library/test_search segfault on master

[jprobichaud]

Perhaps caused by the same issue as issue "Hook interface" #1413, building vw and then doing

cd library
make
./test_search

Also results in a core dump

This is on ArchLinux, 64bit, with GCC 8.2.1 and Boost 1.68

[jprobichaud]

adding print statements, I identified that the segfault happens "between" these two lines, just after the get_task_data:

     HookTask::task_data* d = sch.get_task_data<HookTask::task_data>();
     d->run_f = _search_run_fn;

It seems to be the case also when executing from python

[jprobichaud]

With the recent updates to master, we now get some extra details:

$ library/test_search 


##### train() #####

warning: task should not set options except in initialize function!
terminate called after throwing an instance of 'VW::vw_exception'
  what():  using LDF and actions costs is not yet implemented; turn off action costs
Aborted (core dumped)

[jprobichaud]

I coulnd't have a chance to spend some time on this again, will anyone from the "official" team be able to have a look?

[JohnLangford]

We've been making some headway on issues, so they may be dealt with soonish.

…

-Jon

On Mon, Jan 28, 2019 at 1:55 PM Jp ***@***.***> wrote: I coulnd't have a chance to spend some time on this again, will anyone from the "official" team be able to have a look? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#1671 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAE25uBcCgiKkC1HauS1Fx5XYGp1rM4Cks5vH0edgaJpZM4YNHKl> .

[lokitoth]

I've been looking at the other hook task issue and it seems like this may be the same issue.

The higher-level issue is that we explicitly disable "hook" - which seems to have been around for a while. That is fixed by removing the check in search.cc's setup(). The next-level issue is that this uses the old API for multi-line example, which fails the multi-line test. Switching to multi_ex as input fixes that. The underlying issue is that setting the right function-pointers in SearchTask's init (in libsearch.h) does not seem to actually set them in the task-object that hook-task is looking into.

Still trying to debug why that is going on.

[lokitoth]

I believe I have a fix for this - just need to do a bit more validation under Linux. After that will check whether there is more work (I expect there will be, to propagate the API changes through the Python library).

Here's the branch in my fork: https://github.com/lokitoth/vowpal_wabbit/tree/dev/Fix-SearchHookTaskCppLibrary
cc @jprobichaud

[jprobichaud]

wonderful, thanks a lot for looking into it! I will try to recompile my C++ code later this week on your branch. I'm working exclusively on linux so this will hopefully test that side of the house a bit more. thanks again!

…

On Tue, Feb 26, 2019 at 4:32 PM Jacob Alber ***@***.***> wrote: I believe I have a fix for this - just need to do a bit more validation under Linux. After that will check whether there is more work (I expect there will be, to propagate the API changes through the Python library). Here's the branch in my fork: lokitoth/vowpal_wabbit:dev/Fix-SearchHookTaskCppLibrary cc @jprobichaud <https://github.com/jprobichaud> — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#1671 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AI0166l_TrDuLHSBY5co1NNWlYBDHBomks5vRaf8gaJpZM4YNHKl> .