On rare occasion, the ceph checks might not be 100% correct and the user
might decide to force an upgrade anyway.
Correctly, if we are in an upgrade, we perform `ok-to-stop` check for
each daemon, if one of them fails, we retry and eventually fail.
They are corner-cases where users would like to force the upgrade anyway.
Enhance the new CR property: "skipUpgradeChecks: true".
Closes: https://github.com/rook/rook/issues/3872
Signed-off-by: Sébastien Han <seb@redhat.com>
The CSI secret was only being created for new clusters.
Upgraded clusters did not have the secret generated. To support
upgraded clusters to use CSI, we always ensure the csi secret
exists when the orchestration runs.
Signed-off-by: travisn <tnielsen@redhat.com>
Several commands were not using the cli finalizer function which builds
up the right argument to run a ceph command. This was working on the
operator since it used to have an /etc/ceph/ceph.conf file so we could
ran the command without specifying the ceph conf and key paths.
However, we recently removed that and it silently broke this "already"
broken code :).
Signed-off-by: Sébastien Han <seb@redhat.com>
After further consideration and working with the feature for a few
weeks, the `configOverride` functionality seems to be more complexity
than is worth it for the expected benefit. Users still have the most
preferred option of setting configs in Ceph's config database using
`ceph cofig set` or the dashboard, and the override ConfigMap still
exists for fallback scenarios. The config database can still be used
by Rook for setting necessary configs, but the user has no way of
telling Rook to do so.
Signed-off-by: Blaine Gardner <blaine.gardner@suse.com>
Reduce the duplication of detecting the version of the external cluster
by refactoring to a single method that retrieves and validates the
version
Signed-off-by: travisn <tnielsen@redhat.com>
Add various upgrade tests so that the local cluster does not fail if
an upgrade is available on the external cluster.
Also, add better upgrades handling on update and on add. We now rely as
much as possible on a unique function that works both local and external
cluster 'detectAndValidateCephVersion()'. So the code will act the same
way wether a CR update is triggered or if the operator is restarted.
The health check now reports any topology changes on the external
cluster such as odd number of mons and low number. It will also detect
if the external cluster has been upgraded and will log that information,
helping the admin in his/her decision to trigger an image upgrade.
Finally, we prevent the creation of new CRs if the external cluster
version has a higher major number than the local one.
Closes: https://github.com/rook/rook/issues/3731
Signed-off-by: Sébastien Han <seb@redhat.com>
Allow users to specify overrides in the CephCluster CRD. The override
ConfigMap still exists for emergency situations and is mounted into
daemon pods directly instead of being merged into a Rook-created config
file.
Rook Ceph no longer uses a config file for managing daemons with the
exception of the OSDs which still generate a config in an init
container.
Signed-off-by: Blaine Gardner <blaine.gardner@suse.com>
From now on, Rook will only perform check before upgrades when there is
an actual upgrade. So if the Ceph image changed and a new version is
desired Rook will go through all the daemons and update them one by one
and perform checks in between.
Closes: https://github.com/rook/rook/issues/3583
Signed-off-by: Sébastien Han <seb@redhat.com>
this patch adds two new related features. first, it removes the
requirement that all monitor deployments be pinned to nodes using node
selector. this is allowed when not using host networking and when PVC
storage is used. the second feature introduced in this patch is the
elimination of the Rook scheduler for placing monitors that continue to
use the node selector by using canary pods to that are scheduled by k8s,
after which Rook uses the scheduling results to pin the monitors to
nodes.
fixes: #3667
Signed-off-by: Noah Watkins <noahwatkins@gmail.com>
ceph: Use the new NetworkSpec api
ceph: Change all tests to use the NetworkSpec
ceph: Remove TypeMeta field from NetworkSpec
ceph: Change IsHost network func signature
Signed-off-by: giovanism <giovanism@outlook.co.id>
saveMonConfig already calls WriteConnectionConfig so there is no need to
call WriteConnectionConfig again after saveMonConfig is done.
Signed-off-by: Sébastien Han <seb@redhat.com>
With this change, the admin can specify a single mon and the operator
will add any other mons to both ClusterInfo and the configmap of mon
endpoints.
It will periodically check for changes too and update both ClusterInfo
and the configmap accordingly.
Signed-off-by: Sébastien Han <seb@redhat.com>
By exporting cluster to Cluster we can now reach it from the controller
health check so that it can be populated with 'cluster' data, see the
next commit for better understanding.
Signed-off-by: Sébastien Han <seb@redhat.com>
When configuring an external cluster the orchestration of discover and all the ceph
daemons will be skipped. The operator will still watch for the creation of
crds for this namespace. When a filesystem, object, object user, or
ganesha crd are created, the operator will simply print an error
to the log.
Signed-off-by: travisn <tnielsen@redhat.com>
Signed-off-by: Sébastien Han <seb@redhat.com>
With the 1.1 release we should not be creating bluestore OSDs
with rook's legacy partitioning scheme. All new bluestore OSDs
should be created with ceph-volume. Therefore, the miminum
version allowed is a version of mimic supporting ceph-volume.
This commit also adds consistent version checking between
the Add and Update methods for the operator handling the
CR events.
Signed-off-by: travisn <tnielsen@redhat.com>
Instead of requring csi instances for every ceph cluster that rook
manages, maintain one config map for all ceph clusters. This allows
for one set of csi drivers per rook operator.
The config is intialized as empty when Ceph CSI components are first
created and then updated every time a cluster's mons change.
As before clusters are identified by the k8s namespace the ceph cluster
is running in.
Signed-off-by: John Mulligan <jmulligan@redhat.com>
Create a mutex at the rook operator level that will be used when
updating the shared-between-clusters config map that Ceph CSI consumes.
Pass the mutex down to the function that will update the CM.
Signed-off-by: John Mulligan <jmulligan@redhat.com>
Add a new function that, given existing config string, key, and current
mons, will return an updated version of the Ceph CSI config string
having updated the data with the new mons.
Add a unit test function for the update function.
Signed-off-by: John Mulligan <jmulligan@redhat.com>
when the root of a filesystem mode pvc is mounted as the base data directory
for the monitor it will not be empty: `lost+found` from ext4 will be present.
this causes the mon mkfs to fail because it expects an empty directory.
the solution is to mount under a directory in the source volume.
Signed-off-by: Noah Watkins <noahwatkins@gmail.com>
the life cycle post-start hook is used to chown the ceph daemon log
directory. however, hooks are not guaranteed to run before a container
entry point.
"Kubernetes sends the postStart event immediately after the Container is
created. There is no guarantee, however, that the postStart handler is
called before the Container’s entrypoint is called"
Signed-off-by: Noah Watkins <noahwatkins@gmail.com>
The CSI driver requires the ceph key for administrating
the cluster. The key is stored in a secret and specified in the
storage class. The operator will now automatically generate
a secret named rook-ceph-csi that can be specified by
default in the storage class instead of requiring the
admin to create it.
Signed-off-by: travisn <tnielsen@redhat.com>
a helper flag indicating the exitence of an empty zone (empty = no
monitor pods) is computed when checking the health of the cluster in
order to determine if mons from overloaded zones have some new location
they can be scheduled. the computation of this flag did not check that
the empty zone contains valid nodes that pods can be scheduled on to.
this lead to a mon creation loop.
fixes: #3545
Signed-off-by: Noah Watkins <noahwatkins@gmail.com>
Before each pod starts, the daemon changes the ownership of
/var/log/ceph which is bindmounted on a hostPath.
On OpenShift envionments, the container must be privileged in order to
successfully change the owernship.
Closes: https://github.com/rook/rook/issues/3485
Signed-off-by: Sébastien Han <seb@redhat.com>
OpenShift environment requires privileged containers when interacting
with the host filesystem, typically via dataDirHostPath.
The init container that changes the ownership of dataDirHostPath was
missing the privileged flag and users were getting permission denied
errors on the chown command.
Fixes: https://github.com/rook/rook/issues/3485
Signed-off-by: Sébastien Han <seb@redhat.com>
At the beginning of every container startup we now change the ownership
of every known ceph directories to ceph:ceph.
This will allow daemons to start properly as well as allowing daemon to
writes their logs on files.
This is needed because ceph daemon now run as 'ceph' instead of 'root'
Signed-off-by: Sébastien Han <seb@redhat.com>
We do not determine the daemon type and name by string parsing anymore
but passing that info directly in the update call, it's more reliable.
Also removing a bunch of unused functions.
Signed-off-by: Sébastien Han <seb@redhat.com>
This commit makes OnAdd() and OnUpdate() consistent in behavior. They
will do their best detecting if we can upgrade, continue an upgrade or
refuse to do one. It also refuses to downgrade to any version.
It also load ClusterInfo earlier in the controller so that we can
compare the spec version versus the cluster running one.
Signed-off-by: Sébastien Han <seb@redhat.com>
Do not consider clusterInfo Ceph version since it's the one picked by
the spec image.
The CephVersion provided here needs to be the current running version
of Ceph for when it is later passed to OkToStop and checked with
IsAtLeast. If you are upgrading from a pre 14.2.1 version to a
version 14.2.1 or newer, IsAtLeast will incorrectly return true and
then fail trying to run mon ok-to-stop on a mon pod running a pre 14.2.1 version.
Signed-off-by: Sébastien Han <seb@redhat.com>
When a cluster is updated with a different image version, this triggers
a serialized restart of all the pods. Prior to this commit, no safety
check were performed and rook was hoping for the best outcome.
Now before doing restarting a daemon we check it can be restarted. Once
it's restarted we also check we can pursue with the rest of the
platform. For instance, with monitors we check that they are in quorum,
for OSD we check that PGs are clean and for MDS we make sure they are
all active.
Fixes: https://github.com/rook/rook/issues/2889
Signed-off-by: Sébastien Han <seb@redhat.com>
this patch introduces failure-domain aware scheduling for monitor pods.
previously the only policy was to avoid scheduling monitors on the same
nodes. the new policy tries to spread monitors across zones and nodes.
NOTE: this patch only enforces the well-known k8s "zone"
failure-domain label (i.e. it does not look for region labels). this
decision is made for two reasons. first, eliminating explicit
scheduling is an important goal and will be done as part of upcoming
work. this will enable all supported failure-domain labels to be
handled. second, the "zone" domain is likely more appropriate for a
larger portion of users (as opposed to the geographical distinction
made with region), and supporting multiple failure domain types would
add significant complexity for a temporary resolution to failure
domain scheduling.
at a high-level this patch introduces a data structure `NodeUsage` that
is a pairing of a v1.Node and metadata relevant to scheduling: number of
monitor pods on the node, and its status w.r.t. to being scheduable for
new monitor pods.
each scheduling event computes a unified data structure that organizes
for each node a `NodeUsage` structure into per-failure-zone groups.
subsequent algorithms rely soley on this unified structure to make
scheduling decisions.
there are two entry points to the changes made:
1) during orchestration mon:Cluster:assignMons is invoked to make a
place new monitor pods onto nodes. the core scheduling logic is now
self-contained in mon:scheduleMonitor which implements node and zone
aware scheduling policies.
2) periodic health checks resolve any conflicts that may occur due to
changes to the cluster (e.g. new nodes, configuration changes, etc...).
the two existing checks: mons on invalid nodes, and overloaded nodes are
both handled.
fixes: #2603
Signed-off-by: Noah Watkins <noahwatkins@gmail.com>
The check for whether to set the OwnerReferences is no longer
necessary. When passing the correct value to the apiVersion,
the OwnerReferences can be successfully set for OpenShift.
Signed-off-by: travisn <tnielsen@redhat.com>
Signed-off-by: Guy Margalit <guymguym@gmail.com>
Co-Authored-By: Sébastien Han <seb@redhat.com>
Co-Authored-By: Travis Nielsen <tnielsen@redhat.com>
This change is meant to allow running operators locally on a developer machine.
The idea is to allow faster development cycles by reducing the time and complexity of building -> deploying -> debugging on cluster.
For operators that rely only on kubernetes API this works easily - see cockroachdb and minio examples in development-flow doc.
The change includes:
- rook.NewContext() - Refactored to remove repeating initialization code that was copy-pasted in most of the operators in order to create the clusterd.Context and the Clientsets. Also it detects the mode of working in-cluster vs external and sets up the external mode with standard user config (~/.kube/config) and a job executor.
- rook.GetOperatorImage() - Refactor this repeating code in many operators to detect the operator pod image. Also added a global flag --operator-image that developers can use to override this when running locally.
- rook.TerminateOnError() - Added a convenient function.
Create and maintain a config map that meets the requirements of
the ceph csi such that Rook can maintain the contents of config
map with up-to-date mon information to be used later by csi.
Signed-off-by: John Mulligan <jmulligan@redhat.com>
The mons require the --public-bind-addr when host networking is not in use
since the pod ip will be different from the service ip that will be
advertised to clients. When host networking is enabled, there is no
need for this argument since the endpoints will be the same to bind
inside the pod as what is advertised to clients.
Signed-off-by: travisn <tnielsen@redhat.com>
Clusters using host networking will not work if the non-default port is
being used by the mons. Previous to rook 1.0 the mons were all
using the non-default port 6790 and now they are using the default port
of 6789 from 1.0. Now host networking will preserve the non-default port
to allow upgrades to continue working.
Signed-off-by: travisn <tnielsen@redhat.com>
this adds a structure to hold various settings related to executing ceph
cli commands, and introduces a common interface for configuring and
running such commands.
Signed-off-by: Noah Watkins <noahwatkins@gmail.com>
This commits allows the upgrade from Mimic to Nautilus to work by:
* removing the v2 brackets on the operator ceph config file generation,
we stick with v1 and can revert this back once we deprecate mimic there
is no rush since mons keep on listening to v1.
* enable messengers 2 when the cluster runs on Nautilus
Fixes: https://github.com/rook/rook/issues/2973
Signed-off-by: Sébastien Han <seb@redhat.com>
Add a 'ceph-version' label to application controllers in the same manner
as the prior 'rook-version' label to help with upgrades.
Signed-off-by: Blaine Gardner <blaine.gardner@suse.com>
Sébastien Han
Travis Nielsen
Blaine Gardner
Noah Watkins
giovanism
John Mulligan
Guy Margalit
Louis