As it turns out, certain metadata is only present in the ec2 flavor
of metadata (e.g. public_ipv4) and other data is only present in
the openstack flavor (e.g. network_config). For now, just read the
openstack metadata.
In some circumstances (e.g. nova-agent-watcher) cloudconfig files will
be created where the EtcdEnvironment is an empty map, and hence != nil.
If this is the case we should not do anything at all (because the user
hasn't explicitly asked us to configure etcd). This change standardises
behaviour with the check that we already do for FleetEnvironment.
To maintain the behavior of the coreos-setup-environment that has
started to move into cloudinit we need to write out /etc/environment
with the public and private addresses, if known. The file is updated so
that other contents are not replaced. This behavior is disabled entirely
if /etc/environment was written by a write_files entry.
This attempts to retrieve cloudconfigs from two sources: the meta-data
service, and the user-data service. If only one cloudconfig is found,
that is applied to the system. If both services return a cloudconfig,
the two are merged into a single cloudconfig which is then applied to
the system.
Only a subset of parameters are merged (because the meta-data service
currently only partially populates a cloudconfig). In the event of any
conflicts, parameters in the user-data cloudconfig take precedence over
those in the meta-data cloudconfig.
This uses the new MetadataService implementation to retrieve values for
$private_ipv4 and $public_ipv4 substitutions, instead of using
environment variables.
Move the old metadata-service datasource to url datasource. This new datasource
checks for the existance of meta-data.json and if it doesn't exist, walks the
meta-data directory to build a metadata blob.
Supports retries with exponential backoff as well as connection
timeouts and the ability to skip SSL/TLS verification.
This commit also refactors datasource and initialize packages
in order to use the new HTTP client.
This change creates a few simple interfaces for coreos-specific
configuration options and moves things to them wherever possible; so if
an option needs to write a file, or create a unit, it is acted on
exactly the same way as every other file/unit that needs to be written
during the cloud configuration process.
daemon-reload should be fixed now and the latest CoreOS with locksmith
is causing the etcd unit to get lazy-loaded before all the cloudinit
processes have finished configuring etcd via dropin files. In short,
the luck we were relying on to get by without daemon-reload has
officially run out. Cross your fingers!
This reverts commit 580460ff3f.
