Chu Cheng Li created HDDS-17462:
-----------------------------------
Summary: Send structured errors from the OM gRPC server instead of
exception class names
Key: HDDS-17462
URL: https://issues.apache.org/jira/browse/HDDS-17462
Project: Apache Ozone
Issue Type: Improvement
Reporter: Chu Cheng Li
h3. Problem
* The OM gRPC server sends an OM exception as {{Status.INTERNAL}} with the
description {{class: message}}. {{GrpcOmTransport#unwrapException}} splits this
text at the first {{:}}. Thus, a change in the text format breaks the client.
* The client creates the OM exception from the class name, with reflection in
{{RemoteException#unwrapRemoteException}}. Thus, the server response selects
the class that the client loads.
* The suggested leader in {{OMNotLeaderException}} does not get to the retry
policy. The client creates the exception with the {{String}} constructor, so
the suggested node ID and address are {{null}}. The Hadoop RPC client has the
same problem.
* The OM sets the suggested leader address with {{InetAddress#toString}}, for
example {{host/10.0.0.1}}. The retry policy compares it with the {{host:port}}
address of the OM proxy. Thus, the address does not match, even when it gets to
the client.
h3. Changes
* Add an {{OMGrpcError}} message to {{OmClientProtocol.proto}}. It has an error
code ({{NOT_LEADER}}, {{LEADER_NOT_READY}}, {{READ_INDEX_FAILED}},
{{READ_FAILED}} or {{INTERNAL}}), the message and the suggested leader node ID.
* {{OzoneManagerServiceGrpc}}: map the OM exception to an {{OMGrpcError}} in
one place, and send it in a binary gRPC trailer. Keep the old description for
old clients.
* {{GrpcOmTransport}}: map the error code to the OM exception in one place. Do
not parse the description, and do not use reflection.
* {{OMFailoverProxyProviderBase}}: accept the OM exception as a direct cause,
not only in a {{RemoteException}}. Find the suggested leader by its node ID.
h3. Compatibility
* An old gRPC client ignores the trailer and reads the description.
* A new gRPC client and an old OM: there is no trailer. The client handles the
error as a plain failure and fails over to the next OM.
The Hadoop RPC client is out of scope.
A first version with exception class names in the trailers was in
https://github.com/apache/ozone/pull/11760/changes/dad4b1d32700a500a6331af70117825d671e5360
(HDDS-17448).
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
