We need to watch for changes to peerings and update the server addresses which get served by the ring buffer.
Also, if there is an active connection for a peer, we are getting up-to-date server addresses from the replication stream and can safely ignore the token's addresses which may be stale.
* Avoid logging StreamSecretID
* Wrap additional errors in stream handler
* Fix flakiness in leader test and rename servers for clarity. There was
a race condition where the peering was being deleted in the test
before the stream was active. Now the test waits for the stream to be
connected on both sides before deleting the associated peering.
* Run flaky test serially
When we receive a FailedPrecondition error, retry that more quickly
because we expect it will resolve shortly. This is particularly
important in the context of Consul servers behind a load balancer
because when establishing a connection we have to retry until we
randomly land on a leader node.
The default retry backoff goes from 2s, 4s, 8s, etc. which can result in
very long delays quite quickly. Instead, this backoff retries in 8ms
five times, then goes exponentially from there: 16ms, 32ms, ... up to a
max of 8152ms.
Ensure that the peer stream replication rpc can successfully be used with TLS activated.
Also:
- If key material is configured for the gRPC port but HTTPS is not
enabled now TLS will still be activated for the gRPC port.
- peerstream replication stream opened by the establishing-side will now
ignore grpc.WithBlock so that TLS errors will bubble up instead of
being awkwardly delayed or suppressed
Previously, public referred to gRPC services that are both exposed on
the dedicated gRPC port and have their definitions in the proto-public
directory (so were considered usable by 3rd parties). Whereas private
referred to services on the multiplexed server port that are only usable
by agents and other servers.
Now, we're splitting these definitions, such that external/internal
refers to the port and public/private refers to whether they can be used
by 3rd parties.
This is necessary because the peering replication API needs to be
exposed on the dedicated port, but is not (yet) suitable for use by 3rd
parties.
- Use some protobuf construction helper methods for brevity.
- Rename a local variable to avoid later shadowing.
- Rename the Nonce field to be more like xDS's naming.
- Be more explicit about which PeerID fields are empty.
Peer replication is intended to be between separate Consul installs and
effectively should be considered "external". This PR moves the peer
stream replication bidirectional RPC endpoint to the external gRPC
server and ensures that things continue to function.
When our peer deletes the peering it is locally marked as terminated.
This termination should kick off deleting all imported data, but should
not delete the peering object itself.
Keeping peerings marked as terminated acts as a signal that the action
took place.
Once a peering is marked for deletion a new leader routine will now
clean up all imported resources and then the peering itself.
A lot of the logic was grabbed from the namespace/partitions deferred
deletions but with a handful of simplifications:
- The rate limiting is not configurable.
- Deleting imported nodes/services/checks is done by deleting nodes with
the Txn API. The services and checks are deleted as a side-effect.
- There is no "round rate limiter" like with namespaces and partitions.
This is because peerings are purely local, and deleting a peering in
the datacenter does not depend on deleting data from other DCs like
with WAN-federated namespaces. All rate limiting is handled by the
Raft rate limiter.
1. Fix a bug where the peering leader routine would not track all active
peerings in the "stored" reconciliation map. This could lead to
tearing down streams where the token was generated, since the
ConnectedStreams() method used for reconciliation returns all streams
and not just the ones initiated by this leader routine.
2. Fix a race where stream contexts were being canceled before
termination messages were being processed by a peer.
Previously the leader routine would tear down streams by canceling
their context right after the termination message was sent. This
context cancelation could be propagated to the server side faster
than the termination message. Now there is a change where the
dialing peer uses CloseSend() to signal when no more messages will
be sent. Eventually the server peer will read an EOF after receiving
and processing the preceding termination message.
Using CloseSend() is actually not enough to address the issue
mentioned, since it doesn't wait for the server peer to finish
processing messages. Because of this now the dialing peer also reads
from the stream until an error signals that there are no more
messages. Receiving an EOF from our peer indicates that they
processed the termination message and have no additional work to do.
Given that the stream is being closed, all the messages received by
Recv are discarded. We only check for errors to avoid importing new
data.
