Fixing TSAN data races of the kind -
https://source.cloud.google.com/results/invocations/c3f02253-0d0b-44e6-917d-07e4bc0d3d62/targets/%2F%2Ftest%2Fcpp%2Fperformance:writes_per_rpc_test@poller%3Dpoll/log
I think the original issue was the non-atomic load in
writes_per_rpc_test.cc but also making the change to use barriers
instead of relaxed atomics (probably unimportant but I'll prefer safety
in the absence of otherwise comments).
Could probably also use std::atomic but feeling a bit lazy to make the
changes throughout.
<!--
If you know who should review your pull request, please assign it to
that
person, otherwise the pull request would get assigned randomly.
If your pull request is for a specific language, please add the
appropriate
lang label.
-->
<!--
If you know who should review your pull request, please assign it to
that
person, otherwise the pull request would get assigned randomly.
If your pull request is for a specific language, please add the
appropriate
lang label.
-->
This PR adds the view `grpc.io/client/api_latency` for GCP Observability
which aims to collect the end-to-end time taken by a call.
Changes made to support this -
1) A global interceptor factory registration is created for stats
plugins.
2) OpenCensus plugin now provides a new interceptor that's responsible
for collecting the new latency.
3) Gcp Observability registers this plugin.
4) A new OpenCensus measurement and view is created for api latency.
Note that this is internal as of now, since it's not clear if it should
be exposed as public experimental API. Leaving that decision for the
future.
<!--
If you know who should review your pull request, please assign it to
that
person, otherwise the pull request would get assigned randomly.
If your pull request is for a specific language, please add the
appropriate
lang label.
-->
This PR adds annotations to client attempt spans and server spans on
messages of the form -
* `Send message: 1026 bytes`
* `Send compressed message: 31 bytes` (if message was compressed)
* `Received message: 31 bytes`
* `Received decompressed message: 1026 bytes` (if message needed to be
decompressed)
Note that the compressed and decompressed annotations are not present if
compression/decompression was not performed.
<!--
If you know who should review your pull request, please assign it to
that
person, otherwise the pull request would get assigned randomly.
If your pull request is for a specific language, please add the
appropriate
lang label.
-->
This is a big rewrite of global config.
It does a few things, all somewhat intertwined:
1. centralize the list of configuration we have to a yaml file that can
be parsed, and code generated from it
2. add an initialization and a reset stage so that config vars can be
centrally accessed very quickly without the need for caching them
3. makes the syntax more C++ like (less macros!)
4. (optionally) adds absl flags to the OSS build
This first round of changes is intended to keep the system where it is
without major changes. We pick up absl flags to match internal code and
remove one point of deviation - but importantly continue to read from
the environment variables. In doing so we don't force absl flags on our
customers - it's possible to configure grpc without the flags - but
instead allow users that do use absl flags to configure grpc using that
mechanism. Importantly this lets internal customers configure grpc the
same everywhere.
Future changes along this path will be two-fold:
1. Move documentation generation into the code generation step, so that
within the source of truth yaml file we can find all documentation and
data about a configuration knob - eliminating the chance of forgetting
to document something in all the right places.
2. Provide fuzzing over configurations. Currently most config variables
get stashed in static constants across the codebase. To fuzz over these
we'd need a way to reset those cached values between fuzzing rounds,
something that is terrifically difficult right now, but with these
changes should simply be a reset on `ConfigVars`.
<!--
If you know who should review your pull request, please assign it to
that
person, otherwise the pull request would get assigned randomly.
If your pull request is for a specific language, please add the
appropriate
lang label.
-->
---------
Co-authored-by: ctiller <ctiller@users.noreply.github.com>
We currently take named metrics recorded per-request only. Instead we
should merge field-wise.
<!--
If you know who should review your pull request, please assign it to
that
person, otherwise the pull request would get assigned randomly.
If your pull request is for a specific language, please add the
appropriate
lang label.
-->
1. Make channel creation lazy. This allows test cases to update the
configuration before the connection is made.
2. Pass load reports tracker when creating the policy. This way other
test cases do not see any changes to ChannelArguments.
Using grpc_core::CoreConfiguration::RunWithSpecialConfiguration was
considered but did not work as it removes other builders setup prior to
starting the test cases.
<!--
If you know who should review your pull request, please assign it to
that
person, otherwise the pull request would get assigned randomly.
If your pull request is for a specific language, please add the
appropriate
lang label.
-->
---------
Co-authored-by: ctiller <ctiller@users.noreply.github.com>
<!--
If you know who should review your pull request, please assign it to
that
person, otherwise the pull request would get assigned randomly.
If your pull request is for a specific language, please add the
appropriate
lang label.
-->
It is reported in https://github.com/grpc/grpc/issues/32356 that there
is a race on vptr for `UnimplementedAsyncRequest` which would cause
crashes for multi-threaded server if clients send unimplemented RPC
request to the server.
The cause is that the server requests a call for
`UnimplementedAsyncRequest` in its base class `GenericAsyncRequest` when
the `vptr` still points to the base class's `vtable`. If the call went
in and another server thread picks up the tag before the `vptr` points
back to the derived class's `vtable`, it would call the wrong virtual
function and also this is a data race. This fix makes the request of the
call inside the derived class's constructor.
<!--
If you know who should review your pull request, please assign it to
that
person, otherwise the pull request would get assigned randomly.
If your pull request is for a specific language, please add the
appropriate
lang label.
-->
<!--
If you know who should review your pull request, please assign it to
that
person, otherwise the pull request would get assigned randomly.
If your pull request is for a specific language, please add the
appropriate
lang label.
-->
PR #32215 added the verified root cert subject to the lower level
`tsi_peer`. This PR is a companion to that and completes the feature by
bubbling the information up to the `TsiCustomVerificationCheckRequest`
which is part of the user facing API for implementing custom
verification callbacks.
Third try for #32466.
This adds an interop client / server for GCP Observability integration
testing.
Everything is new here with no refactor. Plan is to get this in first
before trying to refactor out the flags.
Refactor C++ interop test client flags into the common
`client_helper.h/cc`. This is needed by the observability testing PR
#32466
We need the `ABSL_DECLARE_FLAG` in the header file so that we can share
that across different implementation.
<!--
If you know who should review your pull request, please assign it to
that
person, otherwise the pull request would get assigned randomly.
If your pull request is for a specific language, please add the
appropriate
lang label.
-->
This filter was originally written only for the C++ wrapped layer, but
we have plans to use this for Python (and maybe other wrapped languages
too in the future.)
<!--
If you know who should review your pull request, please assign it to
that
person, otherwise the pull request would get assigned randomly.
If your pull request is for a specific language, please add the
appropriate
lang label.
-->
<!--
If you know who should review your pull request, please assign it to
that
person, otherwise the pull request would get assigned randomly.
If your pull request is for a specific language, please add the
appropriate
lang label.
-->
For stats, the StackDriver/OpenCensus API allows setting the
MonitoredResource directly, so use that.
For tracing, there is no explicit MonitoredResource to use, so just
insert it into the attributes for a span.
<!--
If you know who should review your pull request, please assign it to
that
person, otherwise the pull request would get assigned randomly.
If your pull request is for a specific language, please add the
appropriate
lang label.
-->
This test is flaky only with iomgr, this fix will likely fix this.
<!--
If you know who should review your pull request, please assign it to
that
person, otherwise the pull request would get assigned randomly.
If your pull request is for a specific language, please add the
appropriate
lang label.
-->
The `XdsFaultInjectionMaxFault` test has seen a few flakes since #32326
was merged. I believe the flakiness is caused by the fact that when a
large number of RPCs are queued up before the resolver result comes in,
those RPCs are now re-processed in parallel instead of sequentially,
which can cause us to delay more RPCs than we should due to the
`max_faults` setting. To fix this, we change the test to ensure that the
channel is connected (i.e., the resolver result has already been
returned) before we start sending a large number of concurrent RPCs.
Although this is the only test that I've seen flakes in, I've made this
same change consistently to all fault injection tests that are creating
a large number of concurrent RPCs, since the same flake could affect any
of them.
This code is not plumbed through yet, but it provides the core
infrastructure needed to detect the proper GCP environment resources
needed to set up the labels/attributes/resources for stats, tracing and
logging.
Details on how the various environment resources are setup has been
derived by looking at java's cloud logging library and OpenTelemetry's
future plans. (Could be better explained in an offline review since some
links are internal).
Requesting @veblush for a full review and @markdroth for a structural
review.
<!--
If you know who should review your pull request, please assign it to
that
person, otherwise the pull request would get assigned randomly.
If your pull request is for a specific language, please add the
appropriate
lang label.
-->
Cleanup and remove ios cpp test cronet
To test manually:
./tools/bazel test //src/objective-c/tests:CppCronetTests
@sampajano
<!--
If you know who should review your pull request, please assign it to
that
person, otherwise the pull request would get assigned randomly.
If your pull request is for a specific language, please add the
appropriate
lang label.
-->
The pooled allocator currently has an ABA issue in the allocation path.
This change should fix that - algorithm is described reasonably well in
the PR.
<!--
If you know who should review your pull request, please assign it to
that
person, otherwise the pull request would get assigned randomly.
If your pull request is for a specific language, please add the
appropriate
lang label.
-->
These tests are failing because they're running with too few threads,
however if we give them sufficient threads to catch bugs they're flaky.
Remove them and get the team some bandwidth back.
* Revert "Revert "Revert "Revert "server: introduce ServerMetricRecorder API and move per-call reporting from a C++ interceptor to a C-core filter (#32106)" (#32272)" (#32279)" (#32293)"
This reverts commit 1f960697c5.
* Do not create CallMetricRecorder if call is null.
* Revert "Revert "server: introduce ServerMetricRecorder API and move per-call reporting from a C++ interceptor to a C-core filter (#32106)" (#32272)"
This reverts commit deb1e25543.
* Fix by caching call metric recording stuff in async request
PR #32106 caused msan errors in some tests while de-referencing the
server object where async calls are active after the server is
destroyed. Instead cache the ServerMetricRecorder pointer.
* copyright headers fixed
* clang fixes.
There was a ~1% flake in grpclb end2end tests that was reproducible in opt builds, manifesting as a hang, usually in a the SingleBalancerTest.Fallback test. Through experimentation, I found that by skipping the death test in the grpclb end2end test suite, the hang was no longer reproducible in 10,000 runs. Similarly, moving this test to the end of the suite, or making it run first (as is the case in this PR) resulted in 0 failures in 3000 runs.
It's unclear to me yet why the death test causes things to be unstable in this way. It's clear from the logs that one test does affect the rest, grpc_init is done once for all tests, so all tests utilize the same EventEngine ... until the death test completes, and a new EventEngine is created for the next test.
I think this death test is sufficiently artificial that it's fine to change the test ordering itself, and ignore the wonky intermediate state that results from it.
Reproducing the flake:
```
tools/bazel --bazelrc=tools/remote_build/linux.bazelrc test \
-c opt \
--test_env=GRPC_TRACE=event_engine \
--runs_per_test=5000 \
--test_output=summary \
test/cpp/end2end/grpclb_end2end_test@poller=epoll1
```
Yash Tibrewal