[Gandiva] Segfault when the validity bitmap has not been allocated

[asfimport]

This is because the validity buffer could be None:

>>> df = pd.DataFrame(np.random.randint(0, 100, size=(2**12, 10)))
>>> pa.Table.from_pandas(df).to_batches()[0].column(0).buffers()
[None, <pyarrow.lib.Buffer object at 0x110c1a228>]
>>> df = pd.DataFrame(np.random.randint(0, 100, size=(2**12, 10))*1.0)
>>> pa.Table.from_pandas(df).to_batches()[0].column(0).buffers()
[<pyarrow.lib.Buffer object at 0x11a2b3030>, <pyarrow.lib.Buffer object at 0x11a2b3228>]

But Gandiva has not implemented it yet, thus accessing a nullptr:

void Annotator::PrepareBuffersForField(const FieldDescriptor& desc, const arrow::ArrayData& array_data, EvalBatch* eval_batch) { 
    int buffer_idx = 0;
    // TODO:  
    // - validity is optional 
    uint8_t* validity_buf = const_cast<uint8_t*>(array_data.buffers[buffer_idx]->data());
    eval_batch->SetBuffer(desc.validity_idx(), validity_buf);
    ++buffer_idx;

 

Reproduce code:

frame_data = np.random.randint(0, 100, size=(2**22, 10))
table = pa.Table.from_pandas(df)
filt = ...  # Create any gandiva filter
r = filt.evaluate(table.to_batches()[0], pa.default_memory_pool()) # segfault

 Backtrace:

* thread #2, queue = 'com.apple.main-thread', stop reason = EXC_BAD_ACCESS (code=1, address=0x10)
 * frame #0: 0x00000001060184fc libarrow.12.dylib`arrow::Buffer::data(this=0x0000000000000000) const at buffer.h:162
 frame #1: 0x0000000106fbed78 libgandiva.12.dylib`gandiva::Annotator::PrepareBuffersForField(this=0x0000000100624dc8, desc=0x000000010101e138, array_data=0x000000010061f8e8, eval_batch=0x0000000100796848) at annotator.cc:65
 frame #2: 0x0000000106fbf4ed libgandiva.12.dylib`gandiva::Annotator::PrepareEvalBatch(this=0x0000000100624dc8, record_batch=0x00000001007a45b8, out_vector=size=1) at annotator.cc:94
 frame #3: 0x00000001071449b7 libgandiva.12.dylib`gandiva::LLVMGenerator::Execute(this=0x0000000100624da0, record_batch=0x00000001007a45b8, output_vector=size=1) at llvm_generator.cc:102
 frame #4: 0x0000000107059a4f libgandiva.12.dylib`gandiva::Filter::Evaluate(this=0x000000010079c668, batch=0x00000001007a45b8, out_selection=std::__1::shared_ptr<gandiva::SelectionVector>::element_type @ 0x00000001007a43e8 strong=2 weak=1) at filter.cc:106
 frame #5: 0x000000010948e002 gandiva.cpython-36m-darwin.so`__pyx_pw_7pyarrow_7gandiva_6Filter_3evaluate(_object*, _object*, _object*) + 1986
 frame #6: 0x0000000100140e8b Python`_PyCFunction_FastCallDict + 475
 frame #7: 0x00000001001d28ca Python`call_function + 602
 frame #8: 0x00000001001cf798 Python`_PyEval_EvalFrameDefault + 24616
 frame #9: 0x00000001001d3cf9 Python`fast_function + 569
 frame #10: 0x00000001001d2899 Python`call_function + 553
 frame #11: 0x00000001001cf798 Python`_PyEval_EvalFrameDefault + 24616
 frame #12: 0x00000001001d34c6 Python`_PyEval_EvalCodeWithName + 2902
 frame #13: 0x00000001001c96e0 Python`PyEval_EvalCode + 48
 frame #14: 0x00000001002029ae Python`PyRun_FileExFlags + 174
 frame #15: 0x0000000100201f75 Python`PyRun_SimpleFileExFlags + 277
 frame #16: 0x000000010021ef46 Python`Py_Main + 3558
 frame #17: 0x0000000100000e08 Python`___lldb_unnamed_symbol1$$Python + 248
 frame #18: 0x00007fff6ea72085 libdyld.dylib`start + 1

 

Reporter: Siyuan Zhuang / @suquark
Assignee: Siyuan Zhuang / @suquark

PRs and other links:

• GitHub Pull Request #2967

_{Note: This issue was originally created as ARROW-3765. Please see the migration documentation for further details.}

[asfimport]

Wes McKinney / @wesm:
Yes, in a number of places we don't allocate a validity bitmap for data without nulls. For large datasets, it would be wasteful to have to create a bitmap with all 1's

[asfimport]

Philipp Moritz / @pcmoritz:
Issue resolved by pull request 2967
#2967