[BUG] Unexpected OOM writing a DataFrame w/ strings to ORC files

[randerzander]

On a 32GB V100, the below snippet completes successfully with the RAPIDS 0.18 release, but fails w/ an OOM in the latest 0.19 nightly:

import cudf

df = cudf.datasets.randomdata(nrows=20_000_000)
df['teststr'] = 'teststr'
df.to_orc('test.orc', compression='snappy')

Trace:

---------------------------------------------------------------------------
MemoryError                               Traceback (most recent call last)
<ipython-input-1-33735a6b5c55> in <module>
      3 df = cudf.datasets.randomdata(nrows=20_000_000)
      4 df['teststr'] = 'teststr'
----> 5 df.to_orc('test.orc', compression='snappy')

~/conda/envs/rapids-gpu-bdb/lib/python3.7/site-packages/cudf/core/dataframe.py in to_orc(self, fname, compression, *args, **kwargs)
   7396         from cudf.io import orc as orc
   7397 
-> 7398         orc.to_orc(self, fname, compression, *args, **kwargs)
   7399 
   7400     def stack(self, level=-1, dropna=True):

~/conda/envs/rapids-gpu-bdb/lib/python3.7/site-packages/cudf/io/orc.py in to_orc(df, fname, compression, enable_statistics, **kwargs)
    329             liborc.write_orc(df, file_obj, compression, enable_statistics)
    330     else:
--> 331         liborc.write_orc(df, path_or_buf, compression, enable_statistics)
    332 
    333 

cudf/_lib/orc.pyx in cudf._lib.orc.write_orc()

cudf/_lib/orc.pyx in cudf._lib.orc.write_orc()

MemoryError: std::bad_alloc: CUDA error at: /home/rgelhausen/conda/envs/rapids-gpu-bdb/include/rmm/mr/device/cuda_memory_resource.hpp:69: cudaErrorMemoryAllocation out of memory

Interestingly, w/ nrows=100_000_000 doesn't OOM.