Read requests performed during anti antropy full sync currently target
the leader only. This generates a non-negligible load on the leader when
the DC is large enough and can be offloaded to the followers following
the "eventually consistent" policy for the agent state.
We switch the AE read calls to use stale requests with a small (2s)
MaxStaleDuration value and make sure we do not read too fast after a
write.
The observed bug was that a full restart of a consul datacenter (servers
and clients) in conjunction with a restart of a connect-flavored
application with bring-your-own-service-registration logic would very
frequently cause the envoy sidecar service check to never reflect the
aliased service.
Over the course of investigation several bugs and unfortunate
interactions were corrected:
(1)
local.CheckState objects were only shallow copied, but the key piece of
data that gets read and updated is one of the things not copied (the
underlying Check with a Status field). When the stock code was run with
the race detector enabled this highly-relevant-to-the-test-scenario field
was found to be racy.
Changes:
a) update the existing Clone method to include the Check field
b) copy-on-write when those fields need to change rather than
incrementally updating them in place.
This made the observed behavior occur slightly less often.
(2)
If anything about how the runLocal method for node-local alias check
logic was ever flawed, there was no fallback option. Those checks are
purely edge-triggered and failure to properly notice a single edge
transition would leave the alias check incorrect until the next flap of
the aliased check.
The change was to introduce a fallback timer to act as a control loop to
double check the alias check matches the aliased check every minute
(borrowing the duration from the non-local alias check logic body).
This made the observed behavior eventually go away when it did occur.
(3)
Originally I thought there were two main actions involved in the data race:
A. The act of adding the original check (from disk recovery) and its
first health evaluation.
B. The act of the HTTP API requests coming in and resetting the local
state when re-registering the same services and checks.
It took awhile for me to realize that there's a third action at work:
C. The goroutines associated with the original check and the later
checks.
The actual sequence of actions that was causing the bad behavior was
that the API actions result in the original check to be removed and
re-added _without waiting for the original goroutine to terminate_. This
means for brief windows of time during check definition edits there are
two goroutines that can be sending updates for the alias check status.
In extremely unlikely scenarios the original goroutine sees the aliased
check start up in `critical` before being removed but does not get the
notification about the nearly immediate update of that check to
`passing`.
This is interlaced wit the new goroutine coming up, initializing its
base case to `passing` from the current state and then listening for new
notifications of edge triggers.
If the original goroutine "finishes" its update, it then commits one
more write into the local state of `critical` and exits leaving the
alias check no longer reflecting the underlying check.
The correction here is to enforce that the old goroutines must terminate
before spawning the new one for alias checks.
Prevent race between register and deregister requests by saving them
together in the local state on registration.
Also adds more cleaning in case of failure when registering services
/ checks.
Fixes: https://github.com/hashicorp/consul/issues/3676
This fixes a bug were registering an agent with a non-existent ACL token can prevent other
services registered with a good token from being synced to the server when using
`acl_enforce_version_8 = false`.
## Background
When `acl_enforce_version_8` is off the agent does not check the ACL token validity before
storing the service in its state.
When syncing a service registered with a missing ACL token we fall into the default error
handling case (https://github.com/hashicorp/consul/blob/master/agent/local/state.go#L1255)
and stop the sync (https://github.com/hashicorp/consul/blob/master/agent/local/state.go#L1082)
without setting its Synced property to true like in the permission denied case.
This means that the sync will always stop at the faulty service(s).
The order in which the services are synced is random since we iterate on a map. So eventually
all services with good ACL tokens will be synced, this can however take some time and is influenced
by the cluster size, the bigger the slower because retries are less frequent.
Having a service in this state also prevent all further sync of checks as they are done after
the services.
## Changes
This change modify the sync process to continue even if there is an error.
This fixes the issue described above as well as making the sync more error tolerant: if the server repeatedly refuses
a service (the ACL token could have been deleted by the time the service is synced, the servers
were upgraded to a newer version that has more strict checks on the service definition...).
Then all services and check that can be synced will, and those that don't will be marked as errors in
the logs instead of blocking the whole process.
* Proxy Config Manager
This component watches for local state changes on the agent and ensures that each service registered locally with Kind == connect-proxy has it's state being actively populated in the cache.
This serves two purposes:
1. For the built-in proxy, it ensures that the state needed to accept connections is available in RAM shortly after registration and likely before the proxy actually starts accepting traffic.
2. For (future - next PR) xDS server and other possible future proxies that require _push_ based config discovery, this provides a mechanism to subscribe and be notified about updates to a proxy instance's config including upstream service discovery results.
* Address review comments
* Better comments; Better delivery of latest snapshot for slow watchers; Embed Config
* Comment typos
* Add upstream Stringer for funsies
- A new endpoint `/v1/agent/service/:service_id` which is a generic way to look up the service for a single instance. The primary value here is that it:
- **supports hash-based blocking** and so;
- **replaces `/agent/connect/proxy/:proxy_id`** as the mechanism the built-in proxy uses to read its config.
- It's not proxy specific and so works for any service.
- It has a temporary shim to call through to the existing endpoint to preserve current managed proxy config defaulting behaviour until that is removed entirely (tested).
- The built-in proxy now uses the new endpoint exclusively for it's config
- The built-in proxy now has a `-sidecar-for` flag that allows the service ID of the _target_ service to be specified, on the condition that there is exactly one "sidecar" proxy (that is one that has `Proxy.DestinationServiceID` set) for the service registered.
- Several fixes for edge cases for SidecarService
- A fix for `Alias` checks - when running locally they didn't update their state until some external thing updated the target. If the target service has no checks registered as below, then the alias never made it past critical.
* Refactor Service Definition ProxyDestination.
This includes:
- Refactoring all internal structs used
- Updated tests for both deprecated and new input for:
- Agent Services endpoint response
- Agent Service endpoint response
- Agent Register endpoint
- Unmanaged deprecated field
- Unmanaged new fields
- Managed deprecated upstreams
- Managed new
- Catalog Register
- Unmanaged deprecated field
- Unmanaged new fields
- Managed deprecated upstreams
- Managed new
- Catalog Services endpoint response
- Catalog Node endpoint response
- Catalog Service endpoint response
- Updated API tests for all of the above too (both deprecated and new forms of register)
TODO:
- config package changes for on-disk service definitions
- proxy config endpoint
- built-in proxy support for new fields
* Agent proxy config endpoint updated with upstreams
* Config file changes for upstreams.
* Add upstream opaque config and update all tests to ensure it works everywhere.
* Built in proxy working with new Upstreams config
* Command fixes and deprecations
* Fix key translation, upstream type defaults and a spate of other subtele bugs found with ned to end test scripts...
TODO: tests still failing on one case that needs a fix. I think it's key translation for upstreams nested in Managed proxy struct.
* Fix translated keys in API registration.
≈
* Fixes from docs
- omit some empty undocumented fields in API
- Bring back ServiceProxyDestination in Catalog responses to not break backwards compat - this was removed assuming it was only used internally.
* Documentation updates for Upstreams in service definition
* Fixes for tests broken by many refactors.
* Enable travis on f-connect branch in this branch too.
* Add consistent Deprecation comments to ProxyDestination uses
* Update version number on deprecation notices, and correct upstream datacenter field with explanation in docs
There are also a lot of small bug fixes found when testing lots of things end-to-end for the first time and some cleanup now it's integrated with real CA code.
The anti-entropy tests relied on the side-effect of the StartSync()
method to perform a full sync instead of a partial sync. This lead to
multiple anti-entropy go routines being started unnecessary retry loops.
This change changes the behavior to perform synchronous full syncs when
necessary removing the need for all of the time.Sleep and most of the
retry loops.
The state of the service and health check records was spread out over
multiple maps guarded by a single lock. Access to the maps has to happen
in a coordinated effort and the tests often violated this which made
them brittle and racy.
This patch replaces the multiple maps with a single one for both checks
and services to make the code less fragile.
This is also necessary since moving the local state into its own package
creates circular dependencies for the tests. To avoid this the tests can
no longer access internal data structures which they should not be doing
in the first place.
The tests still don't compile but this is a ncessary step in that
direction.
Aestek