This branch defines cubridci/cubridci:develop, the image that compiles CUBRID and runs the
medium and sql tests on the build. For the other test categories, use the
test_rl8.10 branch and its image.
The image definition is docker/ci/ — Dockerfile and docker-entrypoint.sh.
Jenkins multibranch on ci.cubrid.org builds docker/ci on every push to this branch and
pushes the result as cubridci/cubridci:develop. The tag appears on Docker Hub about a
minute after the push. A new branch needs no job registration — Jenkins picks it up on its own.
Jenkins itself is reachable only from the internal network.
| Base | centos:6.10; packages from vault.centos.org |
| Size | 1.73 GB on disk, 785 MB to pull |
| Toolchain | gcc/g++ 8.3.1 and make from devtoolset-8, git 2.12 from sclo-git212, cmake 3.26.3, ninja 1.11.1, bison 3.0.5, flex, ant, ccache 3.7.7 |
| JDK | Temurin 8u442-b06 at /opt/jdk8 (JAVA_HOME); ant pulls in java-1.7.0-openjdk-devel, which is /usr/bin/java and where find_package(JNI) finds the headers |
| Locale, TZ | en_US, Asia/Seoul; ko_KR.UTF-8, ko_KR.EUC-KR and en_US.UTF-8 generated for the test cases |
CC and CXX are preset to gcc and g++. ccache is installed but not used by default; set
CC="ccache gcc" and CXX="ccache g++" to build through it.
devtoolset-8 and sclo-git212 are on PATH through the image environment rather than
scl enable, so every process in the container gets them, whatever starts it.
Four tools come from upstream instead of a repository: cmake, because base carries only 2.8.12;
bison, because base carries only 2.4.1 and the parser needs 3.x; ninja, which base does not
carry; and ccache, because EPEL 6 carries only 3.1.6, which compiles the preprocessed source
and so loses the /* fall through */ comments that -Werror=implicit-fallthrough relies on.
The entrypoint takes a verb. With no argument, or a first argument that starts with -, it
builds. Anything else that is not a verb runs as a command.
/entrypoint.sh checkout [<ref>]
/entrypoint.sh [build] [<build.sh args>...] # build is the default verb
/entrypoint.sh dist [<build.sh args>...]
/entrypoint.sh coverage [<build.sh args>...]
/entrypoint.sh test
/entrypoint.sh <command> [<args>...]
build, dist and coverage look for build.sh in the working directory, then in cubrid/,
and run ./build.sh -p $CUBRID <args> clean build or ./build.sh -p $CUBRID <args> dist. Extra
arguments pass straight through to build.sh; run ./build.sh -h for the list.
| Source tree | checkout puts it in ./cubrid; mounting your own at /home/cubrid also works |
| Install tree | $CUBRID = /home/CUBRID |
| Packages | dist -o <dir> writes the packages into <dir> |
| gcov archives | coverage writes them into $GCOV_OUTPUT_DIR (default: the working directory) |
| Logs | build.log and dist.log in the working directory |
The exit code is 0 on success and non-zero on failure. On a build failure the last 500 lines of
build.log are printed.
checkout clones CUBRID and its three submodules into ./cubrid, relative to the working
directory, which is where build looks second. Bring your own source instead if you prefer —
mount it and skip this verb. CI usually does, because the pipeline checks the source out before
the container starts.
| Which ref | the argument, else $CUBRID_REF, else develop. A branch, a tag or a full 40-character SHA. |
| Submodules | all three, at the commits the ref pins |
| Re-running | idempotent — it fetches, resets hard, and cleans the tree and the submodules |
| On failure | five attempts with a growing delay, then exit 1 |
/entrypoint.sh checkout # develop
/entrypoint.sh checkout release/11.4 # a branch
/entrypoint.sh checkout v11.4.5 # a tag
/entrypoint.sh checkout 138f4881243f4dfe107bf5e11969979b7fd5c33f # a commitThe clone is never shallow. build.sh reads the build serial off
git rev-list --count HEAD, so a --depth 1 tree stamps every package 0001. A full clone
gives the real number — 11.5.0.2496-138f488.
Set BUILD_MIRROR to a directory of bare clones and checkout borrows their objects with
git clone --shared instead of pulling ~471 MB over the network. The mirror is only ever read.
docker run --rm \
-v /home/build-cache/cubrid-mirror:/mirror:ro \
-e BUILD_MIRROR=/mirror \
cubridci/cubridci:develop checkoutBUILD_MIRROR set with no mirror there is an error, not a fallback — a silent 471 MB clone
would hide the missing mount. Submodules are borrowed only where the mirror already holds the
pinned commit; one bumped after the mirror was seeded comes from GitHub.
A borrowed tree needs the mirror on every later run. --shared records the mirror in
.git/objects/info/alternates, so the history is only readable while it stays mounted. Without
it build.sh cannot count the commits, swallows the error, and stamps the package
11.5.0.-<hash> — so build and dist check the history first and refuse to start rather than
producing that quietly. Mount the mirror for the build too, or check out without one.
test runs CTP on the build installed at $CUBRID. It takes no arguments; the environment
sets what runs.
TEST_SUITE |
the suites, colon-separated. Default medium:sql. |
BRANCH_TESTTOOLS |
branch of cubrid-testtools to check out. Default develop. |
BRANCH_TESTCASES |
branch of cubrid-testcases to check out. Default develop. |
TEST_REPORT |
directory that receives CTP's JUnit XML. Default /tmp/tests. |
Each run first checks both repositories out under /home: a clone the first time, a fetch and
reset after that. The suites then run in turn, and one that CTP itself
fails stops the rest. With medium or sql in TEST_SUITE, the verdict reads the
summary_info files CTP wrote in this run, and this run's JUnit XML goes to TEST_REPORT as
the report. That code is the same as in test_rl8.10, so a result reads the same here and in
gha-ci, and a second run in the same /home does not count the first run's failures. The exit
code is 1 when a case failed, nothing ran, or CTP failed; the failed cases are listed at the end.
docker run --rm \
-v "$PWD/workspace:/home" \
-v "$PWD/reports:/tmp/tests" \
cubridci/cubridci:develop \
bash -lc '/entrypoint.sh checkout && /entrypoint.sh build && /entrypoint.sh test'Check the source out and build it, keeping both in a mounted workspace:
docker run --rm \
-v "$PWD/workspace:/home" \
-e MAKEFLAGS=-j \
cubridci/cubridci:develop \
bash -lc '/entrypoint.sh checkout && /entrypoint.sh build'MAKEFLAGS=-j is what the CUBRID Jenkinsfile passes.
With source you already have on the host, mount it and skip checkout:
docker run --rm \
-v "$PWD/cubrid:/home/cubrid" \
-e MAKEFLAGS=-j \
cubridci/cubridci:develop buildgit here is 2.12, older than the ownership check git 2.35.2 added, so a mounted tree reads as
it is, whoever owns it, and needs no safe.directory.
An unreadable history — a tree checked out with BUILD_MIRROR and built without that mirror —
stops at a guard before the build, which quotes git. Its hint about safe.directory is shared
with build_rl8.10 and does not apply here.
To pack release artifacts as well:
docker run --rm \
-v "$PWD/workspace:/home" \
-v "$PWD/packages:/packages" \
cubridci/cubridci:develop dist -o /packagescoverage makes a -m coverage build and packs the two archives the
code coverage guide
asks for. -m coverage is forced, so anything else the caller passes for -m is overridden.
docker run --rm \
-v "$PWD/workspace:/home" \
-v "$PWD/gcov:/gcov" \
-e GCOV_OUTPUT_DIR=/gcov \
cubridci/cubridci:develop \
bash -lc '/entrypoint.sh checkout && /entrypoint.sh coverage'| Archive | Holds |
|---|---|
CUBRID-<version>-gcov-Linux.x86_64.tar.gz |
the install tree, top-level directory CUBRID |
cubrid-<version>-gcov-src-Linux.x86_64.tar.gz |
the source tree with the .gcno and the objects, top-level directory cubrid |
The nightly regression builds it the same way — CTP's common/ext/run_coverage.sh runs
build.sh -m coverage — and packages it with two plain tar calls, which is what this does.
dist cannot package it and is not the thing to fix: it refuses -m coverage outright, and its
source package is built from git ls-files, which by construction leaves out both the .gcno
files and the build directory .gitignore hides.
Unpack the source archive on the test node so its tree lands at the path it was built at.
The .gcda paths are compiled into the binaries. The build and test images both work under
/home, so a checkout here and a tar -C /home -xzf <source archive> there line up exactly,
and lcov then needs neither GCOV_PREFIX nor geninfo_adjust_src_path. coverage prints the
path it built at and the tar line to use.
The names are the nightly's; the layout is not. It tars the trees from the inside, with no
top-level directory, and run_cubrid_install unpacks them into a cubrid-<build id> of its own
making. Keeping the tree's real directory name instead is what lets the paths line up. The
consumer here is the test image, not that installer.
GCOV_OUTPUT_DIR has to be outside the source tree — tar cannot archive a directory it is
writing into. The default working directory /home is fine, since the tree sits at
/home/cubrid; a source tree mounted at /home itself needs this set somewhere else.
Leave room on whatever volume $GCOV_OUTPUT_DIR points at. A full develop build on
build_rl8.10 measured 241 MB for the install archive and 1.1 GB for the source one, from a
3.3 GB tree; the object files are 1.3 GB of that and the .gcno only 222 MB. The build itself
took about 2.5 hours at CMAKE_BUILD_PARALLEL_LEVEL=32, most of it the third-party
dependencies, which build serially.
cc1plus needs roughly 1 GB per compile job. An unbounded -j on a host or cgroup with only
a few GB gets the compiler OOM-killed part way through the build. Cap the job count instead —
build.sh drives the build through cmake, so CMAKE_BUILD_PARALLEL_LEVEL applies to both the
ninja and the make generator:
docker run --rm \
-v "$PWD/workspace:/home" \
-e CMAKE_BUILD_PARALLEL_LEVEL=4 \
cubridci/cubridci:develop buildOne job per GB of memory available to the container is a safe starting point.