Sorry, I think I wasn't clear. I mean the broader issue of C++ extension code. In principle it should be possible to mix C++ extension modules in the same Python process, at least in some cases, especially if they link their own dependencies statically. While "just do it our way, use our tools" is fine for most cases, it might not be fine for some, for a wide variety of reasons. If everyone could agree on C++ compilers and libraries, Linux distros would have standardized ages ago I suppose. This isn't a conda-specific problem; however for conda it becomes a runtime problem rather than a build-time one. But I concede it's a hard problem. Mainly, what I'm suggesting is that the policies and use cases be explicit. Thanks for clarifying.
Regarding Boost particularly, all I meant was that the Boost dependency wasn't apparent in the usual ways, which made debugging problems harder. How would you plan to update Boost even within the ecosystem? Bump the version in the toolchain, rebuild the world, and update all environments? Without explicit dependency, how do you prevent someone from running parquet statically linked to Boost N from running in an environment with boost-cpp==N+1 installed? Thanks, Alex On 02/17/2018 04:20 PM, Wes McKinney wrote:
However, extension modules are always going to have to share the Python process, so this policy kind of says, you can't use external C++ extension code with conda.This is a bit too extreme. What I meant is that you should try not to mix C++ build toolchains. I think this is good advice even without conda/conda-forge in the loop. If conda-forge were supplying the library / build toolchain for the rest of your projects, then everything would be OK.Given the policy, it seems slightly better to link Boost dynamically.We could do this, but it seems like a last resort workaround to the core problem, which is the mixed build toolchain issue. I don't know what Boost's ABI guarantees are, but dynamic linking isn't guaranteed to solve using two libraries built against different versions of Boost in the same process. The boost-cpp package is a pretty chunky runtime dependency also. We could give it a shot and see how it goes in the next release cycle. - Wes On Sat, Feb 17, 2018 at 4:06 PM, Alex Samuel <a...@alexsamuel.net> wrote:OK. I'll probably be able to work around this problem. Just a couple of thoughts for the long term: 1. It seems mostly reasonable to treat conda as closed ecosystem as you describe; other C++ stuff can be deployed by other means. However, extension modules are always going to have to share the Python process, so this policy kind of says, you can't use external C++ extension code with conda. 2. Given the policy, it seems slightly better to link Boost dynamically. I had checked package metadata and shared lib dependencies, and didn't even realize it used Boost, until one of my colleagues actually looked at the symbol table and pointed this out. As it stands, Boost is an undeclared dependency, at least "dependency" in the sense of pinning a version. Thanks for your help. Alex On 02/17/2018 11:32 AM, Wes McKinney wrote:It sounds like for your use case that it would be better for you to build your own Arrow packages that use the same Boost as the rest of your repo. You can possible use the scourge tool that Phillip built to help with this (we're using it to build nightlies). conda-forge is a fairly closed ecosystem under the present circumstances -- the intent is that libraries within it are interoperable with each other, and that packages built with conda-forge binaries as their third party dependencies (e.g. if you were using the boost-cpp conda-forge package) will also be able to work. Using the conda-forge stack as an add-on to a substantial independent C++ library stack is not (IIUC) an intended use case. Note that there are libstdc++-related issues using conda-forge binaries with Anaconda >= 5.0 due to the change in compilers. Hopefully this will get fixed in the next few months. - Wes On Sat, Feb 17, 2018 at 11:14 AM, Alex Samuel <a...@alexsamuel.net> wrote:OK, though if both modules linked Boost statically, I believe they would have distinct copies of global variables. Whether or not this causes problems depends on whether they are purely internal or tied to external state. My hunch is that for Boost::regex, there wouldn't be an issue with two complete copies in the same process, as long as they remained separate. I still suspect our module is picking up some symbols from the copy of Boost statically linked to Parquet rather than the one it pulls in as a shared lib dependency. One thing I could try is to link ours statically as well. I wasn't aware of bcp; I'll take a look at that. It may or may not be possible for us to build our C++ stuff against conda-forge's Boost; I'm not sure. We have a large C++ codebase and distribute parts of it, particularly Python extension modules, via conda. In general, I suspect as conda-forge grows and packages more C++ code, issues like this are likely to become an increasing problem. Do you know if there is a general policy regarding how extension modules in conda packages should link common C++ libraries like Boost? Thanks, Alex On 02/17/2018 11:04 AM, Uwe L. Korn wrote:Static linking does not really solve all problems in the Boost case as there are global variables that are picked up across different Boost versions. Thus if you link it statically it still has an effect on other dependencies. In the case of boost, you can get around this by using the bcp tool http://www.boost.org/doc/libs/1_66_0/tools/bcp/doc/html/index.html to rename your local boost to a different namespace to avoid collisions. We will also do thus in future with our wheel in Arrow, too. But we will not do this for conda-packages as there the assumption is that all artefacts will be linked against the same Boost version. UweAm 17.02.2018 um 16:55 schrieb Alex Samuel <a...@alexsamuel.net>: Yes, we do link our internal code with a different Boost version; we build and (conda) package it ourselves. Why should this matter if Parquet links it statically? If static linking won't allow us to use our own version, why bother linking statically at all? Thanks, AlexOn 02/17/2018 10:52 AM, Uwe L. Korn wrote: Hi, The issue is here no that Boost is linked statically in one and dynamically in another but that you link against two different boost versions. The stacktrace shows links to Boost 1.55 whereas Arrow should be linked against 1.65 or 1.66 (the one coming from conda-forge). Arrow requires at least a Boost version of 1.60+ to work. The most likely guess from my side would be that your internal modules are linked against the system boost, not the conda-provided boost. UweAm 17.02.2018 um 16:47 schrieb Alex Samuel <a...@alexsamuel.net>: Hi there, Sure, I'll append top of the stack. You can see our internal function "asd::infra::util::start_of_date"; everything else is Boost, Python, or libstdc++. My understanding (though I haven't demonstrated this conclusively) is that, because Python loads extension modules RTLD_GLOBAL, an extension module can pick up symbols from another or its dependencies, even if the former "usually" satisfy relocations from their own shared lib dependencies. So, one module linking Boost statically may interfere with another that links it dynamically, by injecting its symbols. If necessary I can try to put together a minimal test case, but no guarantee it will actually trigger the bug. But it might be worth testing my theory above first, with gdb or by some other means. Thanks! Alex #0 0x00007f6ddeba4c8b in std::basic_string<char, std::char_traits<char>, std::allocator<char>::basic_string(std::basic_string<char, std::char_traits<char>,std::allocator<char> > const&) () from /space/asd/conda/envs/rd-20180212-0/lib/python2.7/site-packages/numexpr/../../../libstdc++.so.6 #1 0x00007f6dd4f978d6 in boost::re_detail::cpp_regex_traits_char_layer<char>::init() () from /space/asd/conda/envs/rd-20180212-0/lib/python2.7/site-packages/asd/infra/../../../.././libboost_regex.so.1.55.0 #2 0x00007f6dd4fdbd88 in boost::object_cache<boost::re_detail::cpp_regex_traits_base<char>, boost::re_detail::cpp_regex_traits_implementation<char>::do_get(boost::re_detail::cpp_regex_traits_base<char> const&, unsignedlong) () from /space/asd/conda/envs/rd-20180212-0/lib/python2.7/site-packages/asd/infra/../../../.././libboost_regex.so.1.55.0 #3 0x00007f6dd4fe5bb5 in boost::basic_regex<char, boost::regex_traits<char, boost::cpp_regex_traits<char> >::do_assign(charconst*, char const*, unsigned int) () from /space/asd/conda/envs/rd-20180212-0/lib/python2.7/site-packages/asd/infra/../../../.././libboost_regex.so.1.55.0 #4 0x00007f6dd575a90e in asd::infra::util::start_of_date(std::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, char const*) () at /prod/sys/sysasd/opt/tudor-devtools/v1.3/Linux.el6.x86_64-corei7-avx-gcc4.83-anaconda2.0.1/include/boost/regex/v4/basic_regex.hpp:382 #5 0x00007f6dd2c36734 in boost::python::objects::caller_py_function_impl<boost::python::detail::caller<unsigned long (*)(std::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, char const*), boost::python::default_call_policies, boost::mpl::vector3<unsigned long, std::basic_string<char, std::char_traits<char>, std::allocator<char>const&, char const*> > >::operator()(_object*, _object*) () from /space/asd/conda/envs/rd-20180212-0/lib/python2.7/site-packages/asd/infra/util.so #6 0x00007f6dd52ac71a in boost::python::objects::function::call(_object*, _object*) const () from /space/asd/conda/envs/rd-20180212-0/lib/python2.7/site-packages/asd/infra/../../../../libboost_python.so.1.55.0 #7 0x00007f6dd52aca68 in boost::detail::function::void_function_ref_invoker0<boost::python::objects::(anonymous namespace)::bind_return, void>::invoke(boost::detail::function::function_buffer&) () from /space/asd/conda/envs/rd-20180212-0/lib/python2.7/site-packages/asd/infra/../../../../libboost_python.so.1.55.0 #8 0x00007f6dd52b4cd3 in boost::python::detail::exception_handler::operator()(boost::function0<void> const&) const () from /space/asd/conda/envs/rd-20180212-0/lib/python2.7/site-packages/asd/infra/../../../../libboost_python.so.1.55.0 #9 0x00007f6dd2c32c03 in boost::detail::function::function_obj_invoker2<boost::_bi::bind_t<bool, boost::python::detail::translate_exception<asd::infra::Exception, void (*)(asd::infra::Exception const&)>, boost::_bi::list3<boost::arg<1>, boost::arg<2>, boost::_bi::value<void (*)(asd::infra::Exception const&)> >, bool, boost::python::detail::exception_handler const&,boost::function0<void> const&>::invoke(boost::detail::function::function_buffer&, boost::python::detail::exception_handler const&, boost::function0<void> const&) () from /space/asd/conda/envs/rd-20180212-0/lib/python2.7/site-packages/asd/infra/util.so #10 0x00007f6dd52b4a9d in boost::python::handle_exception_impl(boost::function0<void>) () from /space/asd/conda/envs/rd-20180212-0/lib/python2.7/site-packages/asd/infra/../../../../libboost_python.so.1.55.0 #11 0x00007f6dd52ab2b3 in function_call () from /space/asd/conda/envs/rd-20180212-0/lib/python2.7/site-packages/asd/infra/../../../../libboost_python.so.1.55.0 #12 0x00007f6df2bc8e93 in PyObject_Call (func=0x2799850, arg=<value optimized out>, kw=<value optimized out>) at Objects/abstract.c:2547 #13 0x00007f6df2c7b80d in do_call (f=<value optimized out>, throwflag=<value optimized out>) at Python/ceval.c:4569 #14 call_function (f=<value optimized out>, throwflag=<value optimized out>) at Python/ceval.c:4374 #15 PyEval_EvalFrameEx (f=<value optimized out>, throwflag=<value optimized out>) at Python/ceval.c:2989On 02/17/2018 10:31 AM, Uwe L. Korn wrote: Hello, I am not sure why we are linking statically in the conda-forge packages, as a gut feeling we should link dynamically there. Wes, can you remember why? Alex, would it be possible for you to send us the part of the segmentation fault that is not private to your modules. That would be a good indicator for us what is going wrong. Typically it is best when you enable coredumps with `ulimit -c unlimited` and then run your program as usual. There should be no performance penalty. When ist segfaults, run `gdb python core` (note that the core file might also be postfixed with the PID but that depends on your system). In gdb type 'thread apply all bt full'. Post thd output pf that command and strip away the parts we should not see. Most relevant will be the stacktrace of the thread that segfaulted. UweAm 16.02.2018 um 23:17 schrieb Alex Samuel <a...@alexsamuel.net>: Hello, I am having some troubles using the Continuum PyArrow conda package dependencies in conjunction with internal C++ extension modules. Apparently, Arrow and Parquet link Boost statically. We have some internal packages containing C++ code that linking Boost libs dynamicaly. If we import Feather as well as our own extension modules into the same Python process, we get random segfaults in Boost. I think what's happening is that our extension modules are picking up Boost's symbols from Arrow and Parquet already loaded into the process, rather than from our own Boost shared libs. Could anyone explain the policy for linking Boost in binary distributions, particularly conda packages? What is your expectation for how other C++ extension modules should be built? Thanks in advance, Alex