HDDS-10784. Multipart upload to encrypted bucket fails with ClassCastException

[adoroszlai]

What changes were proposed in this pull request?

Fix exception:

ClassCastException: class org.apache.hadoop.crypto.CryptoOutputStream cannot be cast to class org.apache.hadoop.ozone.client.io.KeyMetadataAware ...
  at org.apache.hadoop.ozone.s3.endpoint.ObjectEndpoint.createMultipartKey(ObjectEndpoint.java:1003)
  at org.apache.hadoop.ozone.s3.endpoint.ObjectEndpoint.put(ObjectEndpoint.java:239)

Steps to reproduce:

0. Start cluster with Ratis streaming disabled (default setting)
1. Create encrypted bucket in S3 volume
2. Upload key via multipart upload to encrypted bucket

https://issues.apache.org/jira/browse/HDDS-10784

How was this patch tested?

Added acceptance test coverage for S3 with Ratis streaming disabled in encrypted bucket in ozonesecure environment. ozonesecure-ha covers the same with Ratis streaming enabled.

CI:
https://github.com/adoroszlai/ozone/actions/runs/8920571448

[ivandika3]

Thank you for the fix. LGTM +1.

I also saw some other explicit casting instances in ObjectEndpointStreaming:

• ObjectEndpointStreaming#putKeyWithStream
• ObjectEndpointStreaming#copyKeyWithStream
• ObjectEndpointStreaming#createMultipartKey

The wrapping for createStreamKey and createMultipartStreamKey:

• OzoneDataStreamOutput
  • KeyDataStreamOutput (non-secure case)
  • OzoneOutputStream (secure case)
    • CryptoOutputStream (FileEncryptionInfo specified)
      • KeyDataStreamOutput
    • OzoneOutputStream (FileEncryptionInfo not specified, but GDPR enabled)
      • CipherOutputStreamOzone
        • KeyDataStreamOutput

I think should be fine to leave it be since these cases did not cast CryptoOutputStream to KeyMetadataAware directly. However, I think we can remove the unnecessary KeyMetadataAware casting from these methods, but I'm also fine to leave it as it is.

Another possible solution might be to extend CryptoOutputStream to support the KeyMetadataAware and delegate it to the wrapped stream, but it might entails a lot of change.

[adoroszlai]

Thanks @ivandika3 for the review.

I also saw some other explicit casting instances in ObjectEndpointStreaming:

Thanks for pointing those out. I'll remove them as part of HDDS-10791.

[hemantk-12]

Thanks @adoroszlai for the quick fix.

LGTM+1.

[adoroszlai]

Thanks @hemantk-12, @ivandika3 for the review.