Background
==========
LoadBalancer needs to track RPC measurements and status for
load-reporting. We need to introduce a "Tracer" API for that.
Since such API is very close to the current
Census(instrumentation)-based stats reporting mechanism in terms of what
are recorded, we will migrate the Census-based stats reporting under the
new Tracer API.
Alternatives
============
We considered plumbing the LB-related information from the LoadBalancer
to the core, and recording those information along with the currently
recorded stats to Census. The LB-related information, such as LB_ID,
reason for dropping reqeusts etc, would be added to the Census
StatsContext as tags.
Since tags are held by StatsContext before eventually being recorded by
providing the measurements, and StatsContext is immutable, this would
require a way for LoadBalancer to override the StatsContext, which means
LoadBalancer API would has direct reference to the Census StatsContext.
This is undesirable because Census API is not stable yet.
Part of the LB-related information is whether the client has received
the initial headers from the server. While such information can be
grabbed by implementing a ClientInterceptor, it must be recorded along
with other information such as LB_ID to be useful, and LB_ID is only
available in GrpclbLoadBalancer.
Bottom line, trying to use solely the Census StatsContext API to record
LB load information would require extra data plumbing channel between
ClientInterceptor, LoadBalancer and the gRPC core, as well as exposing
Census API on the gRPC API. Even with those extensive changes, we are
yet to find a working solution. Therefore, we abandoned this idea and
propose this PR.
Summary of changes
==================
API summary
-----------
Introduce "StreamTracer" API, a callback interface for receiving stats
and tracing related updates concerning **a single stream**.
"ClientStreamTracer" and "ServerStreamTracer" add side-specific
events. A stream can have zero or more tracers and report to all of
them.
On the client-side, CallOptions now takes a list of
ClientStreamTracer.Factory. Opon creating a ClientStream, each of the
factory creates a ClientStreamTracer for the stream. This allows
ClientInterceptors to install its own tracer factories by overriding the
CallOptions.
Since StreamTracer only tracks the span of a stream, tracking of a
ClientCall needs to be done in a ClientInterceptor. By installing its
own StreamTracer when a ClientCall is created, ClientInterceptor can
associate the updates for a Call with the updates for the Streams
created for that Call. This is how we keep the existing Census
reporting mechanism in CensusStreamTracerModule.
On the server-side, ServerStreamTracer.Factory is added through the
ServerBuilder, and is used to create ServerStreamTracers for every
ServerStream.
The Tracer API supports propagation of stats/tracing information through
Context and metadata. Both client-side and server-side tracer factories
have access to the headers object. Client-side tracer relies on
interceptor to read the Context, while server-side tracer has
filterContext() method that can override the Context.
Implementation details
----------------------
Only real streams report stats. Pseudo streams such as delayed stream,
failing stream don't report. InProcess transport streams currently
don't report stats.
"StatsTraceContext" which used to receive updates from core and report
directly to Census (StatsContext), now delegates to the StreamTracers of
a stream. On the client-side, the scope of a StatsTraceContext reduces
from ClientCall to a ClientStream to match the scope of StreamTracer.
The Census-specific logic that was in StatsTraceContext is moved into
CensusStreamTracerModule, which produces factories for StreamTracers
that report to Census.
Reporting with StatsTraceContext is moved out of the Channel/Call layer
into Transport/Stream layer, to match the scope change of
StatsTraceContext.
Bug fixed
----------------
The end of a server-side call was reported in ServerCallImpl's
ServerStreamListenerImpl.closed(), which was wrong. Because closed()
receiving OK doesn't necessarily mean the RPC ended with OK. Instead it
means the server has successfully sent the final status, which may be
non-OK, to the client.
Now the end report is done in both ServerStream.close(any Status) and
before calling ServerStreamListener.closed(non-OK). Whichever happens
first is the reported status.
TODOs
=====
A follow-up change to the LoadBalancer API will add a
ClientStreamTracer.Factory to the PickResult to complete the API needed
by load-reporting.
This removes a needless warning, and isn't much slower. Also this
includes a benchmark for StatsTraceContext to measure the overhead
for creation. It adds about 40ns per RPC. Optimization will come
after structural changes are made to break the dependency on
Census.
This appears to have been broken by 3df1446 (which was reverted and
later rolled forward again in 66ab956).
Without this fix, the ServerServiceDefinition.Builder realizes that a
method is registered that isn't in the ServiceDescriptor. Swapping to a
different constructor causes the builder to generate the
ServiceDescriptor for us.
java.lang.IllegalStateException: No entry in descriptor matching bound method E6Cq77iKGNKVCGyVOqq8DqEazX9AcBdPNoMj86c3I5zo4Tv77U/vLe7QS7mhUfaooN7eYdBW7gd9oyV.kc9I0zJumfuUbhyb7SR1u
at io.grpc.ServerServiceDefinition$Builder.build(ServerServiceDefinition.java:164)
at io.grpc.benchmarks.netty.HandlerRegistryBenchmark.setup(HandlerRegistryBenchmark.java:107)
ErrorProne provides static analysis for common issues, including
misused variables GuardedBy locks.
This increases build time by 60% for parallel builds and 30% for
non-parallel, so I've provided a way to disable the check. It is on by
default though and will be run in our CI environments.
Futures almost universally should be handled in some way when being
returned, either to receive the value or to cancel scheduled tasks to
prevent leaks.
Netty is a bit of a special case though, since it constantly returns
futures that you ignore (even adding a listener returns the "this"
future). So we want to suppress the warning for code using Netty instead
of trying to fix it. When we enable ErrorProne in the build, we should
start passing -Xep:FutureReturnValueIgnored:OFF in the compilerArgs.
This is needed because in interceptor tests, often the types cannot
be changed. The void methods stay for users who are writing tests
where they actually don't care about types. The noop methods
require types to be specified. This is for users who don't care
about the implementation. These represent different levels of
commitment.
This eases the transition of code Mocking MethodDescriptor, which
breaks in this release.
If a LoadBalancer2 is passed in, the builder will create ManagedChannelImpl2 instead of ManagedChannelImpl. This allows us to test the LBv2 classes on a large scale.
not necessary to synchronze every time calling
getServiceDescriptor(), if the descriptor has been created already;
go with the double-checked locking idom
Perhaps by accident, each LoadClient channel got its own executor
pool rather than sharing between all of them. The benchmarks are
currently set up with C++ idioms, creating 64 "channels" per
client. C++ uses one thread per channel, while java does not,
leading to a threading model mismatch.
ForkJoinPool is the current executor because it has good contention
characteristics. However, FJP is also very sensitive to the
number of actively running tasks. Too many, and the contention
will go up. Too few, and threads will repeatedly go to sleep and
wake up.
This change basically makes the FJP shared for all clients in the
same process. It doesn't try at all to shut it down properly or
be generally useful; it's just a benchmark. Additionally, this
also groups the Netty and OkHttp specific channel options into
helper functions. This means OkHttp benchmarks will also use
the correct executor now.
A quick test shows QPS going from ~107kqps to 149kqps.
Fixes: #2406
core: adds @Nullable Object getAttachedObject() to ServiceDescriptor
compiler: Plumbing necessary to access proto file descriptors via
the reflection service
This doesn't impact test behavior per-se, but causes it to produce less
useless log output of the form:
java.lang.IllegalStateException: call was half-closed
at com.google.common.base.Preconditions.checkState(Preconditions.java:174)
at io.grpc.internal.ClientCallImpl.sendMessage(ClientCallImpl.java:380)
at io.grpc.stub.ClientCalls$CallToStreamObserverAdapter.onNext(ClientCalls.java:299)
at io.grpc.benchmarks.driver.LoadClient$AsyncPingPongWorker$1.onNext(LoadClient.java:406)
at io.grpc.benchmarks.driver.LoadClient$AsyncPingPongWorker$1.onNext(LoadClient.java:400)
at io.grpc.stub.ClientCalls$StreamObserverToCallListenerAdapter.onMessage(ClientCalls.java:382)
at io.grpc.internal.ClientCallImpl$ClientStreamListenerImpl$1MessageRead.runInContext(ClientCallImpl.java:473)
... 7 more
Fixes#2372
The DefaultHttp2Headers class is a general-purpose Http2Headers implementation
and provides much more functionality than we need in gRPC. In gRPC, when reading
headers off the wire, we only inspect a handful of them, before converting to
Metadata.
This commit introduces a Http2Headers implementation that aims for insertion
efficiency, a low memory footprint and fast conversion to Metadata.
- Header names and values are stored in plain byte[].
- Insertion is O(1), while lookup is now O(n).
- Binary header values are base64 decoded as they are inserted.
- The byte[][] returned by namesAndValues() can directly be used to construct
a new Metadata object.
- For HTTP/2 request headers, the pseudo headers are no longer carried over to
Metadata.
A microbenchmark aiming to replicate the usage of Http2Headers in NettyClientHandler
and NettyServerHandler shows decent throughput gains when compared to DefaultHttp2Headers.
Benchmark Mode Cnt Score Error Units
InboundHeadersBenchmark.defaultHeaders_clientHandler avgt 10 283.830 ± 4.063 ns/op
InboundHeadersBenchmark.defaultHeaders_serverHandler avgt 10 1179.975 ± 21.810 ns/op
InboundHeadersBenchmark.grpcHeaders_clientHandler avgt 10 190.108 ± 3.510 ns/op
InboundHeadersBenchmark.grpcHeaders_serverHandler avgt 10 561.426 ± 9.079 ns/op
Additionally, the memory footprint is reduced by more than 50%!
gRPC Request Headers: 864 bytes
Netty Request Headers: 1728 bytes
gRPC Response Headers: 216 bytes
Netty Response Headers: 528 bytes
Furthermore, this change does most of the gRPC groundwork necessary to be able
to cache higher ordered objects in HPACK's dynamic table, as discussed in [1].
[1] https://github.com/grpc/grpc-java/issues/2217
MessageFramer calls Drainable.drainTo with a special output stream of
OutputStreamAdapter. Currently, ByteBufInputStream writes to this output
stream by allocating a heapBuffer in UnsafeByteBufUtil.getBytes, copying
from the direct byte buffer of BBIS, and then copies to the direct byte
buffer from MessageFramer.writeRaw().
This change is an easy way to cut down on wasted memory, even though
ideally there would be some way to have less copies. The actual data is
only around 10 bytes, but causes O(10)s of megabytes allocation for the
heap pool.
For #2062
The benchmarks today do not have a good way to record metrics with precision
or shutdown safely when the benchmark is over. This change alters the
AbstractBenchmark class to return a latch that can be waited upon when ending
the benchmark.
Benchmarks also would accidentally request way too many messages from the
server by calling request(1) explicitly in addition to the implicit one
in the StreamObserver to Call adapter. This change adds a few outstanding
requests, but otherwise keeps the request count bounded.
Additionally, benchmark calls would ignore errors, and just shutdown in such
cases. This changes them to log the error and just wait for the benchmark to
complete. In the successful case, the benchmark client notifies server by
halfClosing (via onCompleted) where it previously did not. It is also
careful to only do this once.
Lastly, Benchmarks have been changes to enable and disable recording at exact
points in the benchmark method, rather than waiting for teardown to occur.
Also, recording begins inside the recording method, not in Setup. JMH may
do other procressing before, between, and after iterations.
partially resolving #1469
The added option for java_plugin `enable_deprecated` is `true` by default in `java_plugin.cpp`, so the generated code for `TestService.java` (`compiler/build.gradle` not setting this option) has all deprecated interfaces and static bindService method.
`./build.gradle` and `examples/build.gradle` set this option explicitly to `false`, so all the other generated classes do not have deprecated code.
Will set `enable_deprecated` to `false` by default in future PR when we are ready.
first step to address issue #1469:
- leave and deprecate interfaces in codegen
- introduce `ServiceImplBase`,
- `AbstractService` is deprecated and extends `ServiceImplBase`
- static `bindService()` is deprecated
Resolves#1756
The thread-unsafe method `io.grpc.testing.TestUtils.pickUnusedPort` causes flakes (#1756) in windows. Need to avoid use of this method in test as in windows the tests are running in different jvms and concurrent calls of this method in multiple processes tend to return the same port number.
There are some usages of this method in benchmarks, so moved the method to `io.grpc.benchmarks.Utils` and the method will only be used in benchmarks and not in test.
This allows us to play with zero-copy and proto3 support for lite.
Unfortunately, it introduced some warnings, so deprecated warnings are
now ignored for benchmarks and interop-testing.
This reverts commit 3df1446deb.
The commit was adding to the difficulty of integration for testing. By
itself it isn't bad, so this is a temporary revert until the many other
commits are absorbed and then it will be reapplied.
This does have a manual edit for ClientCallsTest.
This removes a reference of io.grpc.internal from io.grpc. It also
optimizes server call handling for outbound grpc-accept-encoding header,
as had already been done for client call.
This now catches a few more places we needed -Xlint:-options.
InProcessSocketAddress is technically already in our stable API, so I
maintained its current serialVersionUID.
See #933
- Create InternalHandlerRegistry, an immutable look-up table. Handlers
passed to ServerBuilder.addService() go to this registry. This covers
the most common use cases. By keeping the registry internal we could
freely change the registry's interface to accommodate optimizations,
e.g., for hpack.
- The internal registry uses a flat fullMethodName -> handler look-up
table instead of a hierarchical one used before. It faster because it
saves one look-up and a substring.
- Introduces the fallback registry, settable by
ServerBuilder.fallbackHandlerRegistry(), for advanced users who want a
dynamic registry. Moved the current MutableHandlerRegistryImpl to
io.grpc.util.MutableHandlerRegistry as a stock implementation of the
fallback registry. The io.grpc.MutableHandlerRegistry interface is now
removed.
This reverts commit 8825f355df.
The commit changed the package name of services that were used across
languages. That broke their functionality pretty severely. The changes
require more coordination with others.
Resolves#1221
Add ClientCall.cancel(String, Throwable) and deprecate
ClientCall.cancel(). Will delete cancel() after all known third-party
overriders have switched to overriding the new one.
Carl Mastrangelo
Łukasz Strzałkowski
Kun Zhang
Eric Anderson
Kenji Kaneda
ZHANG Dapeng
Lukasz Strzalkowski
Carl Mastrangelo
Eric Gribkoff
Jakob Buchgraber
Louis Ryan
nmittler
Kun Zhang
Jan Tattermusch