Andrey Yarovoy created HDDS-16946:
-------------------------------------
Summary: Remove the per-call InfoBucket from rooted-ofs
isFileClosed and recoverLease
Key: HDDS-16946
URL: https://issues.apache.org/jira/browse/HDDS-16946
Project: Apache Ozone
Issue Type: Improvement
Components: OM
Reporter: Andrey Yarovoy
`BasicRootedOzoneClientAdapterImpl.isFileClosed` and `recoverFilePrepare` call
`getBucket()` before their real RPC. That is an `InfoBucket` (plus `InfoVolume`
+ `InfoBucket` per link hop). The fetch does one job:
`OzoneFSUtils.validateBucketLayout`, which throws `IllegalArgumentException`
for OBJECT_STORE. Neither
site reads anything else from the returned `OzoneBucket`, apart from volume and
bucket names it already
has. These are the hsync paths: HBase calls `isFileClosed` while it waits for
lease recovery, and `recoverLease` on region server failover.
* **isFileClosed** issues `GetFileStatus` (`OzoneBucket.getFileStatus` →
`ClientProtocol.getOzoneFileStatus`,
headOp=false). `OmMetadataReader.getFileStatus` has rejected OBJECT_STORE with
`NOT_SUPPORTED_OPERATION`
unconditionally since HDDS-15925
(`OzoneManagerVersion.GET_FILE_STATUS_REJECTS_OBS`, 14). There, the
layout comes from `ResolvedBucket`, so links are resolved.
* **recoverLease** issues `RecoverLease`.
`OzoneManagerRatisUtils.createClientRequest` looks up the
link-resolved layout and rejects every non-FSO bucket (LEGACY too) before
Ratis, with
"Bucket ... is not FSO layout. It does not support lease recovery". This check
dates from HDDS-7769
(2023). `RpcClient.recoverLease` refuses to send the RPC at all unless the OM
is at least
`HBASE_SUPPORT` (7), which was added in 2024. So every OM that can receive this
RPC already has the check.
### Proposed change
*isFileClosed.* When `proxy.getOmVersion() >= GET_FILE_STATUS_REJECTS_OBS`, call
`proxy.getOzoneFileStatus(volume, bucket, key, false)` directly, the same way
HDDS-15925 changed
`getFileStatus`. Map `NOT_SUPPORTED_OPERATION` to `IllegalArgumentException`.
Against an older OM keep the
current `getBucket()` path.
Keep the `isSnapshotPath()` check before the RPC. Today it runs after
`getBucket()`, so a snapshot path in
a missing bucket reports `FileNotFoundException`, and after the change it
reports "file is in a
snapshot". Either keep that order (cheap: the snapshot path check is local) or
call the difference out
in the PR.
*recoverLease.* Drop the `getBucket()` and pass `ofsPath.getVolumeName()` /
`getBucketName()`. Those are
the names `bucket.getVolumeName()` / `getName()` return today, because the link
handle carries the
link's own names.
The one behaviour change is the exception type for an OBJECT_STORE bucket.
Today the client throws
`IllegalArgumentException` before any RPC. Without the fetch, OM's plain
`IOException` reaches the
client as `OMException(INTERNAL_ERROR)`. To keep the exception type unchanged,
follow the HDDS-16760
pattern:
* OM throws `OMException(NOT_SUPPORTED_OPERATION)` instead of `IOException`,
with the same message.
* Add a new `OzoneManagerVersion` constant, placed after
`LIST_STATUS_REJECTS_OBS(101)`. The number is
assigned at merge; coordinate it with sibling Jiras.
* At or above that version the client skips `getBucket()` and maps
`NOT_SUPPORTED_OPERATION` to
`IllegalArgumentException`. Below it the client keeps `getBucket()`.
LEGACY buckets behave as before: the client never rejected them, and OM rejects
them under either code.
### Compatibility
| | isFileClosed | recoverLease |
|---|---|---|
| New client, OM ≥ new constant | 1 RPC, OBS → `IllegalArgumentException` | 1
RPC, OBS → `IllegalArgumentException` |
| New client, older OM | 1 RPC for OM ≥ 14; `getBucket()` path below |
`getBucket()` path, unchanged |
| Old client, new OM | unchanged | unchanged, except LEGACY now gets
`NOT_SUPPORTED_OPERATION` instead of `INTERNAL_ERROR` |
There is no protobuf change and no new configuration.
Point for review: dropping the fetch drops the BUCKET/READ ACL check that
`getBucketDetails` performed.
`GetFileStatus` and `RecoverLease` keep their own key-level ACL checks, and OM
checks READ on each link
bucket it follows, which is the same trade-off accepted in HDDS-15925 and
HDDS-16760.
## Testing
* Adapter (Mockito partial mock, like
`TestBasicRootedOzoneClientAdapterHeadOp`):
* Both operations issue exactly one OM call and no `getBucket()`.
*`NOT_SUPPORTED_OPERATION` → `IllegalArgumentException`.
*`FILE_NOT_FOUND` / `KEY_NOT_FOUND` / `BUCKET_NOT_FOUND` keep today's
`FileNotFoundException` mapping.
* An older OM falls back to `getBucket()`, including an OM that reports `ZDU`
for recoverLease.
* OM: `TestOMRecoverLeaseRequest` (or the `createClientRequest` test) asserts
`NOT_SUPPORTED_OPERATION`
for OBJECT_STORE and LEGACY, and success for FSO.
* `AbstractRootedOzoneFileSystemTest`: `isFileClosed` / `recoverLease` on an
OBJECT_STORE bucket still fail
with `IllegalArgumentException`. Assert the RPC-count property in the adapter
test, not by OMMetrics
deltas. OMMetrics is cluster-global and the trash emptier moves it (HDDS-16513).
* Benchmark: the cost of these calls is too low and they are too rare to be
worth a dedicated
benchmark. The RPC count goes from 2 to 1 per call, and the adapter test
asserts it exactly.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
