Sergei, thanks for creating this new dedicated issue.
The bugreport still has logic issues in relation to the observed problem.
We know OQGraph works in earlier versions of MariaDB, and older distro versions, and that the oqgraph codebase is the same for all of these builds.
Therefore,
a) the assertion that oqgraph requires a higher version of boost to build, and
b) the implicit assertion that oqgraph "magically" requires a different/higher version of boost with later mariadb versions,
are both verifiable untrue. Simple logic.
At this point, my hypothesis is that rather than a newer version, that OQgraph codebase effectively requires a specific or older version of Boost to build. If you can please check what version of Boost was used for MariaDB 5.2, I believe it would have been 1.40 or 1.42 ?
As Boost is not used for other purposes on the build environments, it should be possible to have this older version available in the environments, possible using a single package and general apt/yum methods (no hacks). You can pin to a specific version of a package.
Note that earlier the OQGraph codebase included the Boost and Boost Graph libraries to make building easier, however this was removed on knielsen's request. One possible solution would be to put it back in?
We can possibly tweak the codebase to take newer versions into account, but Boost appears to be a bit of a moving target so essentially we can expect it to keep breaking every time you start building in a newer OS version or the Boost libraries are updated in a distro.
What do you reckon might be the best approach to fix the and minimise hassle in the future?
Arjen, I didn't say that "oqgraph magically requires a different
version of boost with later mariadb versions" 
I know why it does, there is no magic here. We've changed the
requirements (to be at least 1.45, not 1.42 as before)
to solve some build issues (related to rtti, as far as I remember).
This bug report is about the fact that one some platforms OQgraph does
not work with a newer Boost.
I tried to repeat the failure, and it took me quite a while of
experimenting. It does not fail everywhere, so it mostly works.
We don't need to upgrade Boost regularly either, this change (1.42 to
1.45) was caused by changing the build system to CMake, that has its own
set of peculiarities, and required different set of workarounds, as
compared to autotools.I can assure you, we do not plan to change the
build system again anytime soon
Anyway. In a few days we'll install Boost again on all builders, and
then we could see in buildbot what fails where and how.
Regards,
Sergei
Serg, thanks for making sense in this issue, as MDEV-206 definitely doesn't and indeed refers to magic requirements, as I described in my analysis of the logic breakdown there.
Antony Curtis replied earlier on the MariaDB dev email thread - that should resolve the RTTI issue.
Then there could conceivably still be an issue tying OQGraph to a specific Boost version, but please test with the build change and let me know, as you have all the different environments there. thanks!
— compare/5.5/storage/oqgraph/CMakeLists.txt 2012-06-11 22:12:28.000000000 -0700
+++ compare/5.5-wl820/storage/oqgraph/CMakeLists.txt 2012-06-11 22:21:40.000000000 -0700
@@ -12,7 +12,7 @@
ENDIF()
IF(BOOST_OK)
- ADD_DEFINITIONS(-DHAVE_OQGRAPH)
+ ADD_DEFINITIONS(-DHAVE_OQGRAPH -DBOOST_NO_TYPEID -DBOOST_NO_RTTI)
IF(MSVC)
SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /EHsc")
ELSE(MSVC)
Still crashes 
Serg - trying to decode the status changes in JIRA for this issue.
Do I understand correctly that you've successfully resolved the problem?
Does that mean OQGraph will be included in builds again from 5.5.25 ?
That's be awesome!
thanks
Regards,
Arjen.
I made some tests under ubuntu 10.04 32b with 2 boost versions:
boost 1.45 - works on debug and release builds
boost 1.49 - works on debug, crashes on release
Adding -fno-strict-aliasing for oqgraph seems to help on the newer boost versions (1.46+)
Thanks, Patryk!
I'll try to add this as a workaround, and let Arjen decide whether it'll be a final fix or not (in the new upstream of the OQGraph).
Serg - just for reference... there is no long a separate upstream of OQGraph maintained, OQGraph's primary copy lives in the MariaDB source tree.
For 5.0 it was separate and then merged in to the OurDelta builds, but since the introduction to MariaDB (5.2) such a construct is no longer necessary. Because of Oracle's policy of not having community engines and the impossibility of sanely producing a plugin storage engine completely independently of their build system, we're just not bothering to build OQgraph for stock MySQL.
doesn't crash now. let's reopen the issue if it'll start crashing again
I still get the crash in tests with boost > 1.45
okay, I've reopened the issue.
How do I need to compile and test to be able to repeat the crash?
Just tested at home on gentoo 64bit (boost 1.46.1, gcc 4.4.5) - same result:
http://pastebin.com/sy9V3DPz
output from cmake:
http://pastebin.com/WYNuvvur
pushed -fno-strict-aliasing workaround.
Still crashes in buildbot. Look for failures of oqgraph_basic in 5.3
pushed in 5.3 too