[RATIS-1931] Support Zero-Copy in ratis-grpc

• Type: Improvement
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Component/s: gRPC
• Assignee: Duong
• Reporter: Duong
• Created: 07/Nov/23 01:21
• Updated: 20/May/24 14:07

Description

As GRPC has stabilized the API for zero-copy for protobuf (https://github.com/grpc/grpc-java/issues/7387), it's possible to have such memory efficiency feature in Ratis GRPC.

Issue Links

• is duplicated by: RATIS-1405 Avoid buffer copy with Grpc zero-copy marshaller

Sub-Tasks

• Create zero-copy Marshaller: Resolved
• Add TestGrpcZeroCopy: Resolved
• Restore ZeroCopyMessageMarhaller#popStream: Resolved
• Avoid Ratis internal data from referring to external input protobuf data source: Resolved
• Do not store CommitInfoProto in CommitInfoCache: Resolved
• Refactor client request processing to support reference count: Resolved
• Support Zero-Copy in GrpcClientProtocolService: Resolved
• Refactor appendEntries processing to support reference count: Resolved
• Support Zero-Copy in GrpcServerProtocolService: Resolved
• ReferenceCount should work for all LogEntry types.: Resolved
• Zero-copy buffers are not released: Resolved
• Zero-copy buffers are not released - 2nd chunk: Resolved
• Add metrics assertions to verify all zero-copy messages are released properly: Resolved
• Refactor StateMachine interface to use ReferenceCountedObject: Resolved
• Refactor TransactionContext to supply LogEntryProto via a ReferenceCountedObject: Resolved
• Refactor RaftLog to supply log as ReferenceCountedObject: Resolved
• LogAppender to consume log entries with reference count: Resolved
• Allow StateMachine.read to return a ReferentCountedObject: Resolved
• Missing reference count when putting log entries to cache on follower: Resolved
• Timedout StateMachine retainRead is released twice.: Resolved
• Support Zero-Copy in loading log entries from local files: Open
• Benchmark zero-copy in ozone: Open
• Test failure because SimpleStateMachine4Testing is not closed properly: Open
• Add metrics to expose number of zero-copy unclosed messages: Resolved
• Decouple metadata and configuration entries from appendEntries buffer for stateMachineCache: Resolved
• TransactionContext's stateMachineLogEntry and stateMachineContext may cause corruption: Resolved
• Add a conf to enable/disable zero copy.: Resolved
• Corruption due to SegmentedRaftLogWorker queue LogEntry without reference counter: Resolved