can be solved with server
restarts, for example). Randomized nemesis testing with reconfigurations
found the following scenario:
Let C1 = {A, B, C}, C2 = {A}, C3 = {B, C}. We start from configuration
C1, B is the leader. B commits joint (C1, C2), then new C2
configuration. Note that C does not learn about the last entry
(since it's not part of C2) but it keeps believing that B is alive,
so it keeps believing that B is the leader.
We then partition {A} from {B, C}. A appends (C2, C3) joint
configuration to its log. It's not able to append it to B or C due to
the partition. The partition holds long enough for A to revert to
candidate state (or we may restart A at this point). Eventually the
partition resolves. The only node which can become a candidate now is A:
C does not become a candidate because it keeps believeing that B is the
leader, and B does not become a candidate because it saw the C2
non-joint entry being committed. However, A won't become a leader
because C won't grant it a vote due to the sticky leadership rule.
The cluster will remain unavailable until e.g. C is restarted.
Note that this scenario requires allowing configuration changes which
remove and then readd the same servers to the configuration. One may
wonder if such reconfigurations should be allowed, but there doesn't
seem to be any example of them breaking safety of Raft (and the PhD
doesn't seem to mention them at all; perhaps it implicitly accepts
them). It is unknown whether a similar scenario may be produced without
such reconfigurations.
In any case, disabling sticky leadership resolves the problem, and it is
the last currently known availability problem found in randomized
nemesis testing. There is no reason to keep this extension, both because
the original Raft authors' argument does not apply for shared failure
detector, and because one may even argue with the authors in vanilla
Raft given that prevoting is enabled (see end of third paragraph of this
commit message).
Message-Id: <20210921153741.65084-1-kbraun@scylladb.com>