We use a plain ubuntu container as a build environment for ethtool &
conntrack, so a) we don't require onerous tooling on the build host,
and b) don't pollute the build host with build artefacts. The latter
arise because I can only get conntrack to build by installing its
dependencies 'globally'.
Th dir containing the build script is volume-mounted into the
container, and that is where we also expect to see the result, namely
a 'bin' sub-dir containing two executables.
The weavetools image is then constructed from the 'scratch' image and
contains just that dir.
So we complain with "already running" rather than "Found another
running container" when encountering a 'weave' container running a
different image version. Similarly, we remove stopped 'weave'
containers regardless of version.
Ditto for weavedns.
Which is all how it used to work.
- released weave scripts run their corresponding image versions
- unreleased weave scripts run the 'latest'-tagged image versions
The image version to run can also be specified explicitly by setting
VERSION. Useful for testing, mainly.
We would like to have a stable URL for the most recently released
script, so that we can mention it in the README and people can use it
in update scripts.
Since GitHub's releases/latest URL redirects, and anyway doesn't work
for downloads, we have to create (possibly delete first) a release
tagged "latest_release" and update it.
and return an error status rather than exit when containers stop just
before an HTTP call, which in turn allows us to make tell_dns properly
ignore weavedns container absences.
Fixes#300
Rather than looking for the host end of the veth pair, we look for the
expected interface inside the container. This has the advantage of
detecting weave-attached-ness even when the container is sharing a
network namespace with others.
Closes#303, #302.
Matthias Radestock