-
Notifications
You must be signed in to change notification settings - Fork 164
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
raft: advance commit index safely #139
base: main
Are you sure you want to change the base?
Conversation
2584c41
to
eb6bfc6
Compare
@ahrtr @nvanbenschoten PTAL |
f727406
to
f64d156
Compare
} | ||
|
||
for i, tt := range tests { | ||
storage := newTestMemoryStorage(withPeers(1, 2)) | ||
storage.Append([]pb.Entry{{Index: 1, Term: 1}, {Index: 2, Term: 2}, {Index: 3, Term: 3}}) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This test was incorrect previously. The third entry term is 3, while the leader Term
is 2. It was incorrect to update the commit index to 3 in this case, because the leader 2 would never commit this entry.
This basically demonstrates that we fixed the unsafety. Now it's impossible for the follower to advance the commit index to 3 in this case.
30f9d89
to
73c85a0
Compare
This change makes the commit index advancement in handleHeartbeat safe. Previously, a follower would attempt to update the commit index to whichever was sent in the MsgHeartbeat message. Out-of-bound indices would crash the node. It is always safe to advance a commit index if the follower's log is "in sync" with the leader, i.e. when its log is guaranteed to be a prefix of the leader's log. This is always true if the term of last entry in the log matches the leader team, otherwise this guarantee is established when the first MsgApp append message from the leader succeeds. At the moment, the leader will never send a commit index that exceeds the follower's log size. However, this may change in future. This change is a defence-in-depth. The newly added raftLog.leaderTerm field will be used for other safety checks in the future, for example to establish that overriding a suffix of entries in raftLog is safe. Signed-off-by: Pavel Kalinnikov <[email protected]>
73c85a0
to
87ac09e
Compare
// does for the leader. For followers and candidates, when we first learn or | ||
// bump to a new term, we don't have a proof that our log is consistent with | ||
// the new term's leader (current or prospective). The new leader may override | ||
// any suffix of the log after the committed index. Only when the first append |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
"the log", "the committed index"
Are you referring to the replicated log here, or this replica's local log? In other words, should these "the"s be replaced by "our"?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Both, but mostly the "replicated". The leader can override only a suffix of the "replicated" log after the "replicated" commit index.
"Our" log is lagging the leader's log, such as our committed index. For "our" log, this implies that the leader can override a suffix after our committed index (but we have no way of checking by how much it can go back).
// log is guaranteed to be a prefix of this term's leader log. | ||
// | ||
// The leaderTerm can be safely updated to `t` if: | ||
// 1. the last entry in the log has term `t`, or, more generally, |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Are there cases where (2) is insufficient and (1) is needed? I'm curious why we can't make this more specific. You say "We use (1) to initialize leaderTerm" below. Is this important on startup?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Only (2) is necessary. (1) is a cheap "stateful" version of (2).
On a server restart, we forget who did the last append. (1) gives the best guess, and allows recovering to up-to-date state for idle raft groups that haven't seen recent appends except the current leader's empty entry. For non-idle groups, or followers whose log is significantly behind the leader's and doesn't end with entries at its term, recovering to raftLog.lastTerm()
gives no value, and is equivalent to setting leaderTerm = 0
. To recover, these will then have to "wait" for the next append message from this leader.
It would be more ideal if this field was stored in some HardState.LeaderTerm
- then we would always recover to up-to-date state. Note that we can't reuse HardState.Term
because HardState.Term
can be > leaderTerm
(for the same reason why r.Term
can be > leaderTerm
).
To bring analogy with Paxos, the local raftLog
is an acceptor, and raftLog.leaderTerm
is the id of the highest accepted proposal. The election term is sort of orthogonal to this - the election term (r.Term / HardState.Term
) can briefly jump in front of the accepted proposal term until there is an accepted proposal at this new term.
If we ever want to bring this to the next level: caveat: #139 (comment)MsgApp
messages should not be rejected based on r.Term / HardState.Term
. For correctness, it is only necessary to reject MsgApp
if the message term is < raftLog.leaderTerm / HardState.LeaderTerm
. I think this would reduce some unnecessary rejects during the leader election flux time.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I should note the invariant here, and maybe check it in a few places or tests:
RawNode.raft.Term >= raftLog.leaderTerm >= raftLog.lastTerm()
(1) initializes leaderTerm
to lastTerm()
, which is safe because raft.Term >= raftLog.lastTerm()
(2) maintains it
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
allows recovering to up-to-date state for idle raft groups that haven't seen recent appends except the current leader's empty entry. For non-idle groups, or followers whose log is significantly behind the leader's and doesn't end with entries at its term, recovering to raftLog.lastTerm() gives no value, and is equivalent to setting leaderTerm = 0. To recover, these will then have to "wait" for the next append message from this leader.
This is what I was hoping to clarify. Initializing to lastTerm
is an opportunistic way to allow a restarted follower in idle raft groups to immediately advance its commit index on startup without needing to first accept a MsgApp.
Is it anything more than that? For an idle raft group, a follower with an up-to-date log but a lagging commit index may restart and never receive any new entries. If we didn't have (1) and we started discarding commit indexes in heartbeats with terms > raftLog.leaderTerm, would the commit index on the follower get stuck?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
That's a good point. I think we indeed need (1). Also, (1) plays nicely with the invariant that I put above.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
If we ever want to bring this to the next level:
MsgApp
messages should not be rejected based onr.Term / HardState.Term
. For correctness, it is only necessary to rejectMsgApp
if the message term is <raftLog.leaderTerm / HardState.LeaderTerm
. I think this would reduce some unnecessary rejects during the leader election flux time.
Does this mean MsgApp from old leader will be accepted by voters that voted to a new leader but yet received any log from the new leader? In that case, there may be committed entry in old leader but not in the new leader.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@joshuazh-x That's a good point. We don't want a quorum of nodes to accept entries unless we are sure these entries are consistent with the new leader's log.
So it's safest to accept only MsgApp.Term >= raft.Term
. But we could sometimes accept MsgApp.Term < raft.Term
if:
- The
MsgApp
contains the(index, term)
entry for which we voted when electing theraft.Term
leader. If that election wins, we know the new leader will append right after this entry. - We would truncate
MsgApp.Entries
at the aforementionedindex
we voted for, and append it. - This guarantees that this append is consistent with the new leader's log (i.e. the new leader would send us the exact same entries).
A vote is a promise to the leader not to accept any entries that are not in the leader's log. If we can deduce that an entry is in the leader's log (before / other than by getting a MsgApp
directly from this leader), we can always safely accept it.
It's unclear if such an optimization would give any value (like reduce replication latency in some cases; probably it does avoid a duplicate MsgApp
from the new leader when the election races with the old leader appends), so I will leave it as an exercise for later :) Looks like a complication.
Filed #150 with a more general technique that will bring more benefits.
@@ -935,6 +935,8 @@ func (r *raft) becomeLeader() { | |||
// so the preceding log append does not count against the uncommitted log | |||
// quota of the new leader. In other words, after the call to appendEntry, | |||
// r.uncommittedSize is still 0. | |||
|
|||
r.raftLog.leaderTerm = r.Term // the leader's log is consistent with itself |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
nit: is is common to manipulate fields in raftLog
directly? Should we tuck this behind a method? Same question below.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Unfortunately, the direct manipulation on raftLog
and Progress
fields is scattered across the raft.go
too. This goes into the TODO above: all the correctness handling should be done at the level below.
In this particular case/PR though, we can move the leaderTerm
update into the r.appendEntry
call a few lines above, which in turn will delegate this to r.raftLog.append
method. This will be more correct, I agree.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Done this (see second commit).
Unfortunately, too many things have incorrect interfaces: the raftLog
append and commit methods don't take the leader term into consideration. This lead to some tests testing scenarios which can never happen.
Maybe it would be good to reverse the order of commits here: first clean up the raftLog
interfaces, and then this PR will be a small change.
raft.go
Outdated
// TODO(pav-kv): move this logic to r.raftLog, which is more appropriate for | ||
// handling safety. The raftLog can use leaderTerm for other safety checks. | ||
// For example, unstable.truncateAndAppend currently may override a suffix of | ||
// the log unconditionally, but it can only be done if m.Term > leaderTerm. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I'm +1 on these additional safety checks / assertions, but they pose potential availability risk if we get them wrong and start rejecting valid state transitions. Should we make a habit of adding logging in the cases where we drop messages, so that any bugs are observable? For example, logging on an else
branch here about how we're ignoring the commit index.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Otherwise, entries at this index may mismatch.
Did you ever see a real issue or can you create a test to reproduce the issue (mismatch)?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@ahrtr Currently, this will not occur because the leader cuts the commit index at the follower's Progress.Match
. This is explained in #138. The solution to #138 is exactly to start sending commit indices that may be out-of-bound (but handling them correctly here on the follower).
I can confirm, however, that in a test environment in CRDB I've seen out-of-bound commit indices in this line, triggering panic in r.raftLog.commitTo
. This happened in a particular kind of tests in which writes on the leader or other followers are sometimes not durable. It's bad to crash this follower only because writes on other participants are not durable - we should handle it more gracefully. Esp. in CRDB case when multiple raft groups are hosted by the same process.
This PR strengthens this code so that it handles out-of-bound indices correctly, and proceeds only if it's safe.
Re @nvanbenschoten comment about logging: I don't know if we should log dropped messages. I think we should only be reporting safety properties violation - that means there is a bug in raft
or some state corruption. For cases (like this) when dropped messages can be a direct result of asynchrony / races / distributed nature of the system - it should be ok to silently disregard them (moreover, raft
is built with the assumption that messages can be dropped arbitrarily).
I don't expect we would be dropping any messages here currently (except in the test environment case that I described above).
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
On a second thought, we might opt to distinguish legit message drops vs safety violations. Safety violations should crash offending nodes, or have other means of appearing on radars (e.g. a raft group can be "bricked", and this can surface in monitoring). See #18.
Legit messages drops, OTOH, don't have to be loud, but we can still log them for informational purposes (and, indeed, in cases of liveness bugs logs can help with debugging).
// log is guaranteed to be a prefix of this term's leader log. | ||
// | ||
// The leaderTerm can be safely updated to `t` if: | ||
// 1. the last entry in the log has term `t`, or, more generally, |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Only (2) is necessary. (1) is a cheap "stateful" version of (2).
On a server restart, we forget who did the last append. (1) gives the best guess, and allows recovering to up-to-date state for idle raft groups that haven't seen recent appends except the current leader's empty entry. For non-idle groups, or followers whose log is significantly behind the leader's and doesn't end with entries at its term, recovering to raftLog.lastTerm()
gives no value, and is equivalent to setting leaderTerm = 0
. To recover, these will then have to "wait" for the next append message from this leader.
It would be more ideal if this field was stored in some HardState.LeaderTerm
- then we would always recover to up-to-date state. Note that we can't reuse HardState.Term
because HardState.Term
can be > leaderTerm
(for the same reason why r.Term
can be > leaderTerm
).
To bring analogy with Paxos, the local raftLog
is an acceptor, and raftLog.leaderTerm
is the id of the highest accepted proposal. The election term is sort of orthogonal to this - the election term (r.Term / HardState.Term
) can briefly jump in front of the accepted proposal term until there is an accepted proposal at this new term.
If we ever want to bring this to the next level: caveat: #139 (comment)MsgApp
messages should not be rejected based on r.Term / HardState.Term
. For correctness, it is only necessary to reject MsgApp
if the message term is < raftLog.leaderTerm / HardState.LeaderTerm
. I think this would reduce some unnecessary rejects during the leader election flux time.
// does for the leader. For followers and candidates, when we first learn or | ||
// bump to a new term, we don't have a proof that our log is consistent with | ||
// the new term's leader (current or prospective). The new leader may override | ||
// any suffix of the log after the committed index. Only when the first append |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Both, but mostly the "replicated". The leader can override only a suffix of the "replicated" log after the "replicated" commit index.
"Our" log is lagging the leader's log, such as our committed index. For "our" log, this implies that the leader can override a suffix after our committed index (but we have no way of checking by how much it can go back).
@@ -935,6 +935,8 @@ func (r *raft) becomeLeader() { | |||
// so the preceding log append does not count against the uncommitted log | |||
// quota of the new leader. In other words, after the call to appendEntry, | |||
// r.uncommittedSize is still 0. | |||
|
|||
r.raftLog.leaderTerm = r.Term // the leader's log is consistent with itself |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Unfortunately, the direct manipulation on raftLog
and Progress
fields is scattered across the raft.go
too. This goes into the TODO above: all the correctness handling should be done at the level below.
In this particular case/PR though, we can move the leaderTerm
update into the r.appendEntry
call a few lines above, which in turn will delegate this to r.raftLog.append
method. This will be more correct, I agree.
raft.go
Outdated
@@ -1735,6 +1737,7 @@ func (r *raft) handleAppendEntries(m pb.Message) { | |||
return | |||
} | |||
if mlastIndex, ok := r.raftLog.maybeAppend(m.Index, m.LogTerm, m.Commit, m.Entries...); ok { | |||
r.raftLog.leaderTerm = m.Term // the log is now consistent with the leader |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@nvanbenschoten This should move to r.raftLog.maybeAppend
, relatedly to your other comment.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Also done in the second commit.
// During normal operation, leaderTerm matches the node term though. During a | ||
// leader change, it briefly lags behind, and matches again when the first | ||
// append message succeeds. | ||
leaderTerm uint64 |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
My immediate feeling is that it's a little weird to add a leaderTerm
into raftLog
, which shouldn't care about the info (leader's Term). It should be part of the raft
instead of raftLog
.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
See comment https://github.com/etcd-io/raft/pull/139/files#r1469818174, specifically the Paxos analogy.
Semantically, raftLog
is layered under raft.go
. The raftLog
is agnostic to election, and essentially only implements the "Paxos acceptor" role. So it makes quite a good sense to me to have a "last accepted term" notion in raftLog
. Moving this logic up in raft.go
would be a layering violation of sorts.
which shouldn't care about the info (leader's Term)
raftLog
should care about the leader term with which it is consistent (i.e. what I called the leaderTerm
here). It should not care about the RawNode.Term
/ HardState.Term
, because the latter can be bumped arbitrarily when a new election happens, and this bump is not coordinated with raftLog
.
There is an invariant at raft.go
level that is relevant here: RawNode.Term >= raftLog.leaderTerm
. We use RawNode.Term
instead of raftLog.leaderTerm
in a few places (notably, to reject append messages), and have only been safe because of this implicit invariant. I think we should make it more explicit.
This change improves safety of the append operations on raftLog. This helped fixing some tests which made incorrect assumptions about the log. Signed-off-by: Pavel Kalinnikov <[email protected]>
This PR is currently blocked on a clean-up started in #145. Once the |
hi, #145 was merged, any update to this PR? |
This change makes the commit index advancement in
handleHeartbeat
safe. Previously, a follower would attempt to update the commit index to whichever was sent in theMsgHeartbeat
message. Out-of-bound indices would crash the node.It is always safe to advance a commit index if the follower's log is "in sync" with the leader, i.e. when its log is guaranteed to be a prefix of the leader's log. This is always true if the term of last entry in the log matches the leader team, otherwise this guarantee is established when the first
MsgApp
append message from the leader succeeds.At the moment, the leader will never send a commit index that exceeds the follower's log size. However, this may change in future. This change is a defence-in-depth.
The newly added
raftLog.leaderTerm
field will be used for other safety checks in the future, for example to establish that overriding a suffix of entries inraftLog
is safe.Informs #138
Related to #144