======================================================================
__________ ____ ____ _________
/ ____/ __ \/ __ \/ __ \/ ____/ |
/ / __/ /_/ / / / / /_/ / / / /| |
/ /_/ / ____/ /_/ / _, _/ /___/ ___ |
\____/_/ \____/_/ |_|\____/_/ |_|
The Greenplum Query Optimizer
Copyright (c) 2015, Pivotal Software, Inc.
Licensed under the Apache License, Version 2.0
======================================================================
Welcome to GPORCA, the Greenplum Next Generation Query Optimizer!
To understand the objectives and architecture of GPORCA please refer to the following articles:
- Orca: A Modular Query Optimizer Architecture for Big Data.
- Profiling Query Compilation Time with GPORCA
- Improving Constraints In ORCA
Want to Contribute?
GPORCA supports various build types: debug, release with debug info, release. On x86 systems, GPORCA can also be built as a 32-bit or 64-bit library. You'll need CMake 3.1 or higher to build GPORCA. Get it from cmake.org, or your operating system's package manager.
git clone https://github.com/greenplum-db/gporca.git
cd gporca
GPORCA uses the following library:
- GP-Xerces - Greenplum's patched version of Xerces-C 3.1.X
GP-XERCES is available here. The GP-XERCES README gives instructions for building and installing.
ORCA is built with CMake, so any build system supported by CMake can be used. The team uses Ninja because it's really really fast and convenient.
Go into gporca directory:
cmake -GNinja -H. -Bbuild
ninja install -C build
To run all GPORCA tests, simply use the ctest command from the build directory
after build finishes.
ctest
Much like make, ctest has a -j option that allows running multiple tests in
parallel to save time. Using it is recommended for faster testing.
ctest -j8
By default, ctest does not print the output of failed tests. To print the
output of failed tests, use the --output-on-failure flag like so (this is
useful for debugging failed tests):
ctest -j8 --output-on-failure
To run only the previously failed ctests, use the --rerun-failed flag.
ctest -j8 --rerun-failed --output-on-failure
To run a specific individual test, use the gporca_test executable directly.
./server/gporca_test -U CAggTest
To run a specific minidump, for example for ../data/dxl/minidump/TVFRandom.mdp:
./server/gporca_test -d ../data/dxl/minidump/TVFRandom.mdp
Note that some tests use assertions that are only enabled for DEBUG builds, so DEBUG-mode tests tend to be more rigorous.
Most of the regression tests come in the form of a "minidump" file. A minidump is an XML file that contains all the input needed to plan a query, including information about all tables, datatypes, and functions used, as well as statistics. It also contains the resulting plan.
A new minidump can be created by running a query on a live GPDB server:
- Run these in a psql session:
set client_min_messages='log';
set optimizer=on;
set optimizer_enumerate_plans=on;
set optimizer_minidump=always;
set optimizer_enable_constant_expression_evaluation=off;
- Run the query in the same psql session. It will create a minidump file under the "minidumps" directory, in the master's data directory:
$ ls -l $MASTER_DATA_DIRECTORY/minidumps/
total 12
-rw------- 1 heikki heikki 10818 Jun 10 22:02 Minidump_20160610_220222_4_14.mdp
- Run xmllint on the minidump to format it better, and copy it under the data/dxl/minidump directory:
xmllint --format $MASTER_DATA_DIRECTORY/minidumps/Minidump_20160610_220222_4_14.mdp > data/dxl/minidump/MyTest.mdp
- Add it to the test suite, in server/src/unittest/gpopt/minidump/CICGTest.cpp
--- a/server/src/unittest/gpopt/minidump/CICGTest.cpp
+++ b/server/src/unittest/gpopt/minidump/CICGTest.cpp
@@ -217,6 +217,7 @@ const CHAR *rgszFileNames[] =
"../data/dxl/minidump/EffectsOfJoinFilter.mdp",
"../data/dxl/minidump/Join-IDF.mdp",
"../data/dxl/minidump/CoerceToDomain.mdp",
+ "../data/dxl/minidump/Mytest.mdp",
"../data/dxl/minidump/LeftOuter2InnerUnionAllAntiSemiJoin.mdp",
#ifndef GPOS_DEBUG
// TODO: - Jul 14 2015; disabling it for debug build to reduce testing time
Alternatively, it could also be added to the proper test suite in server/CMakeLists.txt as follows:
--- a/server/CMakeLists.txt
+++ b/server/CMakeLists.txt
@@ -183,7 +183,8 @@ CPartTbl5Test:
PartTbl-IsNullPredicate PartTbl-IsNotNullPredicate PartTbl-IndexOnDefPartOnly
PartTbl-SubqueryOuterRef PartTbl-CSQ-PartKey PartTbl-CSQ-NonPartKey
PartTbl-LeftOuterHashJoin-DPE-IsNull PartTbl-LeftOuterNLJoin-DPE-IsNull
-PartTbl-List-DPE-Varchar-Predicates PartTbl-List-DPE-Int-Predicates;
+PartTbl-List-DPE-Varchar-Predicates PartTbl-List-DPE-Int-Predicates
+Mytest;
In some situations, a failing test does not necessarily imply that the fix is wrong. Occasionally, existing tests need to be updated. There is now a script that allows for users to quickly and easily update existing mdps. This script takes in a logfile that it will use to update the mdps. This logfile can be obtained from running ctest as shown below.
Existing minidumps can be updated by runing the following:
-
Run
ctest -j8. -
If there are failing tests, run
ctest -j8 --rerun-failed --output-on-failure | tee /tmp/failures.out
- The output file can then be used with the
fix_mdps.pyscript.
gporca/scripts/fix_mdps.py --logFile /tmp/failures.out
Note: This will overwrite existing mdp files. This is best used after
committing existing changes, so you can more easily see the diff.
Alternatively, you can use gporca/scripts/fix_mdps.py --dryRun to not change
mdp files
- Ensure that all changes are valid and as expected.
GPORCA contains a series of pipeline and task files to run various sets of tests on concourse. You can learn more about deploying concourse with bosh at bosh.io.
Our concourse currently runs the following sets of tests:
- build and ctest on centos6
- build and ctest on centos7
- build and ctest on ubuntu18
All configuration files for our concourse pipelines can be found in the concourse/
directory.
Note: concourse jobs and pipelines for GPORCA are currently experimental and should not be considered ready for use in production-level CI environments.
Here are a few build flavors (commands run from the ORCA checkout directory):
# debug build
cmake -GNinja -D CMAKE_BUILD_TYPE=DEBUG -H. -Bbuild.debug
# release build with debug info
cmake -GNinja -D CMAKE_BUILD_TYPE=RelWithDebInfo -H. -Bbuild.release
It is recommended to use the --prefix option to the Xerces-C configure script
to install GP-Xerces in a location other than the default under /usr/local/,
because you may have other software that depends on Xerces-C, and the changes
introduced in the GP-Xerces patch make it incompatible with the upstream
version. Installing in a non-default prefix allows you to have GP-Xerces
installed side-by-side with unpatched Xerces without incompatibilities.
You can point cmake at your patched GP-Xerces installation using the
XERCES_INCLUDE_DIR and XERCES_LIBRARY options like so:
However, to use the current build scripts in GPDB, Xerces with the gp_xerces patch will need to be placed on the /usr path.
cmake -GNinja -D XERCES_INCLUDE_DIR=/opt/gp_xerces/include -D XERCES_LIBRARY=/opt/gp_xerces/lib/libxerces-c.so ..
Again, on Mac OS X, the library name will end with .dylib instead of .so.
Unless you intend to cross-compile a 32 or 64-bit version of GP-Orca, you can ignore these
instructions. If you need to explicitly compile for the 32 or 64-bit version of
your architecture, you need to set the CFLAGS and CXXFLAGS environment
variables for the configure script like so (use -m32 for 32-bit, -m64 for
64-bit):
CFLAGS="-m32" CXXFLAGS="-m32" ../configure --prefix=/opt/gp_xerces_32
For the most part you should not need to explicitly compile a 32-bit or 64-bit version of the optimizer libraries. By default, a "native" version for your host platform will be compiled. However, if you are on x86 and want to, for example, build a 32-bit version of Optimizer libraries on a 64-bit machine, you can do so as described below. Note that you will need a "multilib" C++ compiler that supports the -m32/-m64 switches, and you may also need to install 32-bit ("i386") versions of the C and C++ standard libraries for your OS. Finally, you will need to build 32-bit or 64-bit versions of GP-Xerces as appropriate.
Toolchain files for building 32 or 64-bit x86 libraries are located in the cmake directory. Here is an example of building for 32-bit x86:
cmake -GNinja -D CMAKE_TOOLCHAIN_FILE=../cmake/i386.toolchain.cmake ../
And for 64-bit x86:
cmake -GNinja -D CMAKE_TOOLCHAIN_FILE=../cmake/x86_64.toolchain.cmake ../
Show all command lines while building (for debugging purpose)
ninja -v -C build
Debug builds of GPORCA include a couple of "extended" tests for features like
fault-simulation and time-slicing that work by running the entire test suite
in combination with the feature being tested. These tests can take a long time
to run and are not enabled by default. To turn extended tests on, add the cmake
arguments -D ENABLE_EXTENDED_TESTS=1.
GPORCA has four libraries:
- libnaucrates --- has all DXL related classes, and statistics related classes
- libgpopt --- has all the code related to the optimization engine, meta-data accessor, logical / physical operators, transformation rules, and translators (DXL to expression and vice versa).
- libgpdbcost --- cost model for GPDB.
- libgpos --- abstraction of memory allocation, scheduling, error handling, and testing framework.
By default, GPORCA will be installed under /usr/local. You can change this by setting CMAKE_INSTALL_PREFIX when running cmake, for example:
cmake -GNinja -D CMAKE_INSTALL_PREFIX=/home/user/gporca -H. -Bbuild
By default, the header files are located in:
/usr/local/include/naucrates
/usr/local/include/gpdbcost
/usr/local/include/gpopt
/usr/local/include/gpos
the library is located at:
/usr/local/lib/libnaucrates.so*
/usr/local/lib/libgpdbcost.so*
/usr/local/lib/libgpopt.so*
/usr/local/lib/libgpos.so*
Build and install:
ninja install -C build
Note that because Red Hat-based systems do not normally look for shared
libraries in /usr/local/lib, it is suggested to add /usr/local/lib to the
/etc/ld.so.conf and run ldconfig to rebuild the shared library cache if
developing on one of these Linux distributions.
Remove the cmake files generated under build folder of gporca repo:
rm -fr build/*
Remove gporca header files and library, (assuming the default install prefix /usr/local)
rm -rf /usr/local/include/naucrates
rm -rf /usr/local/include/gpdbcost
rm -rf /usr/local/include/gpopt
rm -rf /usr/local/include/gpos
rm -rf /usr/local/lib/libnaucrates.so*
rm -rf /usr/local/lib/libgpdbcost.so*
rm -rf /usr/local/lib/libgpopt.so*
rm -rf /usr/local/lib/libgpos.so*
We accept contributions via Github Pull requests only.
Follow the steps below to open a PR:
- Fork the project’s repository
- Create your own feature branch (e.g.
git checkout -b better_orca) and make changes on this branch.- Follow the previous sections on this page to setup and build in your environment.
- Follow the naming and formatting style guide described here.
- Run through all the tests in your feature branch and ensure they are successful.
- Follow the Add tests section to add new tests.
- Follow the Update tests section to update existing tests.
- Make sure that ctest passes in both debug and retail build since there are some tests that do not overlap.
- Push your local branch to your fork (e.g.
git push origin better_orca) and submit a pull request- In some cases, ICG tests in GPDB may also need to be modified or additional ICG tests may need to be added to provide full coverage for the fix (for example, if the patch fixes wrong results or an execution specific issue). If such a situation occurs, please create a GPDB pull request and reference it in the GPORCA pull request.
Your contribution will be analyzed for product fit and engineering quality prior to merging.
Note: All contributions must be sent using GitHub Pull Requests.
Your pull request is much more likely to be accepted if it is small and focused with a clear message that conveys the intent of your change.
Overall we follow GPDB's comprehensive contribution policy. Please refer to it here for details.
Bump the GPORCA_VERSION_MINOR in CMakeLists.txt whenever your changes affect the ORCA functionality. GPORCA_VERSION_PATCH is bumped only in case where
the changes do not affect ORCA functionality e.g. updating the README.md, adding a test case, fixing comments etc.