hi Al, It's definitely wrong. I confirmed the behavior is present on master.
https://issues.apache.org/jira/browse/ARROW-10121 I made this a blocker for the release. Thanks, Wes On Mon, Sep 28, 2020 at 10:52 AM Al Taylor <al.taylor1...@googlemail.com.invalid> wrote: > > Hi, > > I've found that when I serialize two recordbatches which have a > dictionary-encoded field, but different encoding dictionaries to a sequence > of pybytes with a RecordBatchStreamWriter, then deserialize using > pa.ipc.open_stream(), the dictionaries get jumbled. (or at least, on > deserialization, the dictionary for the first RB is being reused for the > second) > > MWE: > ``` > import pyarrow as pa > from io import BytesIO > > pa.__version__ > > schema = pa.schema([pa.field('foo', pa.int32()), pa.field('bar', > pa.dictionary(pa.int32(), pa.string()))] ) > r1 = pa.record_batch( > [ > [1, 2, 3, 4, 5], > pa.array(["a", "b", "c", "d", "e"]).dictionary_encode() > ], > schema > ) > > r1.validate() > r2 = pa.record_batch( > [ > [1, 2, 3, 4, 5], > pa.array(["c", "c", "e", "f", "g"]).dictionary_encode() > ], > schema > ) > > r2.validate() > > assert r1.column(1).dictionary != r2.column(1).dictionary > > > sink = pa.BufferOutputStream() > writer = pa.RecordBatchStreamWriter(sink, schema) > > writer.write(r1) > writer.write(r2) > > serialized = BytesIO(sink.getvalue().to_pybytes()) > stream = pa.ipc.open_stream(serialized) > > deserialized = [] > > while True: > try: > deserialized.append(stream.read_next_batch()) > except StopIteration: > break > > deserialized[0].column(1).to_pylist() > deserialized[1].column(1).to_pylist() > ``` > (The last line of the above prints out `['a', 'a', 'b', 'c', 'd']`. This > behaviour doesn't look right. I was wondering whether I'm simply not using > the library correctly or if this is a bug in pyarrow. > > Thanks, > > Al