What’s the migration path for ray.data.datasource.tfrecords_datasource.TFRecordDatasource?

Ray Data
1

I was using from ray.data.datasource.tfrecords_datasource import TFRecordDatasource in ray 2.23.0. But after upgrading to 2.40.0, this class was moved to ray.data._internal.datasource.tfrecords_datasource.TFRecordDatasource from this change.

similar to What's the migration path for ray.data.aggregate's Max, Mean, Min, and Std functions?

2

@davidxia could you share your use case/why you need to import that? TFRecordDatasource, like our other data sources, is supposed to be a private implementation.

We recommend just maintaining a separate version, or figuring out how we can support your changes to TFRecordDatasource if you’ve made extensions.

3

We were using when it was public but alpha in 2.23.0. It was privatized here. We’ve since stopped using the class directly. I think we were using it before it had the tfx-bsl performance improvement.

Now we use ray.data.read_tfrecords(tfx_read_options=...). Using the tfx_read_options parameter which will read TFRecords with tfx-bsl which is faster.

1 Like