pyarrow.cuda.read_record_batch

pyarrow.cuda.read_record_batch(buffer, schema, *, DictionaryMemo dictionary_memo=None, pool=None)

Construct RecordBatch referencing IPC message located on CUDA device.

While the metadata is copied to host memory for deserialization, the record batch data remains on the device.

Parameters
  • buffer – Device buffer containing the complete IPC message

  • schema (Schema) – The schema for the record batch

  • dictionary_memo (DictionaryMemo, optional) – If message contains dictionaries, must pass a populated DictionaryMemo

  • pool (MemoryPool (optional)) – Pool to allocate metadata from

Returns

batch (RecordBatch) – Reconstructed record batch, with device pointers