The client retries connection without backoff when the server is gone
after the watch stream is established. This results in high CPU usage
in the client process. This change introduces backoff when the stream is
failed and unavailable.
Signed-off-by: Hisanobu Tomari <posco.grubb@gmail.com>
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>
Since http2 spec defines the receive windows's size and max size of
frame in the stream, the underlayer - gRPC client can pre-read data
from server even if the application layer hasn't read it yet.
And the initialized cluster has 20KiB snapshot, which can be pre-read
by underlayer. We should increase the snapshot's size, just in case
that the io.Copy won't return the canceled or timeout error.
Fixes: #14477
Signed-off-by: Wei Fu <fuweid89@gmail.com>
etcdserver starts to using generic in https://github.com/etcd-io/etcd/pull/14515,
so any applications which depend on package `go.etcd.io/etcd/server/v3' must be
built with go version >= 1.18.
Most likely we may depends on go 1.18 new features in other packages, i.e. client, pkg,
so we just simplify the changelog message to require applications depending on any
etcd 3.6 pacakges must be built with go version >= 1.18, although only etcdserver uses
go generic for now.
Signed-off-by: Benjamin Wang <wachao@vmware.com>
This commit simplifies the logic added in 37c7e4d to auto-leave joint
configurations. It does so by making the following adjustments to the
code:
- remove the `oldApplied <= r.pendingConfIndex` condition. This does
not seem necessary. When a node first attempts to auto-leave a joint
config, it will bump `r.pendingConfIndex` when proposing. In cases
where `oldApplied >= r.pendingConfIndex`, the proposal must have
already been applied. Reviewers should double check this.
- use raft.Step instead of custom proposal code. This code was already
present in stepLeader, so there was no reason to duplicate it. This
would have avoided bugs like the one we fixed in #14538.
- use `confChangeToMsg` to generate message, to centralize the creation
of all `MsgProp{EntryConfChange}` messages.
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>
When a key-value store corruption check happens immediately after a
compaction, the revision at which the key-value store hash is computed,
is the compacted revision itself.
In that case, the hash computation logic was incorrect because it
returned an ErrCompacted error; this error should instead be returned when
the revision at which the key-value store is hashed, is strictly lower
than the compacted revision.
Fixes#14325
Signed-off-by: Jeremy Leach <44558776+jbml@users.noreply.github.com>
Some comments in the file are formatted automatically into ugly style,
because the hierarchical structure is missing. After removing the
leading numbers in the comments, `go fmt` will not format the comments
anymore.
Signed-off-by: Benjamin Wang <wachao@vmware.com>
