Move the creation of LoadStatsStore (aka, the stats object) into XdsClient. The XdsClient is responsible for managing the lifetime of stats objects. Creations of LoadStatsStores are reference counted so that multiple EDS policies can retrieve the same stats object for load recording. Counters for recording loads per locality also need to be reference counted, as each EDS policy for the same cluster will receive endpoints for the same group of localities, they will use the same load counters for recording each locality's loads.
A user has reported a GOAWAY with too_many_pings when using BDP. We
aren't certain why it is happening, but want to provide a way to disable
BDP while we continue investigating. b/162162973
Due to historical reasons, protobuf is in the proto2:: namespace
internally instead of google::protobuf. We have been maintaining diffs
that replace each occurence of one with the other. Instead we can simply
create a namespace alias and use that alias instead of the canonical
name. That greatly reduces the size of the diff and its likelihood to
break.
If the names ever align in the future, we can swap back to the canonical
names.
FALLTHROUGH_INTENDED was defined by Abseil, but is now getting an ABSL
prefix and the old name will be removed. Swapping to a new define name
to avoid redefining the existing-but-soon-to-be-deleted
FALLTHROUGH_INTENDED.
It's a lot of code and there are classes in Guava are better. This was noticed
with a lint checker. This commit does change the error-handling behavior, as
previous the code wrongly cancelled the Future instead of setting it to have an
exception.
verifyZeroInteractions has the same behavior as verifyNoMoreInteractions. It
was deprecated in Mockito 3.0.1 and replaced with verifyNoInteractions, which
does not change behavior depending on previous verify() calls. All instances
were replaced with verifyNoInteractions, except those in
ApplicationThreadDeframerTest which were replaced with verifyNoMoreInteractions
since there is a verify() call in `@Before`.
Adding `DelayedClientCall` in preparation of implementing `ConfigSelector` in core.
`DelayedClientCall` is implemented exactly in the same way as `DelayedStream`. Only added logic to monitor initial DEADLINE. Note that `ClientCall.cancel()` is not thread-safe and will cause exceptions if trying to start call after it, which is different from in the stream where cancel() is thread-safe and wouldn't trigger any checkState()s. The initial DEADLINE monitor should not call `ClientCall.cancel()` directly.
Duplicated `XdsClientImptTest` for V3. `XdsClientImptTestV2` and all other tests are still using V2. Even for `XdsClientImptTest`, although the protocol is V3, the test xds server still sends V2 resources in its V3 response.
Most of these are easy "replace X with Y."
The CreateStartScripts changes were because the scripts were being included in
the output zip/tar multiple times. The was because they were all using the same
output directory, and the entire output directory was being included for each.
The output directory tmp/ was particularly poor because other tasks were
dumping things into it, so our zip/tar was including those junk files as well.
Starting in Gradle 6.0 maven-publish began including sha256 and sha512
checksums, in addition to the previous md5 and sha1 checksums. We don't want
.sha256.asc and .sha512.asc files, as they serve no purpose.
The previous 'git ls-remote' was returning the tag-referenced commits via
vF.O.O^{} which was confusing the version check (as v1.30.0 would not match
v1.30.0^{}). Passing --refs filters those ^{} tags.
There was also a missing 'v' in the version check. It was added explicitly to
generate the list but not to check the result.
The routing LB policy should immediately the Channel's picker that delegates picking to the updated routes. Otherwise, new RPCs will keep being sent through old routes even if they are removed. This change also includes the fix for syncing state change for child balancers in deactivated state.
- Add support for send_all_clusters field in LRS response. When it is set to true, just send load reports for clusters that the client is currently tracking (aka, is sending load to).
- The actual load report interval (in each ClusterStats message, which contains the stats for each cluster:eds_service) should be tracked individually.
The top-level LB policy, which is an aggregator for CDS policies. It maintains the lifecycle of CDS LB policy instances. The pick argument taken from the Channel contains the information to determine which child CDS policy instance should the picking operation be delegated to.
The implementation is similar to the action part of what we currently have in the routing policy. The existing routing policy will be refactored to two parts, with the route match part moved into ConfigSelector and action part being this top-level LB policy.
We found that the interceptor approach for `ConfigSelector` would be adding a layer of indirection for no gain: The API Result selectConfig(LoadBalancer.PickSubchannelArgs args) consumes headers among other inputs, because route matching might need to match the headers; and the API produces ClientInterceptor among other outputs. But the headers is not available until clientCall.start(listner, headers), whereas the interceptor need be applied to the call before clientCall.start(). So the input is not available until the output is applied. That means we will need to delay calling the downstream newCall() (which will either be RealChannel or the interceptor) until start() is called.
So we want to change to the other approach similar to what c-core is taking: Have `Result(Object config, CallOptions, Runnable committedCallback)`, where CallOption is the selector modified CallOption, and committedCallback is used to monitor the call lifecycle.
sanjaypujare
Eric Anderson
Chengyuan Zhang
Patrice Chalin
ZHANG Dapeng
cindyxue
Eric Gribkoff
Menghan Li
Ran
susinmotion
ZhenLian