[libvirt] [PATCH] travis: switch to using Ubuntu 16.04 and 18.04
The container images provided by Travis only support Ubuntu 14.04, however, Travis has ability to run docker, which allows the build script to use arbitrary OS images. This takes advantage of that to convert the build over to Ubuntu 16.04 and 18.04 This is using the official Ubuntu provided images and installing extra build deps required, as we previously did with Travis container images. With the switch to Docker though, this can be improved, by building custom Docker images with all the deps pre-installed which will cut down build time. This can be driven from the package lists in libvirt-jenkins-ci repo, to remove the duplication. This work for future improvement though, this just does the minimal conversion to match what we already do, but with newer distro. Signed-off-by: Daniel P. Berrangé <berrange@redhat.com> --- .travis.yml | 212 ++++++++++++++++++++++++++++------------------------ 1 file changed, 116 insertions(+), 96 deletions(-) diff --git a/.travis.yml b/.travis.yml index c71411dcae..15b5e933e9 100644 --- a/.travis.yml +++ b/.travis.yml @@ -1,6 +1,5 @@ -sudo: false -language: c -cache: ccache +sudo: required +language: generic branches: except: @@ -8,17 +7,38 @@ branches: matrix: include: - - compiler: gcc - dist: trusty + - services: + - docker env: - - PYTHON=$(which python2) - - DISTCHECK_CONFIGURE_FLAGS="--with-init-script=upstart" - - compiler: clang - dist: trusty + - DISTCHECK_FLAGS="--with-init-script=upstart" + before_install: + - docker pull berrange/test + - docker run + --privileged + -v `pwd`:/build + -w /build + -e VIR_TEST_DEBUG="$VIR_TEST_DEBUG" + -e PACKAGES="$PACKAGES" + -e DISTCHECK_FLAGS="$DISTCHECK_FLAGS" + ubuntu:18.04 + /bin/sh -c "$DOCKER_CMD" + - services: + - docker env: - - PYTHON=$(which python3) - - DISTCHECK_CONFIGURE_FLAGS="--with-init-script=systemd" + - DISTCHECK_FLAGS="--with-init-script=upstart" + before_install: + - docker pull berrange/test + - docker run + --privileged + -v `pwd`:/build + -w /build + -e VIR_TEST_DEBUG="$VIR_TEST_DEBUG" + -e PACKAGES="$PACKAGES" + -e DISTCHECK_FLAGS="$DISTCHECK_FLAGS" + ubuntu:16.04 + /bin/sh -c "$DOCKER_CMD" - compiler: clang + language: c os: osx env: - PATH="/usr/local/opt/gettext/bin:/usr/local/opt/ccache/libexec:/usr/local/opt/rpcgen/bin:$PATH" @@ -27,80 +47,18 @@ matrix: - brew update - brew upgrade - brew install python ccache rpcgen yajl + before_script: + - ./autogen.sh --prefix=$(pwd)/install-root script: # We can't run make distcheck/syntax-check because they # fail on macOS, but doing 'install' and 'dist' gives us # some useful coverage - make -j3 && make -j3 install && make -j3 dist + after_failure: + - echo '=== LOG FILE(S) START ===' + - find -name test-suite.log | xargs cat + - echo '=== LOG FILE(S) END ===' -addons: - apt: - # Please keep this list sorted alphabetically - packages: - - augeas-tools - - autoconf - - automake - - autopoint - - bash-completion - - ccache - - dnsmasq-base - - dwarves - - ebtables - - gcc - - gettext - - glusterfs-client - - libacl1-dev - - libapparmor-dev - - libattr1-dev - - libaudit-dev - - libavahi-client-dev - - libblkid-dev - - libc6-dev - - libcap-ng-dev - - libc-dev-bin - - libcurl4-gnutls-dev - - libdbus-1-dev - - libdevmapper-dev - - libfuse-dev - - libgnutls-dev - - libnetcf-dev - - libnl-3-dev - - libnl-route-3-dev - - libnuma-dev - - libopenwsman-dev - - libparted-dev - - libpcap-dev - - libpciaccess-dev - - librbd-dev - - libreadline-dev - - libsanlock-dev - - libsasl2-dev - - libselinux1-dev - - libssh2-1-dev - - libssh-dev - - libtirpc-dev - - libtool - - libudev-dev - - libxen-dev - - libxml2-dev - - libxml2-utils - - libyajl-dev - - lvm2 - - make -# - nfs-common # broken on trusty since build #807 (2018-01-09) - - open-iscsi - - parted - - patch - - perl - - pkgconf - - policykit-1 - - qemu-utils - - radvd - - scrub - - sheepdog - - systemtap-sdt-dev - - xsltproc - - zfs-fuse git: submodules: true @@ -108,23 +66,85 @@ git: env: global: - VIR_TEST_DEBUG=1 - -before_script: - - ./autogen.sh --prefix=$(pwd)/install-root - -script: - - make -j3 && make -j3 syntax-check && make -j3 distcheck DISTCHECK_CONFIGURE_FLAGS=$DISTCHECK_CONFIGURE_FLAGS - -after_failure: - - echo '============================================================================' - - 'ver=$(sed -n "s/AC_INIT(\[libvirt\], \[\([0-9.]*\)\],.*/\1/p" configure.ac); - if [ -f $(pwd)/tests/test-suite.log ]; then - cat $(pwd)/tests/test-suite.log; - elif [ -f libvirt-$ver/_build/tests/test-suite.log ]; then - cat libvirt-$ver/_build/tests/test-suite.log; - else - echo "=== NO LOG FILE FOUND ==="; - fi' + - DOCKER_CMD=" + apt-get update && + apt-get install -y \$PACKAGES && + ./autogen.sh --prefix=/build/install-root && + make -j 3 && + make -j 3 syntax-check && + make -j 3 distcheck DISTCHECK_CONFIGURE_FLAGS=\$DISTCHECK_FLAGS || + ( + echo '=== LOG FILE(S) START ===' + find -name test-suite.log | xargs cat + echo '=== LOG FILE(S) END ===' + exit 1 + ) + " + - PACKAGES=" + augeas-tools + autoconf + automake + autopoint + bash-completion + ccache + dnsmasq-base + dwarves + ebtables + gcc + gettext + git + glusterfs-client + libacl1-dev + libapparmor-dev + libattr1-dev + libaudit-dev + libavahi-client-dev + libblkid-dev + libc6-dev + libcap-ng-dev + libc-dev-bin + libdbus-1-dev + libdevmapper-dev + libfuse-dev + libgnutls28-dev + libnetcf-dev + libnl-3-dev + libnl-route-3-dev + libnuma-dev + libopenwsman-dev + libparted-dev + libpcap-dev + libpciaccess-dev + librbd-dev + libreadline-dev + libsanlock-dev + libsasl2-dev + libselinux1-dev + libssh2-1-dev + libssh-dev + libtirpc-dev + libtool + libudev-dev + libxen-dev + libxml2-dev + libxml2-utils + libyajl-dev + lvm2 + make + nfs-common + open-iscsi + parted + patch + perl + pkgconf + policykit-1 + qemu-utils + radvd + scrub + sheepdog + systemtap-sdt-dev + xsltproc + zfs-fuse" notifications: irc: -- 2.17.0
On Tue, 2018-06-05 at 18:47 +0100, Daniel P. Berrangé wrote:
The container images provided by Travis only support Ubuntu 14.04, however, Travis has ability to run docker, which allows the build script to use arbitrary OS images. This takes advantage of that to convert the build over to Ubuntu 16.04 and 18.04
Neat! I was not aware that was possible. TIL :)
This is using the official Ubuntu provided images and installing extra build deps required, as we previously did with Travis container images.
Everything after this doesn't belong in the commit message IMHO.
With the switch to Docker though, this can be improved, by building custom Docker images with all the deps pre-installed which will cut down build time. This can be driven from the package lists in libvirt-jenkins-ci repo, to remove the duplication. This work for future improvement though, this just does the minimal conversion to match what we already do, but with newer distro.
We will probably want to do that, since installing packages takes quite a bit of time and using Docker like this apparently causes jobs to serialize, which makes shaving down time all the more important. I'm wondering if using Ubuntu 16.04 and 18.04 as base image is the best option, however. Pavel is trying to get more hardware assigned to us in the CentOS CI environment, and assuming that's successful we will get those distributions added there as well for full-stack testing; Travis CI is more useful for developers to smoke test their patch series before posting, however, and with that in mind I think it would be much better to run the build on the latest CentOS, Fedora, Debian and Ubuntu to give reasonable coverage, leaving up to CentOS CI to catch issues related to interactions with other components or compatibility with older distros after merge. [...]
+ before_install:
Not sure 'before_install' is the best choice here, I would probably have gone with 'script' instead. On the other hand, it seems like the distinction between job phases is fairly blurry, so it probably doesn't matter after all.
+ - docker pull berrange/test
I'm pretty sure we don't want this :)
+ - docker run + --privileged
Is '--privileged' really needed here? I'm not familiar enough with Docker, but it feels like we shouldn't need it. Feel free to point out exactly why I'm wrong :)
+ -v `pwd`:/build
Please use $(pwd) instead of `pwd`.
+ -w /build + -e VIR_TEST_DEBUG="$VIR_TEST_DEBUG" + -e PACKAGES="$PACKAGES" + -e DISTCHECK_FLAGS="$DISTCHECK_FLAGS" + ubuntu:18.04
It would be great to have the base image name in the environment as well, so that in case the build fails on only one of the Ubuntu versions you could easily tell from a quick look at the build summary which one that is. [...]
env: - - PYTHON=$(which python3)
You got rid of this variable entirely. I think we won't lose any coverage because configure will naturally pick python2 on 16.04 and python3 on 18.04 (I can't check because Travis CI decided to only keep around a truncated log of my test run), but it was kinda nice to have that explicitly in the summary. As long as you can confirm the above is correct and no coverage is lost, though, I'm okay with the removal.
- - DISTCHECK_CONFIGURE_FLAGS="--with-init-script=systemd" + - DISTCHECK_FLAGS="--with-init-script=upstart"
You copied the wrong line here, we're going to test upstart twice. I don't see a good reason to change the variable name, however: let's just stick with matching the name expected by distcheck itself. [...]
- compiler: clang + language: c os: osx
You lost the ccache setting. That's probably okay because we install the package and set up PATH explicitly anyway. [...]
+ after_failure: + - echo '=== LOG FILE(S) START ===' + - find -name test-suite.log | xargs cat + - echo '=== LOG FILE(S) END ==='
This is simpler than the previous approach. I like it :) [...]
+ - DOCKER_CMD=" + apt-get update && + apt-get install -y \$PACKAGES && + ./autogen.sh --prefix=/build/install-root &&
Use --prefix=\$(pwd)/install-root here for consistency with the macOS part.
+ make -j 3 &&
No space between '-j' and '3'.
+ make -j 3 syntax-check && + make -j 3 distcheck DISTCHECK_CONFIGURE_FLAGS=\$DISTCHECK_FLAGS ||
This won't work if we have more than a single flag, will it?
+ ( + echo '=== LOG FILE(S) START ===' + find -name test-suite.log | xargs cat + echo '=== LOG FILE(S) END ===' + exit 1 + ) + " + - PACKAGES="
You lost the comment about this list being sorted alphabetically.
+ augeas-tools [...] + zfs-fuse"
Put the closing quote on a separate line for neater future diffs. -- Andrea Bolognani / Red Hat / Virtualization
On Wed, Jun 06, 2018 at 11:27:38AM +0200, Andrea Bolognani wrote:
On Tue, 2018-06-05 at 18:47 +0100, Daniel P. Berrangé wrote:
The container images provided by Travis only support Ubuntu 14.04, however, Travis has ability to run docker, which allows the build script to use arbitrary OS images. This takes advantage of that to convert the build over to Ubuntu 16.04 and 18.04
Neat! I was not aware that was possible. TIL :)
Yeah, nor was I until yesterday when I found someone saying this is the way to get off the ancient Ubuntu versions in Travis :-)
This is using the official Ubuntu provided images and installing extra build deps required, as we previously did with Travis container images.
Everything after this doesn't belong in the commit message IMHO.
With the switch to Docker though, this can be improved, by building custom Docker images with all the deps pre-installed which will cut down build time. This can be driven from the package lists in libvirt-jenkins-ci repo, to remove the duplication. This work for future improvement though, this just does the minimal conversion to match what we already do, but with newer distro.
We will probably want to do that, since installing packages takes quite a bit of time and using Docker like this apparently causes jobs to serialize, which makes shaving down time all the more important.
When I tested it, the jobs still ran in parallel. The key difference is that it causes the jobs to be scheduled on the VM based infra in Travis instead of the container based infra. There's a slightly longer startup time for jobs but otherwise nothing changes. The jobs still took approx 25-30 minutes to complete, which is the same as before this change. Pre-built docker images will definitely cut it down though, at least 5 minutes off perhaps more.
I'm wondering if using Ubuntu 16.04 and 18.04 as base image is the best option, however. Pavel is trying to get more hardware assigned to us in the CentOS CI environment, and assuming that's successful we will get those distributions added there as well for full-stack testing; Travis CI is more useful for developers to smoke test their patch series before posting, however, and with that in mind I think it would be much better to run the build on the latest CentOS, Fedora, Debian and Ubuntu to give reasonable coverage, leaving up to CentOS CI to catch issues related to interactions with other components or compatibility with older distros after merge.
Yeah, I just picked ubuntu as an initial target since it is close to what we have. I would like to evolve it to have similar options for coverage as the main CI. Ideally at least one of centos, ubuntu, fedora and mingw, so 5 jobs in total including macOS.
[...]
+ before_install:
Not sure 'before_install' is the best choice here, I would probably have gone with 'script' instead.
Yes, cut+paste mistake.
On the other hand, it seems like the distinction between job phases is fairly blurry, so it probably doesn't matter after all.
+ - docker pull berrange/test
I'm pretty sure we don't want this :)
Opps, yes, thats cruft from my testing custom images
+ - docker run + --privileged
Is '--privileged' really needed here? I'm not familiar enough with Docker, but it feels like we shouldn't need it. Feel free to point out exactly why I'm wrong :)
We need privileged because we have to apt-get install extra packages. Once we switch to pre-built images we can drop the --privileged arg.
+ -v `pwd`:/build
Please use $(pwd) instead of `pwd`.
+ -w /build + -e VIR_TEST_DEBUG="$VIR_TEST_DEBUG" + -e PACKAGES="$PACKAGES" + -e DISTCHECK_FLAGS="$DISTCHECK_FLAGS" + ubuntu:18.04
It would be great to have the base image name in the environment as well, so that in case the build fails on only one of the Ubuntu versions you could easily tell from a quick look at the build summary which one that is.
Yeah, I'll see how to get that visible
[...]
env: - - PYTHON=$(which python3)
You got rid of this variable entirely. I think we won't lose any coverage because configure will naturally pick python2 on 16.04 and python3 on 18.04 (I can't check because Travis CI decided to only keep around a truncated log of my test run), but it was kinda nice to have that explicitly in the summary. As long as you can confirm the above is correct and no coverage is lost, though, I'm okay with the removal.
Yeah, I'll check
- - DISTCHECK_CONFIGURE_FLAGS="--with-init-script=systemd" + - DISTCHECK_FLAGS="--with-init-script=upstart"
You copied the wrong line here, we're going to test upstart twice.
I don't see a good reason to change the variable name, however: let's just stick with matching the name expected by distcheck itself.
ok
[...]
- compiler: clang + language: c os: osx
You lost the ccache setting. That's probably okay because we install the package and set up PATH explicitly anyway.
The ccache setting applies to the "host", and so wouldn't have an effect in side the docker container. I guess it would still have affected ox-s though so I can re-add.
[...]
+ after_failure: + - echo '=== LOG FILE(S) START ===' + - find -name test-suite.log | xargs cat + - echo '=== LOG FILE(S) END ==='
This is simpler than the previous approach. I like it :)
Mostly because I couldn't figure out the right damn path to use. It does have the benefit though that if gnulib tests fail, we'll be catting the gnulib log file too.
[...]
+ - DOCKER_CMD=" + apt-get update && + apt-get install -y \$PACKAGES && + ./autogen.sh --prefix=/build/install-root &&
Use
--prefix=\$(pwd)/install-root
here for consistency with the macOS part.
ok
+ make -j 3 &&
No space between '-j' and '3'.
+ make -j 3 syntax-check && + make -j 3 distcheck DISTCHECK_CONFIGURE_FLAGS=\$DISTCHECK_FLAGS ||
This won't work if we have more than a single flag, will it?
Good point, i'll quote it
+ ( + echo '=== LOG FILE(S) START ===' + find -name test-suite.log | xargs cat + echo '=== LOG FILE(S) END ===' + exit 1 + ) + " + - PACKAGES="
You lost the comment about this list being sorted alphabetically.
ok
+ augeas-tools [...] + zfs-fuse"
Put the closing quote on a separate line for neater future diffs.
-- Andrea Bolognani / Red Hat / Virtualization
Regards, Daniel -- |: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|
On Wed, 2018-06-06 at 11:28 +0100, Daniel P. Berrangé wrote:
On Wed, Jun 06, 2018 at 11:27:38AM +0200, Andrea Bolognani wrote:
We will probably want to do that, since installing packages takes quite a bit of time and using Docker like this apparently causes jobs to serialize, which makes shaving down time all the more important.
When I tested it, the jobs still ran in parallel. The key difference is that it causes the jobs to be scheduled on the VM based infra in Travis instead of the container based infra. There's a slightly longer startup time for jobs but otherwise nothing changes. The jobs still took approx 25-30 minutes to complete, which is the same as before this change.
Doesn't match my experience: Ubuntu jobs took 30 minutes each and the second one only started once the first one was done, meaning the whole build took about one hour to finish. The macOS job was running at the same time as the first Ubuntu job. Might have been a coincidence, though.
I'm wondering if using Ubuntu 16.04 and 18.04 as base image is the best option, however. Pavel is trying to get more hardware assigned to us in the CentOS CI environment, and assuming that's successful we will get those distributions added there as well for full-stack testing; Travis CI is more useful for developers to smoke test their patch series before posting, however, and with that in mind I think it would be much better to run the build on the latest CentOS, Fedora, Debian and Ubuntu to give reasonable coverage, leaving up to CentOS CI to catch issues related to interactions with other components or compatibility with older distros after merge.
Yeah, I just picked ubuntu as an initial target since it is close to what we have. I would like to evolve it to have similar options for coverage as the main CI. Ideally at least one of centos, ubuntu, fedora and mingw, so 5 jobs in total including macOS.
I don't like the fact that you didn't include Debian :) Ideally we'd also throw openSUSE into the mix, after adding it to the CentOS CI, but as with Ubuntu that depends on getting acces to extra capacity.
+ - docker run + --privileged
Is '--privileged' really needed here? I'm not familiar enough with Docker, but it feels like we shouldn't need it. Feel free to point out exactly why I'm wrong :)
We need privileged because we have to apt-get install extra packages. Once we switch to pre-built images we can drop the --privileged arg.
Makes sense.
- compiler: clang + language: c os: osx
You lost the ccache setting. That's probably okay because we install the package and set up PATH explicitly anyway.
The ccache setting applies to the "host", and so wouldn't have an effect in side the docker container. I guess it would still have affected ox-s though so I can re-add.
As I said it's probably not strictly necessary, but it certainly won't hurt to keep it around. Your call. -- Andrea Bolognani / Red Hat / Virtualization
On Wed, Jun 06, 2018 at 01:01:05PM +0200, Andrea Bolognani wrote:
On Wed, 2018-06-06 at 11:28 +0100, Daniel P. Berrangé wrote:
On Wed, Jun 06, 2018 at 11:27:38AM +0200, Andrea Bolognani wrote:
We will probably want to do that, since installing packages takes quite a bit of time and using Docker like this apparently causes jobs to serialize, which makes shaving down time all the more important.
When I tested it, the jobs still ran in parallel. The key difference is that it causes the jobs to be scheduled on the VM based infra in Travis instead of the container based infra. There's a slightly longer startup time for jobs but otherwise nothing changes. The jobs still took approx 25-30 minutes to complete, which is the same as before this change.
Doesn't match my experience: Ubuntu jobs took 30 minutes each and the second one only started once the first one was done, meaning the whole build took about one hour to finish. The macOS job was running at the same time as the first Ubuntu job.
Might have been a coincidence, though.
Yeah, I think that is unrelated, because I'm definitely seeing both ubuntu jobs running at the same time. Probably travis just didn't have spare workers to spin up another job at the time you tested.
Yeah, I just picked ubuntu as an initial target since it is close to what we have. I would like to evolve it to have similar options for coverage as the main CI. Ideally at least one of centos, ubuntu, fedora and mingw, so 5 jobs in total including macOS.
I don't like the fact that you didn't include Debian :)
Ideally we'd also throw openSUSE into the mix, after adding it to the CentOS CI, but as with Ubuntu that depends on getting acces to extra capacity.
Sure I don't mind, though note that travis limits you to 5 concurrent jobs, so if we go over that, we effectively double the time to complete the test run as later jobs will have to wait. In terms of core project post-commit CI, the main value is the macOS job, as that's not overlapping with centos OS. The other jobs are largely there for letting developers test changes pre-posting for review/merge. So I don't think we need to go over the top and add every platform into the travis setup, just a reasonable set that will catch common problems before review. Let the centos CI handle the full matrix. Regards, Daniel -- |: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|
On Wed, Jun 06, 2018 at 11:28:27AM +0100, Daniel P. Berrangé wrote:
On Wed, Jun 06, 2018 at 11:27:38AM +0200, Andrea Bolognani wrote:
On Tue, 2018-06-05 at 18:47 +0100, Daniel P. Berrangé wrote:
The container images provided by Travis only support Ubuntu 14.04, however, Travis has ability to run docker, which allows the build script to use arbitrary OS images. This takes advantage of that to convert the build over to Ubuntu 16.04 and 18.04
Neat! I was not aware that was possible. TIL :)
Yeah, nor was I until yesterday when I found someone saying this is the way to get off the ancient Ubuntu versions in Travis :-)
Oh, maybe you just forgot, we had some improvement suggestions like a year ago but we missed some e-mails, I guess. Have a nice day, Martin
On Thu, 2018-06-07 at 09:52 +0200, Martin Kletzander wrote:
On Wed, Jun 06, 2018 at 11:28:27AM +0100, Daniel P. Berrangé wrote:
On Wed, Jun 06, 2018 at 11:27:38AM +0200, Andrea Bolognani wrote:
On Tue, 2018-06-05 at 18:47 +0100, Daniel P. Berrangé wrote:
The container images provided by Travis only support Ubuntu 14.04, however, Travis has ability to run docker, which allows the build script to use arbitrary OS images. This takes advantage of that to convert the build over to Ubuntu 16.04 and 18.04
Neat! I was not aware that was possible. TIL :)
Yeah, nor was I until yesterday when I found someone saying this is the way to get off the ancient Ubuntu versions in Travis :-)
Oh, maybe you just forgot, we had some improvement suggestions like a year ago but we missed some e-mails, I guess.
Sometimes stuff slips through the cracks ¯\_(ツ)_/¯ If you feel like fishing that out and checking whether any of the suggested improvements would still apply, that would be great; that said, outside of the minor issues pointed out during review, Dan's approach looks pretty solid to me. -- Andrea Bolognani / Red Hat / Virtualization
participants (3)
-
Andrea Bolognani -
Daniel P. Berrangé -
Martin Kletzander