arrow-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Wes McKinney <wesmck...@gmail.com>
Subject Re: [C++] Apply Gandiva Filter to a RecordBatch
Date Sun, 05 Apr 2020 20:35:52 GMT
You might pose your question on dev@arrow.apache.org, more of the
people with closer Gandiva knowledge are present there

On Sat, Apr 4, 2020 at 10:45 PM Yue Ni <niyue.com@gmail.com> wrote:
>
> Thanks so much. This is exactly what I was looking for, and I will give it a try.
>
> BTW, as a follow up question, I wonder if there is anyone has any idea that if I want
to apply both selection and projection to a record batch, is there any performance difference
between these two ways:
> 1) use filter to get a selection vector and convert it to an array ==> use arrow::compute::Take
to filter the record batch ==> construct a projector to do projection for the filtered
record batch
> 2) use filter to get a selection vector ==> construct projector to do projection and
apply the selection vector at the same time
>
> The first approach allows me to process a query step by step (first selection then projection)
and the second approach is more concise but seems not clearly separated compared to the first
approach. I prefer the first approach since it is easier to handle a filtering-only query,
but would like to confirm if it will degrade the performance if both selection and projection
are needed.
>
> On Sun, Apr 5, 2020 at 11:08 AM Wes McKinney <wesmckinn@gmail.com> wrote:
>>
>> Try the arrow::compute::Take function
>>
>> https://github.com/apache/arrow/blob/master/cpp/src/arrow/compute/kernels/take.h#L121
>>
>> On Sat, Apr 4, 2020 at 8:19 PM Yue Ni <niyue.com@gmail.com> wrote:
>> >
>> > Hi Wes,
>> >
>> > Thanks for the reply, but I don't think this is what I am looking for.
>> >
>> > It seems to me this `result.to_array()` will only return the array for the selection
vector (https://github.com/apache/arrow/blob/b07c2626cb3cdd3498b41da9feedf7c8319baa27/python/pyarrow/gandiva.pyx#L130),
but it is not clear to me how I can use the selection vector to filter the original record
batch.
>> >
>> > If I understand correctly, in this test case, `result.to_array().equals(pa.array(range(1000),
type=pa.uint32()))` is asserting that the selection vector has integer index values from [0,
1000), but I am looking for to obtain an array in the filtered record batch which should be
an array of floats here. I know I can iterate indices in the selection vector and use it to
retrieve each row in original record batch columns, but I am not certain if this is the right
way to do it. For example, if I have multiple columns in the original record batch, do I need
to iterate the selection vector multiple times to filter each of the column? Since this is
a common task, I expect there is an easy/efficient API to do this.
>> >
>> > Basically, I am looking for something like:
>> > selection_vector = filter.evaluate(record_batch, pa.default_memory_pool())
>> > filtered_column_arrays_in_record_batch = record_batch.filter(selection_vector)
# what is the API for doing this?
>> >
>> > In the C++ test cases, the closest thing I find is to construct a gandiva projector
to use the selection vector, but every test case there requires client to construct a gandiva
expression to build the projector (for example, in this test case, a {sum_expr} is used for
constructing the projector, https://github.com/apache/arrow/blob/967728fe4654e5d53bc0789e64e5a9ba7f27f263/cpp/src/gandiva/tests/filter_project_test.cc#L121
).
>> >
>> > I wonder if the filtering can be done without involving creating a projection
expression. At the same time, if projector is expected to be used for doing this, what projector
expressions should be used if I want to keep all the columns as they are but just with some
rows filtered based on the criteria given?
>> >
>> > On Sun, Apr 5, 2020 at 7:27 AM Wes McKinney <wesmckinn@gmail.com> wrote:
>> >>
>> >> You can see an example of filtering via the Python bindings
>> >>
>> >> https://github.com/apache/arrow/blob/master/python/pyarrow/tests/test_gandiva.py#L89
>> >>
>> >> This creates a gandiva::Filter using gandiva::Filter::Make, which can
>> >> be used to filter a RecordBatch
>> >>
>> >> Is this what you need?
>> >>
>> >> On Fri, Apr 3, 2020 at 7:12 PM Yue Ni <niyue.com@gmail.com> wrote:
>> >> >
>> >> > Hi there,
>> >> >
>> >> > I am using the gandiva C++ library for processing RecordBatch. I would
like to know how I can apply gandiva::Filter for a RecordBatch so that I can do some filtering
without using the projector.
>> >> >
>> >> > Since I don't find any documentation for it, I read some source code
about its usage, and here are the test cases I found about its usage:
>> >> > 1) https://github.com/apache/arrow/blob/967728fe4654e5d53bc0789e64e5a9ba7f27f263/cpp/src/gandiva/tests/filter_test.cc
>> >> > 2) https://github.com/apache/arrow/blob/967728fe4654e5d53bc0789e64e5a9ba7f27f263/cpp/src/gandiva/tests/filter_project_test.cc
>> >> >
>> >> > From my reading, I find it is possible to get a SelectionVector by
using the gandiva::Filter, at the same time, you can use the SelectionVector with the gandiva::Projector
to filter RecordBatch when doing projection. My questions are:
>> >> > 1) if I don't want to do any projection but simply filtering, what
is the recommended way to do it?
>> >> > 2) I am trying to handle the case like "SELECT * FROM table WHERE blah",
is it recommended to apply filtering without projection in this case or is there any alternative
approach doing it?
>> >> >
>> >> > Thanks.
>> >> >
>> >> > Regards,
>> >> > Yue
>> >> >

Mime
View raw message