Adam Hooper created ARROW-6861:
----------------------------------
Summary: With arrow-0.14.1-output Parquet dictionary column:
Failure reading column: IOError: Arrow error: Invalid: Resize cannot downsize
Key: ARROW-6861
URL: https://issues.apache.org/jira/browse/ARROW-6861
Project: Apache Arrow
Issue Type: Bug
Components: C++, Python
Affects Versions: 0.15.0
Environment: debian:buster (in Docker, Linux 5.2.11-200.fc30.x86_64)
Reporter: Adam Hooper
Attachments: fix-dict-builder-capacity.diff
I'll need to jump through hoops to upload the (seemingly-valid) Parquet file
that triggers this bug. In the meantime, here's the error I get, reading the
Parquet file with read_dictionary=true. I'll start with the stack trace:
{{Failure reading column: IOError: Arrow error: Invalid: Resize cannot
downsize}}
{{#0 0x0000000000b9fffd in __cxa_throw ()}}
{{#1 0x00000000004ce7b5 in parquet::PlainByteArrayDecoder::DecodeArrow
(this=0x555556612e50, num_values=67339, null_count=0, valid_bits=0x7f39a764b780
'\377' <repeats 200 times>..., valid_bits_offset=748544,}}
\{{ builder=0x555556616330) at
/src/apache-arrow-0.15.0/cpp/src/parquet/encoding.cc:886}}
{{#2 0x000000000046d703 in
parquet::internal::ByteArrayDictionaryRecordReader::ReadValuesSpaced
(this=0x555556616260, values_to_read=67339, null_count=0)}}
\{{ at /src/apache-arrow-0.15.0/cpp/src/parquet/column_reader.cc:1314}}
{{#3 0x00000000004a13f8 in
parquet::internal::TypedRecordReader<parquet::PhysicalType<(parquet::Type::type)6>
>::ReadRecordData (this=0x555556616260, num_records=67339)}}
\{{ at /src/apache-arrow-0.15.0/cpp/src/parquet/column_reader.cc:1096}}
{{#4 0x0000000000493876 in
parquet::internal::TypedRecordReader<parquet::PhysicalType<(parquet::Type::type)6>
>::ReadRecords (this=0x555556616260, num_records=815883)}}
\{{ at /src/apache-arrow-0.15.0/cpp/src/parquet/column_reader.cc:875}}
{{#5 0x0000000000413955 in parquet::arrow::LeafReader::NextBatch
(this=0x555556615640, records_to_read=815883, out=0x7ffd4b5afab0) at
/src/apache-arrow-0.15.0/cpp/src/parquet/arrow/reader.cc:413}}
{{#6 0x0000000000412081 in parquet::arrow::FileReaderImpl::ReadColumn
(this=0x5555566067a0, i=7, row_groups=..., out=0x7ffd4b5afab0) at
/src/apache-arrow-0.15.0/cpp/src/parquet/arrow/reader.cc:218}}
{{#7 0x00000000004121b0 in parquet::arrow::FileReaderImpl::ReadColumn
(this=0x5555566067a0, i=7, out=0x7ffd4b5afab0) at
/src/apache-arrow-0.15.0/cpp/src/parquet/arrow/reader.cc:223}}
{{#8 0x0000000000405fbd in readParquet(std::__cxx11::basic_string<char,
std::char_traits<char>, std::allocator<char> > const&) ()}}
And now a report of my gdb adventures:
In Arrow 0.15.0, when reading a particular dictionary column
({{read_dictionaries=true}}) with 815883 rows that was written by Arrow 0.14.1,
{{arrow::Dictionary32Builder<arrow::BinaryType>::AppendIndices(...)}} is called
twice (once with 493568 values, once with 254976 values); and then
{{PlainByteArrayDecoder::DecodeArrow()}} is called. (I'm a novice; I don't know
why this column comes in three batches.) On first {{AppendIndices()}} call, the
buffer capacity is equal to the number of values. On second call, that's no
longer the case: the buffer grows using {{BufferBuilder::GrowByFactor}}, so its
capacity is 987136.
But there's a bug: the 987136-capacity buffer is in
{{Dictionary32Builder::indices_builder_}}; so 987136 is stored in
{{Dictionary32Builder::indices_builder_.capacity_}}.
{{Dictionary32Builder::capacity_}} does not change when {{AppendIndices()}} is
called. (Dictionary32Builder behaves like a proxy for its {{indices_builder_}};
but its {{capacity()}} method is not virtual, so things are messy.)
So {{builder.capacity_}} is 0. Then comes the final batch of 67339 values, via
{{DecodeArrow()}}. It calls {{builder->Reserve(num_values)}}. But
{{builder->Reserve(num_values)}} tries to increase the capacity from 0 (its
wrong, cached value) to {{length_ + num_values}} (815883). Since
{{indicies_builder->capacity_}} is 987136, that's a downsize – which throws an
exception.
The only workaround I can find: use {{read_dictionaries=false}}.
This affects Python, too.
I've attached a patch that fixes the issue for my file. I don't know how to
formulate a reduction, though, so I haven't contributed unit tests. I'm also
not certain how FinishInternal is meant to work, so this definitely needs
expert review. (FinishInternal was _definitely_ buggy before my patch; after my
patch it _might_ be buggy but I don't know.)
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
