Apps typically maintain the raft configuration as part of the state
machine. As a result, they want to be able to reject configuration change
entries at apply time based on the state on which the entry is supposed
to be applied. When this happens, the app should not call
ApplyConfChange, but the comments did not make this clear.
As a result, it was tempting to pass an empty pb.ConfChange or it's V2
version instead of not calling ApplyConfChange.
However, an empty V1 or V2 proto aren't noops when the configuration is
joint: an empty V1 change is treated internally as a single
configuration change for NodeID zero and will cause a panic when applied
in a joint state. An empty V2 proto is treated as a signal to leave a
joint state, which means that the app's config and raft's would diverge.
The comments updated in this commit now ask users to not call
ApplyConfState when they reject a conf change. Apps that never use joint
consensus can keep their old behavior since the distinction only matters
when in a joint state, but we don't want to encourage that.
I changed `(*RawNode).Ready`'s behavior in #10892 in a problematic way.
Previously, `Ready()` would create and immediately "accept" a Ready
(i.e. commit the app to actually handling it). In #10892, Ready() became
a pure read-only operation and the "accepting" was moved to
`Advance(rd)`. As a result it was illegal to use the RawNode in certain
ways while the Ready was being handled. Failure to do so would result in
dropped messages (and perhaps worse). For example, with the following
operations
1. `rd := rawNode.Ready()`
2. `rawNode.Step(someMsg)`
3. `rawNode.Advance(rd)`
`someMsg` would be dropped, because `Advance()` would clear out the
outgoing messages thinking that they had all been handled by the client.
I mistakenly assumed that this restriction had existed prior, but this
is incorrect.
I noticed this while trying to pick up the above PR in CockroachDB,
where it caused unit test failures, precisely due to the above example.
This PR reestablishes the previous behavior (result of `Ready()` must
be handled by the app) and adds a regression test.
While I was there, I carried out a few small clarifying refactors.
This change introduces joint quorums by changing the Node and RawNode
API to accept pb.ConfChangeV2 (on top of pb.ConfChange).
pb.ConfChange continues to work as today: it allows carrying out a
single configuration change. A pb.ConfChange proposal gets added to
the Raft log as such and is thus also observed by the app during Ready
handling, and fed back to ApplyConfChange.
ConfChangeV2 allows joint configuration changes but will continue to
carry out configuration changes in "one phase" (i.e. without ever
entering a joint config) when this is possible.
We are worried about breaking backwards compatibility for any
application out there that may have relied on the old behavior. Their
RawNode invocation would have been broken by the removal of the peers
argument so it would not have changed silently; an associated comment
tells callers how to fix it.
This is the first (maybe not last) step in cleaning up the bootstrap
code around StartNode.
Initializing a Raft group for the first time is awkward, since a
configuration has to be pulled from thin air. The way this is solved
today is unclean: The app is supposed to pass peers to StartNode(),
we add configuration changes for them to the log, immediately pretend
that they are applied, but actually leave them unapplied (to give the
app a chance to observe them, though if the app did decide to not apply
them things would really go off the rails), and then return control to
the app. The app will then process the initial Readys and as a result
the configuration will be persisted to disk; restarts of the node then
use RestartNode which doesn't take any peers.
The code that did this lived awkwardly in two places fairly deep down
the callstack, though it was really only necessary in StartNode(). This
commit refactors things to make this more obvious: only StartNode does
this dance now. In particular, RawNode does not support this at all any
more; it expects the app to set up its Storage correctly.
Future work may provide helpers to make this "preseeding" of the Storage
more user-friendly. It isn't entirely straightforward to do so since
the Storage interface doesn't provide the right accessors for this
purpose. Briefly speaking, we want to make sure that a non-bootstrapped
node can never catch up via the log so that we can implicitly use one
of the "skipped" log entries to represent the configuration change into
the bootstrap configuration. This is an invasive change that affects
all consumers of raft, and it is of lower urgency since the code (post
this commit) already encapsulates the complexity sufficiently.
It has always bugged me that any new feature essentially needed to be
tested twice due to the two ways in which apps can use raft (`*node` and
`*RawNode`). Due to upcoming testing work for joint consensus, now is a
good time to rectify this somewhat.
This commit removes most logic from `(*node).run` and uses `*RawNode`
internally. This simplifies the logic and also lead (via debugging) to
some insight on how the semantics of the approaches differ, which is now
documented in the comments.
Now that a Config is also added to the full status, the old name
did not convey the intention, which was to get a Status without
an associated allocation.
Put all the logic related to applying a configuration change in one
place in preparation for adding joint consensus.
This inspired various TODOs.
I had to rewrite TestSnapshotSucceedViaAppResp since it was relying
on a snapshot applied to the leader, which is now prevented.
Mechanically extract `progressTracker`, `Progress`, and `inflights`
to their own package named `tracker`. Add lots of comments in the
progress, and take the opportunity to rename and clarify various
fields.
This doesn't completely eliminate access to prs.nodes, but that's not
really necessary. This commit uses the existing APIs in a few more
places where it's convenient, and also sprinkles some assertions.
The Progress maps contain both the active configuration and information
about the replication status. By pulling it into its own component, this
becomes easier to unit test and also clarifies the code, which will see
changes as etcd-io/etcd#7625 is addressed.
More functionality will move into `prs` in self-contained follow-up commits.
The suggested pattern for Raft proposals is that they be retried
periodically until they succeed. This turns out to be an issue
when a leader cannot commit entries because the leader will continue
to append re-proposed entries to its log without committing anything.
This can result in the uncommitted tail of a leader's log growing
without bound until it is able to commit entries.
This change add a safeguard to protect against this case where a
leader's log can grow without bound during loss of quorum scenarios.
It does so by introducing a new, optional ``MaxUncommittedEntriesSize
configuration. This config limits the max aggregate size of uncommitted
entries that may be appended to a leader's log. Once this limit
is exceeded, proposals will begin to return ErrProposalDropped
errors.
See cockroachdb/cockroach#27772
In #9982, a mechanism to limit the size of `CommittedEntries` was
introduced. The way this mechanism worked was that it would load
applicable entries (passing the max size hint) and would emit a
`HardState` whose commit index was truncated to match the limitation
applied to the entries. Unfortunately, this was subtly incorrect
when the user-provided `Entries` implementation didn't exactly
match what Raft uses internally. Depending on whether a `Node` or
a `RawNode` was used, this would either lead to regressing the
HardState's commit index or outright forgetting to apply entries,
respectively.
Asking implementers to precisely match the Raft size limitation
semantics was considered but looks like a bad idea as it puts
correctness squarely in the hands of downstream users. Instead, this
PR removes the truncation of `HardState` when limiting is active
and tracks the applied index separately. This removes the old
paradigm (that the previous code tried to work around) that the
client will always apply all the way to the commit index, which
isn't true when commit entries are paginated.
See [1] for more on the discovery of this bug (CockroachDB's
implementation of `Entries` returns one more entry than Raft's when the
size limit hits).
[1]: https://github.com/cockroachdb/cockroach/issues/28918#issuecomment-418174448
Scanning the uncommitted portion of the raft log to determine whether
there are any pending config changes can be expensive. In
cockroachdb/cockroach#18601, we've seen that a new leader can spend so
much time scanning its log post-election that it fails to send
its first heartbeats in time to prevent a second election from
starting immediately.
Instead of tracking whether a pending config change exists with a
boolean, this commit tracks the latest log index at which a pending
config change *could* exist. This is a less expensive solution to
the problem, and the impact of false positives should be minimal since
a newly-elected leader should be able to quickly commit the tail of
its log.
TickQuiesced allows the caller to support "quiesced" Raft groups which
do not perform periodic heartbeats and elections. This is useful in a
system with thousands of Raft groups where these periodic operations can
be overwhelming in an otherwise idle system.
It might seem possible to avoid advancing the logical clock at all in
such Raft groups, but doing so has an interaction with the CheckQuorum
functionality. If a follower is not quiesced while the leader is the
follower can call an election that will fail because the leader's lease
has not expired (electionElapsed < electionTimeout). The next time the
leader sends a heartbeat to this follower the follower will see that the
heartbeat is from a previous term and respond with a MsgAppResp. This in
turn will cause the leader to step down and become a follower even
though there isn't a leader in the group. By allowing the leader's
logical clock to advance via TickQuiesced, the leader won't reject the
election and there will be a smooth transfer of leadership to the
follower.
Tobias Schottdorf