These are nonsensical and a network implementation is not required
to handle them correctly, so panic instead of sending them out.
Signed-off-by: Nathan VanBenschoten <nvanbenschoten@gmail.com>
This commit ensures that the raft leader eagerly broadcasts a MsgApp to
each follower when initiating an automatic transition out of a joint
configuration. This had been missed previously, which could lead to
delayed completion of an auto-transition.
Signed-off-by: Nathan VanBenschoten <nvanbenschoten@gmail.com>
This commit updates the `RawNode`'s `prevHardSt` to the new HardState in
`acceptReady` instead of on `Advance`. This aligns the handling of
`prevHardSt` with that of `prevSoftSt` (and other fields like `msgs`)
and simplifies the logic in `Advance`.
Signed-off-by: Nathan VanBenschoten <nvanbenschoten@gmail.com>
This reverts commit d73a986e4edb15ef9dbfc994f1cbf5e96694d877, which
was added only for benchmarking purposes.
Signed-off-by: Tobias Grieger <tobias.b.grieger@gmail.com>
This makes the alternative option of implementing the leader's self-ack
of entry append the default.
Signed-off-by: Tobias Grieger <tobias.b.grieger@gmail.com>
This is tested directly at the level of `RawNode` in
`TestRawNodeReadIndex`. `*node` is a thin wrapper around `RawNode` so
this is sufficient.
The reason to remove the test is that it now incurs data races
since it's not possible to adjust the `readStates` and `step`
fields while the node is running, and there is no primitive
to synchronize with its goroutine. This could all be fixed
but it's not worth it.
Signed-off-by: Tobias Grieger <tobias.b.grieger@gmail.com>
This now needs an additional Ready cycle to apply the previous conf change,
so the finalizing conf change does too.
Signed-off-by: Tobias Grieger <tobias.b.grieger@gmail.com>
Leader only acks to itself on `(*raft).advance` so we have to
make this test a bit more like the real thing.
Signed-off-by: Tobias Grieger <tobias.b.grieger@gmail.com>
This was expecting the progress of the leader to be updated as a
result of MsgProp but it is now happening in `(*raft).advance`.
Signed-off-by: Tobias Grieger <tobias.b.grieger@gmail.com>
This fixes essentially all tests using this, since now they don't have
to do anything special about the extra cycle introduced for single
nodes.
Signed-off-by: Tobias Grieger <tobias.b.grieger@gmail.com>
Switched this to baking the conf changes into the initial state
to have fewer cycles to walk through in the test.
Signed-off-by: Tobias Grieger <tobias.b.grieger@gmail.com>
This needed to apply entries from CommittedEntries, not Entries.
Previously the test got away with it because the two slices were
equal. Now it was hanging because when it proposed the second
conf change the first one hadn't applied yet, and so it got dropped,
and the test would hang.
Signed-off-by: Tobias Grieger <tobias.b.grieger@gmail.com>
`StartNode` runs a naked goroutine, so it's impossible to test against
it in a way that will reliably produce contained test failures when
assertions are hit on the `(*node).run` goroutine.
This commit introduces a harness that we can use in tests to wrap
this goroutine and allow it to defer errors to `*testing.T`.
Note that tests of `Node` still need to be architected carefully
since it's easy to produce a deadlock in them should things not
go exactly as planned.
Signed-off-by: Tobias Grieger <tobias.b.grieger@gmail.com>
See https://github.com/etcd-io/etcd/issues/14370.
When run in a single-voter configuration, prior to this PR
raft would emit `HardState`s that would emit a proposed `Entry`
simultaneously in `CommittedEntries` and `Entries`.
To be correct, this requires users of the raft library to enforce an
ordering between appending to the log and notifying the client about
`CommittedEntries` also present in `Entries`. This was easy to miss.
Walk back this behavior to arrive at a simpler contract: what's
emitted in `CommittedEntries` is truly committed, i.e. present
in stable storage on a quorum of voters.
This in turn pessimizes the single-voter case: rather than fully
handling an `Entry` in just one `Ready`, now two are required,
and in particular one has to do extra work to save on allocations.
We accept this as a good tradeoff, since raft primarily serves
multi-voter configurations.
Signed-off-by: Tobias Grieger <tobias.b.grieger@gmail.com>
Show-cases the current behavior and changes made in future commits for [^1].
The test demonstrates that a single-voter raft instance will emit an
entry as committed while it still needs to be appended to the log.
[^1]: https://github.com/etcd-io/etcd/issues/14370
Signed-off-by: Tobias Grieger <tobias.b.grieger@gmail.com>
This is a speed-of-light benchmark that uses an in-memory single-voter
RawNode to sequentially propose and process entries.
As a bonus, it also measures the number of calls to the underlying
Storage. Calls to the Storage are cheap since the benchmark is in-
memory, but in a real-world implementation, especially one that doesn't
cache results, additional calls to the Storage interface can translate
to a heavy hit as they might involve additional I/O.
Signed-off-by: Tobias Grieger <tobias.b.grieger@gmail.com>
RecentActive is now initialized to true in `becomeLeader`. Both
configuration changes and CheckQuorum make sure not to break this,
so we now now that the leader is always RecentActive.
Signed-off-by: Tobias Grieger <tobias.b.grieger@gmail.com>
In the test `testNonleaderElectionTimeoutRandomized`, the possible values of election timeout randomized out should be 10-19(inclusive), but the current test only tests the possible values as 11-19(inclusive), which is incorrect here. We need to expand the scope of the tests to ensure robustness, and also to make the logic clearer.
Signed-off-by: Zike Yang <zike@apache.org>
