Commit Graph

22 Commits

Author SHA1 Message Date
Matt Keeler 8a968299dd
Fix flaky tests in the agent/grpc/public/services/serverdiscovery package (#13173)
Occasionally we had seen the TestWatchServers_ACLToken_PermissionDenied be flagged as flaky in circleci. This change should fix that.

Why it fixes it is complicated. The test was failing with a panic when a mocked ACL Resolver was being called more times than expected. I struggled for a while to determine how that could be. This test should call authorize once and only once and the error returned should cause the stream to be terminated and the error returned to the gRPC client. Another oddity was no amount of running this test locally seemed to be able to reproduce the issue. I ran the test hundreds of thousands of time and it always passed.

It turns out that there is nothing wrong with the test. It just so happens that the panic from unexpected invocation of a mocked call happened during the test but was caused by a previous test (specifically the TestWatchServers_StreamLifecycle test)

The stream from the previous test remained open after all the test Cleanup functions were run and it just so happened that when the EventPublisher eventually picked up that the context was cancelled during cleanup, it force closes all subscriptions which causes some loops to be re-entered and the streams to be reauthorized. Its that looping in response to forced subscription closures that causes the mock to eventually panic. All the components, publisher, server, client all operate based on contexts. We cancel all those contexts but there is no syncrhonous way to know when they are stopped.

We could have implemented a syncrhonous stop but in the context of an actual running Consul, context cancellation + async stopping is perfectly fine. What we (Dan and I) eventually thought was that the behavior of grpc streams such as this when a server was shutting down wasn’t super helpful. What we would want is for a client to be able to distinguish between subscription closed because something may have changed requiring re-authentication and subscription closed because the server is shutting down. That way we can send back appropriate error messages to detail that the server is shutting down and not confuse users with potentially needing to resubscribe.

So thats what this PR does. We have introduced a shutting down state to our event subscriptions and the various streaming gRPC services that rely on the event publisher will all just behave correctly and actually stop the stream (not attempt transparent reauthorization) if this particular error is the one we get from the stream. Additionally the error that gets transmitted back through gRPC when this does occur indicates to the consumer that the server is going away. That is more helpful so that a client can then attempt to reconnect to another server.
2022-05-23 08:59:13 -04:00
Matt Keeler 2a4ca71d3f
Move to using a shared EventPublisher (#12673)
Previously we had 1 EventPublisher per state.Store. When a state store was closed/abandoned such as during a consul snapshot restore, this had the behavior of force closing subscriptions for that topic and evicting event snapshots from the cache.

The intention of this commit is to keep all that behavior. To that end, the shared EventPublisher now supports the ability to refresh a topic. That will perform the force close + eviction. The FSM upon abandoning the previous state.Store will call RefreshTopic for all the topics with events generated by the state.Store.
2022-04-12 09:47:42 -04:00
Dan Upton e48c1611ee
WatchRoots gRPC endpoint (#12678)
Adds a new gRPC streaming endpoint (WatchRoots) that dataplane clients will
use to fetch the current list of active Connect CA roots and receive new
lists whenever the roots are rotated.
2022-04-05 15:26:14 +01:00
Dan Upton ebdda4848f
streaming: split event buffer by key (#12080) 2022-01-28 12:27:00 +00:00
R.B. Boyer 1cef3c99c2
state: adjust streaming event generation to account for partitioned nodes (#10860)
Also re-enabled some tests that had to be disabled in the prior PR.
2021-08-17 16:49:26 -05:00
Daniel Nephin 20f7a72792 stream: remove bufferItem.NextLink
Both NextLink and NextNoBlock had the same logic, with slightly
different return values. By adding a bool return value (similar to map
lookups) we can remove the duplicate method.
2021-06-07 17:04:46 -04:00
Daniel Nephin a29b848e3b stream: fix a snapshot cache bug
Previously a snapshot created as part of a resumse-stream request could have incorrectly
cached the newSnapshotToFollow event. This would cause clients to error because they
received an unexpected framing event.
2021-02-16 12:52:23 -05:00
Daniel Nephin e4a78c977d stream: document that Payload must be immutable
If they are sent to EventPublisher.Publish.

Also document that PayloadEvents is expected to come from a subscription and that it is
not immutable.
2020-11-06 13:00:33 -05:00
Daniel Nephin b57c7afcbb stream: include the namespace in the snap cache key
Otherwise the wrong snapshot could be returned when the same key is used in different namespaces
2020-10-30 14:34:04 -04:00
Daniel Nephin 61ce0964a4 stream: remove Event.Key
Makes Payload a type with FilterByKey so that Payloads can implement
filtering by key. With this approach we don't need to expose a Namespace
field on Event, and we don't need to invest micro formats or require a
bunch of code to be aware of exactly how the key field is encoded.
2020-10-28 16:48:04 -04:00
Daniel Nephin f857aef4a8 submatview: add a test for handling of NewSnapshotToFollow
Also add some godoc
Rename some vars and functions
Fix a data race in the new cache test for entry closing.
2020-10-06 13:22:02 -04:00
Daniel Nephin 9c5181c897 stream: full test coverage for EventPublisher.Subscribe 2020-10-02 13:46:24 -04:00
Daniel Nephin 0769f54fe1 stream: refactor to support change in framing events
Removing EndOfEmptySnapshot, add NewSnapshotToFollow
2020-10-02 13:41:31 -04:00
Daniel Nephin d192b0a080 stream: move goroutine out of New
This change will make it easier to manage goroutine lifecycle from the caller.

Also expose EventPublisher from state.Store
2020-09-28 18:40:10 -04:00
Daniel Nephin 81cc3daf69 stream: have SnapshotFunc accept a non-pointer SubscribeRequest
The value is not expected to be modified. Passing a value makes that explicit.
2020-09-03 16:54:02 -04:00
Daniel Nephin decba06b7d stream: close all subs when EventProcessor is shutdown. 2020-07-22 19:04:10 -04:00
Daniel Nephin a2f8605c66 stream: Add forceClose and refactor subscription filtering
Move the subscription context to Next. context.Context should generally
never be stored in a struct because it makes that struct only valid
while the context is valid. This is rarely obvious from the caller.
Adds a forceClosed channel in place of the old context, and uses the new
context as a way for the caller to stop the Subscription blocking.

Remove some recursion out of bufferImte.Next. The caller is already looping so we can continue
in that loop instead of recursing. This ensures currentItem is updated immediately (which probably
does not matter in practice), and also removes the chance that we overflow the stack.

NextNoBlock and FollowAfter do not need to handle bufferItem.Err, the caller already
handles it.

Moves filter to a method to simplify Next, and more explicitly separate filtering from looping.

Also improve some godoc

Only unwrap itemBuffer.Err when necessary
2020-07-14 15:57:47 -04:00
Daniel Nephin 2595436f62 stream: Improve docstrings
Also rename ResumeStrema to EndOfEmptySnapshot to be more consistent with other framing events

Co-authored-by: Paul Banks <banks@banksco.de>
2020-07-14 15:57:47 -04:00
Daniel Nephin 16a2b3fafc stream: change Topic to an interface
Consumers of the package can decide on which type to use for the Topic. In the future we may
use a gRPC type for the topic.
2020-07-14 15:57:47 -04:00
Daniel Nephin aa571bd0ce state: Move change processing out of EventPublisher
EventPublisher was receiving TopicHandlers, which had a couple of
problems:

- ChangeProcessors were being grouped by Topic, but they completely
  ignored the topic and were performed on every change
- ChangeProcessors required EventPublisher to be aware of database
  changes

By moving ChangeProcesors out of EventPublisher, and having Publish
accept events instead of changes, EventPublisher no longer needs to
be aware of these things.

Handlers is now only SnapshotHandlers, which are still mapped by Topic.

Also allows us to remove the small 'db' package that had only two types.
They can now be unexported types in state.
2020-07-14 15:57:47 -04:00
Daniel Nephin e1305fe80c stream: unexport identifiers
Now that EventPublisher is part of stream a lot of the internals can be hidden
2020-07-14 15:57:47 -04:00
Daniel Nephin 9e37894778 stream: Move EventPublisher to stream package
The EventPublisher is the central hub of the PubSub system. It is toughly coupled with much of
stream. Some stream internals were exported exclusively for EventPublisher.

The two Subscribe cases (with or without index) were also awkwardly split between two packages. By
moving EventPublisher into stream they are now both in the same package (although still in different files).
2020-07-14 15:57:47 -04:00
Renamed from agent/consul/state/event_publisher.go (Browse further)