Add support for anonymous in-process servers. #8589

markb74 · 2021-10-10T18:49:21Z

Anonymous servers aren't registered statically, meaning they can't be referenced by name.
Only the InProcessSocketAddress, fetched via Server.getListenSockets()
can be used to connect to the server.

We also support the creation of in-process channels via forTarget(String), which
is particularly useful for production Android usage of in-process servers,
where process startup latency is crucial. A custom name resolver can be
used to create the server instance on demand without directly impacting
the startup latency of in-process gRPC clients.

This supports a more-standard approach to "OnDeviceServer" referenced in gRFC L73.
https://github.com/grpc/proposal/blob/master/L73-java-binderchannel.md#ondeviceserver

markb74 · 2021-10-10T18:50:23Z

core/src/main/java/io/grpc/inprocess/InProcessServer.java

    }
  }

  @Override
  public SocketAddress getListenSocketAddress() {
-    return new InProcessSocketAddress(name);
+    return new InProcessSocketAddress(name, anonymous ? this : null);


This choice is notable. The address instance will only reference the server if it's anonymous.

Yes, I think that is the correct behavior.

At first I was like, "oh, definitely," then I realized "but that might produce weird behavior/semantics" because depending on how the address was created you'd get different behavior. But I actually think it works well. Essentially it is the same as the anonymous server having a randomly-generated ID that is never reused.

Anonymous servers aren't registered statically, meaning they can't be referenced by name. Only the InProcessSocketAddress, fetched via Server.getListenSockets() can be used to connect to the server. This is particularly useful for production Android usage of in-process servers, where process startup latency is crucial, since a custom name resolver can be used to create the server instance on demand without directly impacting the startup latency of in-process gRPC clients. This approach supports a more-standard approach to "OnDeviceServer" referenced in gRFC L73. https://github.com/grpc/proposal/blob/master/L73-java-binderchannel.md#ondeviceserver

Tests require it.

ejona86 · 2021-10-13T22:05:00Z

This is particularly useful for production Android usage of in-process servers,
where process startup latency is crucial, since a custom name resolver can be
used to create the server instance on demand without directly impacting
the startup latency of in-process gRPC clients.

That's technically possible without anonymous in-process servers. That just requires InProcessChannelBuilder.forTarget().

ejona86

Overall, looks good and is surprisingly minor tweaks. The SocketAddress identity thing though is an important bug to resolve.

ejona86 · 2021-10-13T22:08:14Z

core/src/main/java/io/grpc/inprocess/InProcessServer.java

    }
  }

  @Override
  public SocketAddress getListenSocketAddress() {
-    return new InProcessSocketAddress(name);
+    return new InProcessSocketAddress(name, anonymous ? this : null);


Yes, I think that is the correct behavior.

At first I was like, "oh, definitely," then I realized "but that might produce weird behavior/semantics" because depending on how the address was created you'd get different behavior. But I actually think it works well. Essentially it is the same as the anonymous server having a randomly-generated ID that is never reused.

core/src/main/java/io/grpc/inprocess/InProcessServerBuilder.java

core/src/main/java/io/grpc/inprocess/InProcessServer.java

markb74 · 2021-10-14T12:55:16Z

This is particularly useful for production Android usage of in-process servers,
where process startup latency is crucial, since a custom name resolver can be
used to create the server instance on demand without directly impacting
the startup latency of in-process gRPC clients.

That's technically possible without anonymous in-process servers. That just requires InProcessChannelBuilder.forTarget().

Good point, I'd meant to mention that change explicitly. Updated.

ejona86

Looks pretty good, I have two concerns:

Always using "anonymous" in the InProcessSocketAddress. If I use InProcessServer.forName("anonymous") the two socket addresses will look identical, but won't be equal/interchangable. Essentially, there's no reliable way to learn the address is anonymous.
InProcessServer.anonymous() seems slightly weak, in that it isn't obvious it would create a new builder instance. We may want newAnonymous() or the like. I've added this to the API review schedule to discuss.

ejona86 · 2021-10-14T21:59:13Z

API review:

No client-side concerns
InProcessServer.anonymous()
- forAnonymous()
- newAnonymous()
- forAnonymousServer()
InProcessSocketAddress.getName()
- “Anonymous”
  - Could disallow forName(“anonymous”)
- “unlikely.to.be.used.anonymous”
  - Could disallow forName(“unlikely.to.be.used.anonymous”)
  - +.5
- Random string
- Null
  - +4
  - +.5
Some concerns about how the address holds a reference to the server. Discussed how the current implementation is pretty similar to WeakHashMap<InProcessSocketAddress, InProcessServer> (although it can be slow to drop the strong references after the weak ref disappears) and there was mention that such a registry made more sense.

We ran out of time in the meeting. It seems we should maybe schedule a one-off meeting to discuss this, since I sense some resistance to concepts/behavior that make anonymous memory-only; if anonymous servers had to be shut down, it seems that'd simplify the conversation.

Created a new AnonymousInProcessSocketAddress class which is distinct from the existing InProcessSocketAddress. This simplifies the server builder API since we can have forAddress() instead of anonymous(), and removes the need for InProcessSocketAddress to handle anonymous servers, which had some tricy implications like getName() returning null.

markb74 · 2021-10-21T10:13:26Z

core/src/main/java/io/grpc/inprocess/AnonymousInProcessSocketAddress.java

+    }
+    InProcessServer otherServer = ((AnonymousInProcessSocketAddress) obj).getServer();
+    synchronized (this) {
+      return otherServer == server;


Notable choice: I chose to let two instances to be equal if they share the same server instance. There's currently no way that can happen (given InProcessServer's impl), but I figured it was a reasonable choice from this classes perspective.

Since this address is the central registration location, this should use instance equality of the address. So we shouldn't have equals/hashCode at all here. We also don't want hashCode() to change over time.

Oh good point re: hashcode, I missed that. Done.

markb74 · 2021-10-21T10:17:22Z

core/src/main/java/io/grpc/inprocess/InProcessServer.java

@@ -40,11 +40,16 @@
  private static final ConcurrentMap<String, InProcessServer> registry
      = new ConcurrentHashMap<>();

-  static InProcessServer findServer(String name) {


The possibility of creating a base AbstractInProcessSocketAddress class with package visible findServer(), registerServer(InProcessServer) and unregisterServer(InProcessServer) methods occurred to me.

That would allow this method, and registerInstance() and unregisterInstance() below to just be methods on the address, removing the need for instance checks.

Optimizing this method seems like a micro-optimization. If this method is the only one that benefits, then avoiding public API seems clearly superior. If we think user's code would improve, then common base class is more interesting. Seems we can always do it in the future, so let's favor less API for now.

Your suggestion would also mean moving the registry to InProcessSocketAddress, which propagates the "strangeness" of anonymous socket address. Seems counter to our preferences.

Makes sense. I was attracted to the approach making everything more consistent, even if that does mean "embracing the strangeness" to some degree. But I like philosophy that improving user code is the key goal, so agreed this probably isn't worth it.

core/src/main/java/io/grpc/inprocess/InProcessServer.java

core/src/main/java/io/grpc/inprocess/AnonymousInProcessSocketAddress.java

ejona86 · 2021-10-21T19:41:17Z

core/src/main/java/io/grpc/inprocess/AnonymousInProcessSocketAddress.java

+    }
+    InProcessServer otherServer = ((AnonymousInProcessSocketAddress) obj).getServer();
+    synchronized (this) {
+      return otherServer == server;


Since this address is the central registration location, this should use instance equality of the address. So we shouldn't have equals/hashCode at all here. We also don't want hashCode() to change over time.

ejona86 · 2021-10-21T19:44:30Z

core/src/main/java/io/grpc/inprocess/InProcessServer.java

@@ -40,11 +40,16 @@
  private static final ConcurrentMap<String, InProcessServer> registry
      = new ConcurrentHashMap<>();

-  static InProcessServer findServer(String name) {


Optimizing this method seems like a micro-optimization. If this method is the only one that benefits, then avoiding public API seems clearly superior. If we think user's code would improve, then common base class is more interesting. Seems we can always do it in the future, so let's favor less API for now.

Your suggestion would also mean moving the registry to InProcessSocketAddress, which propagates the "strangeness" of anonymous socket address. Seems counter to our preferences.

core/src/main/java/io/grpc/inprocess/InProcessServer.java

core/src/test/java/io/grpc/inprocess/AnonymousInProcessTransportTest.java

core/src/test/java/io/grpc/inprocess/InProcessTransportTest.java

ejona86 · 2021-10-21T20:03:39Z

API review notes:

To side-step naming and getName() issues, should we make anonymous first-class in the address and have the address power the anonymous feature?
Use a different class AnonymousInProcessSocketAddress extends SocketAddress
- +many (we were already well over time)
Given the above, drop explicit anonymous API from server API. Use forAddress(SocketAddress) on server-side instead
- Mirrors NettyServerBuilder.forAddress()

There was enough agreement for anonymous at this point that everyone felt comfortable with the feature moving forward. We agreed any nits could be handled after this PR went in.

ejona86

Only one comment, and it is probably-good-to-have-but-not-a-blocker.

core/src/main/java/io/grpc/inprocess/InProcessServer.java

markb74 requested a review from ejona86 October 10, 2021 18:49

markb74 commented Oct 10, 2021

View reviewed changes

markb74 force-pushed the anoninprocess branch from ee09e13 to 8cd4bc5 Compare October 10, 2021 20:22

markb74 force-pushed the anoninprocess branch from 8cd4bc5 to babfb2b Compare October 11, 2021 09:17

markb74 added 3 commits October 11, 2021 14:52

Restore overridden forAddress(name, int) method.

28fc525

Tests require it.

Add an additional test for anon in-process servers.

361c9f8

Add tests for changes to InProcessSocketAddress

df89d02

ejona86 reviewed Oct 13, 2021

View reviewed changes

Address review comments.

026113a

ejona86 reviewed Oct 14, 2021

View reviewed changes

markb74 commented Oct 21, 2021

View reviewed changes

markb74 added 3 commits October 21, 2021 11:21

minor fixes

619902f

fix style issue

506b46e

fix style issues in tests

4ac409d

sanjaypujare reviewed Oct 21, 2021

View reviewed changes

core/src/main/java/io/grpc/inprocess/InProcessServer.java Outdated Show resolved Hide resolved

ejona86 reviewed Oct 21, 2021

View reviewed changes

markb74 mentioned this pull request Oct 22, 2021

AnonymousInProcessSocketAddress is experimental. #8626

Open

Address review comments.

4d5a1e3

markb74 requested a review from ejona86 October 25, 2021 15:58

ejona86 approved these changes Oct 25, 2021

View reviewed changes

core/src/main/java/io/grpc/inprocess/InProcessServer.java Show resolved Hide resolved

Additional assertions

fc61151

markb74 merged commit 607362a into grpc:master Oct 25, 2021

github-actions bot locked as resolved and limited conversation to collaborators Jan 24, 2022

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add support for anonymous in-process servers. #8589

Add support for anonymous in-process servers. #8589

markb74 commented Oct 10, 2021 •

edited

markb74 Oct 10, 2021

ejona86 Oct 13, 2021

ejona86 commented Oct 13, 2021

ejona86 left a comment

ejona86 Oct 13, 2021

markb74 commented Oct 14, 2021

ejona86 left a comment •

edited

ejona86 commented Oct 14, 2021

markb74 Oct 21, 2021

ejona86 Oct 21, 2021

markb74 Oct 22, 2021

markb74 Oct 21, 2021 •

edited

ejona86 Oct 21, 2021

markb74 Oct 22, 2021

ejona86 Oct 21, 2021

ejona86 Oct 21, 2021

ejona86 commented Oct 21, 2021 •

edited

ejona86 left a comment

Add support for anonymous in-process servers. #8589

Add support for anonymous in-process servers. #8589

Conversation

markb74 commented Oct 10, 2021 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ejona86 commented Oct 13, 2021

ejona86 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

markb74 commented Oct 14, 2021

ejona86 left a comment • edited

Choose a reason for hiding this comment

ejona86 commented Oct 14, 2021

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

markb74 Oct 21, 2021 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ejona86 commented Oct 21, 2021 • edited

ejona86 left a comment

Choose a reason for hiding this comment

markb74 commented Oct 10, 2021 •

edited

ejona86 left a comment •

edited

markb74 Oct 21, 2021 •

edited

ejona86 commented Oct 21, 2021 •

edited