This is to better reflect the intent of the user when node ports are
specified explicitly (as opposed to being assigned by Kubernetes). The
`last-applied-configuration` annotation added by `kubectl apply` is one
such indicator we are now leveraging.
We still default to omitting the node ports when the annotation is
missing.
Signed-off-by: Timo Reimann <ttr314@googlemail.com>
Refactor plugin management:
- support multiple plugins per executable
- support restarting a plugin process in the event it terminates
- simplify plugin lifecycle management by using separate managers for
each scope (server vs backup vs restore)
Signed-off-by: Andy Goldstein <andy.goldstein@gmail.com>
Mirror pods are pods created from static manifest files on a node.
They're mirrored to the apiserver so they're visible when querying the
apiserver for a list of pods, but it's not possible to send a pod
containing the mirror pod annotation to the apiserver and have it be
created successfully. Instead of trying to do this, log a message that
we're skipping restoring the pod because it's a mirror pod.
Signed-off-by: Andy Goldstein <andy.goldstein@gmail.com>
If a PV has a reclaim policy of Delete and we didn't create a snapshot
of it, don't restore the PV, as doing so would create a PV whose
underlying volume is incorrect.
Also "reset" any PVCs bound to the PV so they'll be dynamically
provisioned when restored.
Signed-off-by: Andy Goldstein <andy.goldstein@gmail.com>
All properties from a backup will be merged into the ServiceAccount
except for the default token secret.
Signed-off-by: Nolan Brubaker <nolan@heptio.com>
Completed jobs and pods may be useful in the backup for auditing
purposes, but don't recreate them when restoring.
Signed-off-by: Nolan Brubaker <nolan@heptio.com>
When restoring resources that raise an already exists error, check their
equality before logging a message on the restore. If they're the same
except for some metadata, don't generate a message.
The restore process was modified so that if an object had an empty
namespace string, no namespace key is created on the object. This was to
avoid manipulating the copy of the current cluster's object by adding
the target namespace.
There are some cases right now that are known to not be equal via this
method:
- The `default` ServiceAccount in a namespace will not match, primarily
because of differing default tokens. These will be handled in their own
patch
- IP addresses for Services are recorded in the backup object, but are
either not present on the cluster object, or different. An issue for
this already exists at https://github.com/heptio/ark/issues/354
- Endpoints have differing values for `renewTime`. This may be
insubstantial, but isn't currently handled by the resetMetadataAndStatus
function.
- PersistentVolume objects do not match on spec fields, such as
claimRef and cloud provider persistent disk info
Signed-off-by: Nolan Brubaker <nolan@heptio.com>
The main Ark code was hard-coding specific support for AWS, GCE, and
Azure volume snapshots and restores, and anything else was considered
unsupported.
Add GetVolumeID and SetVolumeID to the BlockStore interface, to allow
block store plugins to handle volume snapshots and restores.
Signed-off-by: Andy Goldstein <andy.goldstein@gmail.com>
Deleting the clusterIP field when the service should be headless will
cause it to be assigned a new IP on restore; instead it should retain
the headless state after restoration.
Fixes#168
Signed-off-by: Nolan Brubaker <nolan@heptio.com>
