https://issues.apache.org/jira/browse/ARROW-16131

⚠️ Ticket has not been started in JIRA, please click 'Start Progress'.

I checked out all the build failures reported , including arrow-gcsfs-test (Timeout) and arrow-s3-test (Failed), and they don't seem to be related with my change.

I can build and run the S3 test locally, and it ran into segfault locally as the CI server did for the test case GetFileInfoRoot, but this seems not related with the PR change:

TEST_F(TestS3FS, GetFileInfoRoot) { AssertFileInfo(fs_.get(), "", FileType::Directory); }

For the gcs test, I failed to build the google cloud cpp lib and I am not able to run it locally so far, but it seems happened in other PR as well (for example, #12816)

Hmm, I don't think this is the right way to do this. The IPC message's custom_metadata is independent from record batch (i.e. schema) metadata, one should not influence the other.

(also, you could have a custom_metadata for other message types such as Tensor)

cc @wesm for opinions.

The IPC message's custom_metadata is independent from record batch (i.e. schema) metadata

This is actually my initial understanding, according to this documentation (https://arrow.apache.org/docs/format/Columnar.html#custom-application-metadata), it says This includes Field, Schema, and Message. and it seems to indicate schema metadata is different from message metadata.

But for the documentation here (https://arrow.apache.org/docs/python/data.html#custom-schema-and-field-metadata), it says Arrow supports both schema-level and field-level custom key-value metadata allowing for systems to insert their own application defined metadata to customize behavior. and record batch (message level) metadata is not mentioned.

And I only see API/tests allowing users to modify schema's metadata (probably I am missing something), so I assumed my initial understanding is incorrect and came up with such a PR change.

If message metadata is different from schema metadata, the question here may be slightly more complex since in this case, each record batch may provide 1) this record batch's schema specific metadata, this seems not serialized/deserialized 2) this record batch's message specific metadata, we don't seem to provide API for reading/writing this (internally we have AppendCustomMetadata but it is not called and there seems no higher level API allowing users to pass this information)

Please advice what the correct behavior should be so that I could give it a try. Thanks.

So, basically, per-message metadata is an Arrow IPC feature that's independent from record batches or schemas. That's why you don't see it mentioned in the schema metadata documentation.

If message metadata is different from schema metadata, the question here may be slightly more complex since in this case, each record batch may provide 1) this record batch's schema specific metadata, this seems not serialized/deserialized 2) this record batch's message specific metadata, we don't seem to provide API for reading/writing this

Indeed the message metadata wouldn't be attached to the record batch or schema, but passed separately.

One possible API on the write side:

class ARROW_EXPORT RecordBatchWriter {
 public:
  // (default implementation could call `WriteRecordBatch(batch)` while ignoring metadata?)
  virtual Status WriteRecordBatch(const RecordBatch& batch, const KeyValueMetadata& custom_metadata);

and on the read side:

class ARROW_EXPORT RecordBatchReader {
 public:
  // (default implementation could call `ReadNext(batch)` and reset `*custom_metadata` to null)
  virtual Status ReadNext(std::shared_ptr<RecordBatch>* batch, std::shared_ptr<KeyValueMetadata>* custom_metadata);

One possible API on the write side

Thanks for the advice. Let me give it a try.

@pitrou @lidavidm
I submitted a new commit following the suggestion above. I tried to keep the changes compatible with existing public APIs so several new APIs are added. Could you please help to review? Let me know if there is any issue. Thanks.

For the CI:

1. I found one test case 80 - arrow-compute-hash-join-node-test failed in C++ / AMD64 Conda C++ (pull_request), but I guess this is probably not caused by my PR
2. another test case 51 - arrow-flight-test (Failed) failed in continuous-integration/appveyor/pr, I am not able to find the detailed test report in the CI job, is this an issue?

@niyue Indeed, these CI failures are unrelated. I'll let @lidavidm take a look at the Flight one; the hash-join-node-test crash is at https://issues.apache.org/jira/browse/ARROW-15361

I filed ARROW-16215

I rebased the PR onto the latest master branch and resolved the conflict in read_write_test.cc.

@pitrou @lidavidm could you please help to review if there is anything else that needs to be revised? Thanks.

Thanks a lot for your contribution @niyue !

@pitrou @lidavidm thanks so much for the detailed review

Benchmark runs are scheduled for baseline = 52698f3 and contender = 942f77e. 942f77e is a master commit associated with this PR. Results will be available as each benchmark for each run completes.
Conbench compare runs links:
[Finished ⬇️0.0% ⬆️0.0%] ec2-t3-xlarge-us-east-2
[Failed] test-mac-arm
[Failed ⬇️0.0% ⬆️0.36%] ursa-i9-9960x
[Finished ⬇️0.88% ⬆️0.0%] ursa-thinkcentre-m75q
Buildkite builds:
[Finished] 942f77e5 ec2-t3-xlarge-us-east-2
[Failed] 942f77e5 test-mac-arm
[Failed] 942f77e5 ursa-i9-9960x
[Finished] 942f77e5 ursa-thinkcentre-m75q
[Finished] 52698f35 ec2-t3-xlarge-us-east-2
[Failed] 52698f35 test-mac-arm
[Finished] 52698f35 ursa-i9-9960x
[Finished] 52698f35 ursa-thinkcentre-m75q
Supported benchmarks:
ec2-t3-xlarge-us-east-2: Supported benchmark langs: Python, R. Runs only benchmarks with cloud = True
test-mac-arm: Supported benchmark langs: C++, Python, R
ursa-i9-9960x: Supported benchmark langs: Python, R, JavaScript
ursa-thinkcentre-m75q: Supported benchmark langs: C++, Java