Andrey Yarovoy created HDDS-17715:
-------------------------------------
Summary: Avoid Enum.valueOf round trips in OzoneAcl
toProtobuf/fromProtobuf
Key: HDDS-17715
URL: https://issues.apache.org/jira/browse/HDDS-17715
Project: Apache Ozone
Issue Type: Improvement
Components: OM, Ozone Client
Reporter: Andrey Yarovoy
Assignee: Andrey Yarovoy
*Description*
`OzoneAcl.toProtobuf` and `OzoneAcl.fromProtobuf`
(`hadoop-ozone/common/.../OzoneAcl.java`) map between the Java ACL enums and
their
protobuf counterparts through a name round trip:
{code:java}
.setType(OzoneAclType.valueOf(acl.getType().name()))
.setAclScope(OzoneAclScope.valueOf(acl.getAclScope().name()))
...
new OzoneAcl(ACLIdentityType.valueOf(protoAcl.getType().name()),
protoAcl.getName(),
AclScope.valueOf(protoAcl.getAclScope().name()), aclRights);
{code}
That is `name()` + `Enum.valueOf` (a `HashMap` lookup in
`Enum.enumConstantDirectory`) twice per ACL entry, in both directions. It runs
for every
ACL of every volume, bucket and key that OM serializes (getKeyInfo,
getFileStatus, getBucketInfo, list*), and again on the client when parsing.
Evidence:
- OM: the ACL-conversion frames showed up in the OM read-path jstack censusas
the one self-contained Ozone-code item.
HDDS-16289's projection removed the ACL rebuild from intermediate FSO path
segments. The final-segment key and every response still pay it.
- RS client (HBase YCSB on EC): `OzoneAcl.fromProtobuf` is ~0.1% of RS CPU, of
which `valueOf` is ~0.06%. **The client-side win is small; the value is OM CPU
per read RPC.**
*Proposed fix*
Four ordinal-indexed lookup arrays, built once in a `static {}` block. Iterate
the **Java** enums' `values()` and resolve each counterpart
once with the existing `name()`/`valueOf` contract:
{code:java}
private static final OzoneAclType[] TYPE_TO_PROTO = new
OzoneAclType[ACLIdentityType.values().length];
private static final ACLIdentityType[] TYPE_FROM_PROTO = ...; // indexed by
proto ordinal
private static final OzoneAclScope[] SCOPE_TO_PROTO = ...;
private static final AclScope[] SCOPE_FROM_PROTO = ...;
// call sites: TYPE_TO_PROTO[acl.getType().ordinal()],
TYPE_FROM_PROTO[protoAcl.getType().ordinal()], ...
{code}
- Iterate the Java enums, not the protobuf ones, so the protobuf `UNRECOGNIZED`
sentinel never enters the tables. The proto file is proto2,
so ordinals are stable.
- Behavior-preserving, including a class-load-time failure if a name ever
diverges.
- `parseAcl` (CLI/config string parsing) stays on `valueOf`: it handles
arbitrary user input and is cold.
*Compatibility / side effects*
None. No wire or API change.
*Tests*
- `TestOzoneAcls`, `TestOzoneAclUtil`, `TestOzoneObjInfo` round-trip tests.
- New unit test: every `ACLIdentityType`/`AclScope` value round-trips through
to/fromProtobuf to the same constant.
- Optional: JMH micro-benchmark of `toProtobuf`/`fromProtobuf` for a 3-entry
ACL list, plus OM CPU per `getKeyInfo` in the existing OM read benchmark.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
