each+configure serves no purpose, because the task has already been
realized with the each. This was just a faulty conversion in 0ff9f37b9e,
although it was because I thought the each was specialized API from the
protobuf plugin, not the normal container each.
We have two versions of multiple projects, the non-shadow and the
shadow. When using project() references, things are mostly fine,
although we may test with a different version than our users would use.
However, when a dependency from Maven Central depends on a grpc
artifact, it pulls in the non-shadow configuration and messes up
gradle's tracking. We shouldn't be using the non-shadow configuration,
so this issue is sort of a blessing to improve the test reliability.
Gradle 7 will notice that there's a missing dependency, but it is not
deterministic and is very slow to test. Starting in Gradle 8 it is much
more reliably triggered which greatly helped the testing efforts. The
error in Gradle 7 looks like:
```
> Task :grpc-gcp-observability:interop:startScripts
Execution optimizations have been disabled for task ':grpc-gcp-observability:interop:startScripts' to ensure correctness due to the following reasons:
- Gradle detected a problem with the following location: 'grpc-java/netty/shaded/build/libs/grpc-netty-shaded-1.54.0-SNAPSHOT-original.jar'. Reason: Task ':grpc-gcp-observability:interop:startScripts' uses this output of task ':grpc-netty-shaded:jar' without declaring an explicit or implicit dependency. This can lead to incorrect results being produced, depending on what order the tasks are executed. Please refer to https://docs.gradle.org/7.6/userguide/validation_problems.html#implicit_dependency for more details about this problem.
```
Fixes#9992
* xds: handle the handlerRemoved callback to skip updateSslContext processing
In handlerAdded we submit a callback to updateSslContext but before the
callback is executed the handler could be removed (e.g. bad connection)
in which case the callback should skip all of the processing.
Also added a unit test to check there is no exception.
If provided with the new PickFirstLoadBalancerConfig,
PickFirstLoadBalancer will shuffle the list of addresses it receives
from the name resolver.
PickFirstLoadBalancerProvider will now support the new config
if enabled by an env variable.
When the subchannel is transitioning from TRANSIENT_FAILURE to either
IDLE or CONNECTING we will not update the LB state. Additionally, if
the subchannel becomes idle we request a new connection so that the
subchannel will keep on trying to establish a connection.
All BinderTransport transactions are oneway which means uncaught
Exceptions during processing are merely logged locally and not
propagated to the peer. Instead, we add a top level catch block
that handles the unexpected by shutting down the whole transport. This
makes our peer aware of the problem immediately (instead of relying on a
deadline) and gives clients a fresh transport instance to handle any
retries.
Under normal conditions the child LB of `ClusterImplLoadBalancer` does
not fluctuate, based on the field used to configure load balancing in
the xDS `Cluster` proto it is either:
1. `WrrLocalityLoadBalancer` if the newer `load_balancing_policy` field
is used
2. `WeightedTargetLoadBalancer` if the legacy `lb_policy` field is used
`ClusterImplLoadBalancer` currently assumes that this child does not
change and so does not change the child LB when the name resolver sends an
update. If the control plane does switch to using a different field for
LB config, that update will have an LB config meant for the other child
LB type. This will result in a ClassCastException and a channel panic.
To address this, `ClusterImplLoadBalancer` will now use
`GracefulSwitchLoadBalancer` and makes sure if the child policy changes
the correct LB implementation is switched to.
The internal build rule for this client already adds these dependencies, and is what related tests for these dependencies have been relying on for a while.
We're starting to use the per-release published interop client docker images for those tests too now, and these use the gradle build rules.
The xds server can take a really long time to start if the xds resources
are slow to load. Ideally the management server would be available
during this time so we can inspect the server. The server health still
won't go to SERVING until the xds server starts, which is appropriate.
There's still plenty more that could be done, but I want to keep this on
the simpler/less-invasive side and it'd just delay these changes for no
real benefit.
Previously builds were done with Ubuntu 16.04, and now we are using
18.04. Thus the generated binaries will no longer work for
Ubuntu 16.04 and Debian 9 users, both of which are outside of their
support window and aren't supported by Abseil. RHEL users are
unaffected, as the binaries already didn't work on RHEL 7 and they will
remain working with RHEL 8. FWIW, Ubuntu 18.04 will leave its support
window in June.
The point of the sorting is to reduce the chances of merge conflicts. I
greatly prefer verboseness over cleverness in examples, but the tasks
can only be sorted manually and there's so many of them.
It is counter-productive to do this for the examples that have their own
project folder, as there's so few tasks in that case that they don't
need to be ordered.
The version used by protoc-gen-grpc-java will be upgraded separately,
because of large C++ build changes necessary. But that won't impact
users at all. We are upgrading to protoc 22.3; only the grpc plugin is
not upgraded.
Bazel is upgraded for both Java and C++.
If a child load balancer rejects the addresses it if given all we can do
is to trigger a name resolution refresh and hope for a better set of
addresses.
This removes some steps from the release process. These two locations
aren't special enough in way that deserves manually changing the version
each release.
Eric Anderson
Terry Wilson
yifeizhuang
sanjaypujare
Kaloyan Simitchiyski
chenwei321
Larry Safran
John Cormie
apolcyn
Ashitha Santhosh
Sergii Tkachenko
Carl Mastrangelo