Enable Zstandard decompression only when all nvcomp integrations are enabled

[vuule]

Closes #9055
Limiting due to known nvcomp issues.

[codecov]

Codecov Report

Merging #10944 (abb966f) into branch-22.06 (54789ee) will increase coverage by 0.02%.
The diff coverage is 93.75%.

❗ Current head abb966f differs from pull request most recent head af35471. Consider uploading reports for the commit af35471 to get more accurate results

@@               Coverage Diff                @@
##           branch-22.06   #10944      +/-   ##
================================================
+ Coverage         86.30%   86.32%   +0.02%     
================================================
  Files               144      144              
  Lines             22665    22668       +3     
================================================
+ Hits              19560    19569       +9     
+ Misses             3105     3099       -6     

Impacted Files	Coverage Δ
python/cudf/cudf/core/indexed_frame.py	91.70% <ø> (ø)
python/cudf/cudf/utils/ioutils.py	79.47% <87.50%> (-0.13%)	⬇️
python/cudf/cudf/io/avro.py	78.57% <100.00%> (ø)
python/cudf/cudf/io/csv.py	91.80% <100.00%> (ø)
python/cudf/cudf/io/json.py	97.56% <100.00%> (ø)
python/cudf/cudf/io/orc.py	92.77% <100.00%> (ø)
python/cudf/cudf/io/parquet.py	90.83% <100.00%> (ø)
python/cudf/cudf/io/text.py	100.00% <100.00%> (ø)
python/cudf/cudf/core/dataframe.py	93.78% <0.00%> (+0.04%)	⬆️
python/cudf/cudf/core/column/string.py	88.78% <0.00%> (+0.12%)	⬆️
... and 4 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 35f9d5e...af35471. Read the comment docs.

[beckernick]

Despite caveats, this (and the prior PR) should close #9055 right?

[vuule]

Despite caveats, this (and the prior PR) should close #9055 right?

I think it should. Thanks for the reminder.

[jbrennan333]

I tested this with our spark-rapids integration test (after building with nvcomp-2.3).
With LIBCUDF_NVCOMP_POLICY=STABLE, the zstd compression tests fail with:

ai.rapids.cudf.CudfException: cuDF failure at: /rapids/spark-rapids-jni/thirdparty/cudf/cpp/src/io/comp/nvcomp_adapter.cpp:34: Unsupported compression type

When I set it to LIBCUDF_NVCOMP_POLICY=ALWAYS, the zstd compression tests pass.

../../src/main/python/parquet_test.py::test_parquet_compress_read_round_trip[parquet-reader_confs5-none] PASSED [ 93%]
../../src/main/python/parquet_test.py::test_parquet_compress_read_round_trip[parquet-reader_confs5-uncompressed] PASSED [ 95%]
../../src/main/python/parquet_test.py::test_parquet_compress_read_round_trip[parquet-reader_confs5-snappy] PASSED [ 96%]
../../src/main/python/parquet_test.py::test_parquet_compress_read_round_trip[parquet-reader_confs5-gzip] PASSED [ 98%]
../../src/main/python/parquet_test.py::test_parquet_compress_read_round_trip[parquet-reader_confs5-zstd] temp bytes 27973
temp bytes 27931
temp bytes 27955
temp bytes 27940
PASSED [100%]

[jbrennan333]

I tested this with our spark-rapids integration test (after building with nvcomp-2.3).
With LIBCUDF_NVCOMP_POLICY=STABLE, the zstd compression tests fail with:

ai.rapids.cudf.CudfException: cuDF failure at: /rapids/spark-rapids-jni/thirdparty/cudf/cpp/src/io/comp/nvcomp_adapter.cpp:34: Unsupported compression type

When I set it to LIBCUDF_NVCOMP_POLICY=ALWAYS, the zstd compression tests pass.

../../src/main/python/parquet_test.py::test_parquet_compress_read_round_trip[parquet-reader_confs5-none] PASSED [ 93%]
../../src/main/python/parquet_test.py::test_parquet_compress_read_round_trip[parquet-reader_confs5-uncompressed] PASSED [ 95%]
../../src/main/python/parquet_test.py::test_parquet_compress_read_round_trip[parquet-reader_confs5-snappy] PASSED [ 96%]
../../src/main/python/parquet_test.py::test_parquet_compress_read_round_trip[parquet-reader_confs5-gzip] PASSED [ 98%]
../../src/main/python/parquet_test.py::test_parquet_compress_read_round_trip[parquet-reader_confs5-zstd] temp bytes 27973
temp bytes 27931
temp bytes 27955
temp bytes 27940
PASSED [100%]

[jbrennan333]

+1 lgtm. I tested with this change locally as well.

[bdice]

The CUDF_FAIL macro provides helpful context, like line numbers where the error is thrown, which are lost when the macro is replaced by fail_unsupported. I would continue using CUDF_FAIL in all cases even though the message is repeated.

cudf/cpp/include/cudf/utilities/error.hpp

Line 110 in 379cc9f

throw cudf::logic_error("cuDF failure at: " __FILE__ ":" CUDF_STRINGIFY(__LINE__) ": " reason)

[vyasr]

Not sure that we need the extra level of indirection of fail_unsupported instead of just calling CUDF_FAIL directly, but this looks fine to me.

[vuule]

Reworked the errors completely. Should be more informative now.

[bdice]

Discussed error handling offline. Future work can avoid testing for specific error messages, which was part of the motivation for wrapping the CUDF_FAIL macro in a common function.

[vuule]

@gpucibot merge