[jira] [Commented] (ARROW-4316) Reusing arrow.so for both Python and R

["Jeffrey Wong (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/ARROW-4316?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16748403#comment-16748403 ] 

Jeffrey Wong commented on ARROW-4316:
-------------------------------------

I just built arrow 0.12.0 from github using the tag apache-arrow-0.12.0. I compiled arrow with the flags


cmake .. -DCMAKE_BUILD_TYPE=Release -DARROW_PARQUET=ON -DARROW_PYTHON=ON
make install

which produces the .so files in /usr/local/lib. I noticed that the libarrow.so file is 8.3 Mb, but the one that is shipped with pyarrow is 9.7 Mb. What else is inside pyarrow's .so files? When I link against the .so's from github (and remove the command 
sed -i "s/PKG_CXXFLAGS=/PKG_CXXFLAGS=-D_GLIBCXX_USE_CXX11_ABI=0 /g" src/Makevars.infrom my script) the R package will build successfully. To debug further, I think I need to know the difference between the .so files in pyarrow and the .so files that are built from source

> Reusing arrow.so for both Python and R
> --------------------------------------
>
>                 Key: ARROW-4316
>                 URL: https://issues.apache.org/jira/browse/ARROW-4316
>             Project: Apache Arrow
>          Issue Type: Bug
>          Components: Python, R
>    Affects Versions: 0.12.0
>         Environment: Ubuntu 16.04, R 3.4.4, pyarrow 0.12, cmake 3.12
>            Reporter: Jeffrey Wong
>            Priority: Major
>
> My team uses both pyarrow and R arrow, we'd like both libraries to link to the same arrow.so file for consistency. pyarrow ships both arrow.so and parquet.so, if I can reuse those .so's to  link R that would guarantee consistency. 
>  Under arrow v0.11.1 I was able to link R against libarrow.so found under pyarrow by passing LIB_DIR to the R [configure file|https://github.com/apache/arrow/blob/master/r/configure]. However, in v0.12.0 I am no longer able to do that. Here is a reproducible example on Ubuntu 16.04 which produces the error:
>  
> {code:java}
> sh: line 1: 5404 Segmentation fault (core dumped) '/usr/lib/R/bin/R' --no-save --slave 2>&1 < '/tmp/RtmpyOuz4g/file14716feda8fc'
> *** caught segfault ***
> address 0x7f160f026250, cause 'invalid permissions'
> An irrecoverable exception occurred. R is aborting now ...
> {code}
>  
>  Reproducible example:
> {code:java}
>  # get the parquet headers which are not shipped with pyarrow
>   
> tee /etc/apt/sources.list.d/apache-arrow.list <<APT_LINE
> deb [arch=amd64] https://dl.bintray.com/apache/arrow/$(lsb_release --id --short | tr 'A-Z' 'a-z')/ $(lsb_release --codename --short) main
> deb-src [] https://dl.bintray.com/apache/arrow/$(lsb_release --id --short | tr 'A-Z' 'a-z')/ $(lsb_release --codename --short) main
> APT_LINE
> apt-get update
> mkdir /tmp/arrow_headers; cd /tmp/arrow_headers
> apt-get download --allow-unauthenticated libparquet-dev
> ar -x libparquet-dev_0.12.0-1_amd64.deb
> tar -xJvf data.tar.xz
>   
>  #get pyarrow v0.12
>   
>  pip3 install pyarrow --upgrade
>  #figure out where pyarrow is
>  PY_ARROW_PATH=$(python3 -c "import pyarrow, os; print(os.path.dirname(pyarrow.__file__))")
>  PY_ARROW_VERSION=$(python3 -c "import pyarrow; print(pyarrow.__version__)")
>  PYTHON_LIBDIR=$(python3 -c "import sysconfig; print(sysconfig.get_config_var('LIBDIR'))")
>   
>  # pyarrow doesn't ship parquet headers. Copy the ones from apt into the pyarrow dir
>  mkdir $PY_ARROW_PATH/include/parquet
>  cp -r /tmp/arrow_headers/usr/include/parquet/* $PY_ARROW_PATH/include/parquet/
>   
>  #install R arrow
>  echo "export LD_LIBRARY_PATH=\"\${LD_LIBRARY_PATH}:${PYTHON_LIBDIR}:${PY_ARROW_PATH}\"" | tee -a /usr/lib/R/etc/ldpaths
>  git clone https://github.com/apache/arrow.git /tmp/arrow
>  cd /tmp/arrow/r
>  git checkout "apache-arrow-${PY_ARROW_VERSION}"
>  sed -i "/Depends: R/c\Depends: R (>= 3.4)" DESCRIPTION
>  sed -i "s/PKG_CXXFLAGS=/PKG_CXXFLAGS=-D_GLIBCXX_USE_CXX11_ABI=0 /g" src/Makevars.in
>  R CMD INSTALL ./ --configure-vars="INCLUDE_DIR=$PY_ARROW_PATH/include LIB_DIR=$PY_ARROW_PATH" {code}



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)