Sergey Soldatov created HDDS-15998:
--------------------------------------
Summary: Adopt Arrow Flight ZeroCopy grpc approach to readBlock
Key: HDDS-15998
URL: https://issues.apache.org/jira/browse/HDDS-15998
Project: Apache Ozone
Issue Type: Improvement
Components: Ozone Client, server-framework
Affects Versions: 2.3.0
Reporter: Sergey Soldatov
Assignee: Sergey Soldatov
Adds a zero-copy streaming read path for block/chunk data served over gRPC
between the datanode and the client, eliminating the redundant heap allocations
that the stock gRPC/protobuf path incurs on every large payload.
Today a streaming ReadBlock/ReadChunk response materializes the full payload as
a heap ByteString inside a ContainerCommandResponseProto on both the serialize
(datanode → wire) and parse (wire → client) hops - for MB-sized payloads at
high message rates this becomes GC pressure and throughput cost that scales
with payload size rather than useful work.
The optimization borrows the "sidecar" technique proven in Apache Arrow Flight:
the opaque block bytes are kept at the highest protobuf field tag so they land
last on the wire, and custom gRPC marshallers hand-encode only the small
envelope fields (cmdType, traceID, result, checksum, offset) while carrying the
payload in a pooled direct Netty ByteBuf that is passed by reference instead of
copied.
The wire format stays byte-identical to stock protobuf in every flag
combination, so any client interoperates with any datanode.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
