The OSD daemon has been missing critical flags for logging to stderr
where k8s can capture the logs. Without the --log-to-stderr=true,
all the OSD logging was essentially lost until now.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
Update the base operator image and cluster examples to use the latest
octopus release v15.2.4. Also cleanup some old examples and unit tests
that were still based on mimic.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
The stuck mons were only being force deleted in the mon health checker.
After an operator restart, the health check may not be executed
for 20 minutes until after the timeout during the main reconcile
which is holding onto a lock for mon reconcile. Now the stuck mon
will be deleted during the reconcile as well.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
If the mon is backed by a pvc, ensure the pvc is deleted when
the mon is failed over. The pvc is being left in a bound state
even though it will no longer be used in the cluster.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
During mon failover if the mon fails to be removed, the new mon
config was not being saved. The mon removal anyway is not retried
if it fails, so it should make a best effort to remove all the
resources related to the mon, and then always update the mon
config after the failover.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
This commit fixes an argument of the CreateCrashCollectorSecret function as
it is expected to be the namespace of the cluster.
Signed-off-by: binoue <banji-inoue@cybozu.co.jp>
When the operator loads the mons and checks for a valid maxMonId
the check of the id against the currently running mons was outdated
from legacy mon ids. The check now has to parse the new id which is
just the letter instead of having a number from legacy mon ids.
Without this fix, the maxMonId would stay incorrect if it was somehow
corrupted.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
We can now explicitly set any property on a given pool by using the new
Property field in the CephBlockPool Spec.
Also, this fixes the case where both `CephBlockPool` and `CephCluster`
are created at the same time. When Rook creates the pool, the cluster is
still being bootstrapped and the global option
`osd_pool_default_pg_autoscale_mode` has not bee set yet. So the pool
gets created but its `pg_autoscale_mode` property is set to `warn`
instead of `on`.
Closes: https://github.com/rook/rook/issues/5608V
Signed-off-by: Sébastien Han <seb@redhat.com>
The volume sources for an OSD on PVC were relying on the ordering
of the volume sources slice. While this was working, it was causing
headaches and potential future developer error when being
interpreted incorrectly. The addition of the wal pvc device
would make the issue even more pronounced. Now each PVC used
for the OSD will be identified in a map by name such as data
and metadata.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
This commit contains CR implementations for:
CephObjectRealm
CephObjectZoneGroup
CephObjectZone
Also there are changes made to the objectstore
to add rgws in the object-store to zones and
zone groups in a multisite configuration and
the removal of the --default parameter for any
realms/zonegroups/zones that are created.
Signed-off-by: Ali Maredia <amaredia@redhat.com>
When running Octopus, we still need to turn on the balancer module. The
distinction is tricky, the module is "enabled" but it has to be "turned
on" and configured.
This was only working when the `pg_autoscaler` module was explicitly
listed in the CephCluster spec mgr section module section but not
automatically configured when running Octopus (the `pg_autoscaler`
module is already enabled).
Signed-off-by: Sébastien Han <seb@redhat.com>
The mgr crash module is always enabled so we don't need to run the
command again. It saves some orchestration time.
Signed-off-by: Sébastien Han <seb@redhat.com>
When the PVC template increases in size for the storageClassDeviceSet
in the cluster CR, we need to increase the size of the PVC and restart
the OSD. The OSD is restarted by changing a variable on the pod spec
that indicates the desired size of the OSD.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
When integrating rook with istio, service port names need to be prefixed by their protocol.
Trying to fix this by removing static port names in favor of configurable names.
Closes: #5587
add protocol prefix to service port names, hardcoded as discussed. renamed container ports
Signed-off-by: Alex Flom <alexander.flom@gmail.com>
When the cluster is using HostNetworking, Rook was either picking up the
external or internal IP of the node. This resulted in mon endpoints have
public IP addresses. Those IP are not reachable from within the cluster
so OSD/CSI couldn't access the monitors from the configmap endpoint.
Also, exposing the cluster on a public network does not seem realistic,
so sticky with private/internal IP addresses is better.
Closes: https://github.com/rook/rook/issues/5495
Signed-off-by: Sébastien Han <seb@redhat.com>
If a mon is stuck in terminating state on a failed node,
force delete the pod so it can attempt to restart. If
the mon is backed by a PV the mon can start on another
node as well.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
The operator should only print helpful info messages
when an OSD is going to be updated. If the OSD hasn't
changed it is just a debug message.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
The osd health check every 5 min is not frequent enough to respond
quickly enough to osds going down. Reducing the time interval
to one minute for faster response time.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
The pod deletion is needed by other daemons besides the osds,
so we move the helper method into the k8sutil package.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
Now, we can not only cleanup monitor data, logs and crashes but the
disks too. As part of the cleanupPolicy CR spec, we have a new setting
called sanitizeDisks which holds more details:
* confirmation: the confirmation message to sanitize disks,
use "yes-really-sanitize-disks" to confirm
This will **only** wipe the metadata, so it's a fast cleanup allowing
you to re-install later but won't remove all the data from the drive.
Signed-off-by: Sébastien Han <seb@redhat.com>
The crash collector was missing the security context to properly run on
OCP, thus crashing when trying to post a crash dump.
Signed-off-by: Sébastien Han <seb@redhat.com>
As of Nautilus Ceph auto-tunes its osd_memory_target on the fly so we don't need to force it.
ROok does not support Mimic anymore so this can be removed.
Signed-off-by: Sébastien Han <seb@redhat.com>
In 1.3 we added support for the metadata pvcs, thus requiring the name
to be set on the templates to differentiate between data and metadata.
For backward compatibility we need to assume the template is for data
if it is not set.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
During cluster delection, the finalizer will not be removed until
all the PVCs are removed. Since this could wait for a long time,
as long as the deletion event is re-queued, we need to cancel
the goroutine that is watching for cluster deletion so it can
avoid running multiple goroutines as the event runs again
and again.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
When a cluster is requested for deletion there is no longer
a reason to monitor the cluster health such as whether mons
are in quorum, osds are up and in, or monitor the ceph status.
The cluster will wait potentially for a long time if the
PVCs are not deleted by the admin, therefore we stop monitoring
the cluster even before the finalizer is removed.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
If the admin key is available, set the external creds struct
to use that admin username and key instead of requiring the
callers to check every time if they should use the admin
key or the external key.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
The Ceph version information is not needed to be passed
to the osd health checker since mimic support was dropped.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
The ceph username is all that needs to be passed to the goroutines
that are monitoring the health of Ceph. This simplifies the
initialization of the goroutines.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
The monitoring goroutines that check for mon and osd health,
the ceph cluster status, and the bucket provisioner should
only be started once for each cluster. Previous to this change
and since the controller runtime conversion the goroutines were
being started again every time a reconcile was run. This means
you could potentially have many goroutines monitoring the same
thing.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
During cluster deletion, we currently only retry for a couple minutes
to wait for the pvcs to be deleted. After the timeout, we proceed
with the cluster deletion. To properly protect the pvcs for proper
cleanup, the finalizer should not be removed until the pvcs
are all confirmed to be deleted. In order to not block other cluster
events, we re-queue the deletion event to run again every 10s
until the pvcs are deleted or the finalizer is manually removed.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
The controller setting for requeuing an event moved to the
opcontroller package and was no longer needed in the main
controller package.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
When Rook-Ceph is connected to an external cluster and the admin tries
to create resources such as CephFilesystem or CephObjectStore, we must
create the rook-ceph-config secret so that the daemons can get their mon
information.
Closes: https://github.com/rook/rook/issues/5439
Signed-off-by: Sébastien Han <seb@redhat.com>
An OSD on a PVC when portable=false is assigned to a node
with a node selector. The same node assignment is expected
for the lifetime of the cluster. On subsequent reconciles,
the operator was looking up the node assignment from the
nodeName on the pod spec, which is not set if the pod is down.
The operator needs to retrieve the assignment from the
deployment spec node selector.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
When an OSD pod is stuck in terminating state, there is no need
to check if the OSD is marked as out before force deleting.
As long as the osd is down we will force delete the pod.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
There are many duplicated code to judge whether an OSD is on PVC
or not. Instead, let's introduce a function for this purpose.
It can also be applied to the logic of judging whether an OSD
on PVC has metadata device or not.
Signed-off-by: Satoru Takeuchi <satoru.takeuchi@gmail.com>
If the OSD pod is portable and is on a failed node, the node
will be stuck indefinitely in the terminating state. The pod
can be force deleted in this scenario in order to allow the
underlying volume to be detached and then mounted on another
node.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
This is the final conversion to controller-runtime conversion. This time the
CephCluster CRD has been converted to use the controller-runtime
library.
The controller incorporates all the previous watchers too, so the Node
and hot-plug configmap are been watched too.
Only the operator setting configmap is not being watcher since it's not
related to the CephCluster CRD.
Not only the patch converts to controller-runtime but also tries to
re-organize the tree of the repo to actually make the code more
readable and have better functions/methods/tests separations.
Closes: https://github.com/rook/rook/issues/4939
Signed-off-by: Sébastien Han <seb@redhat.com>
There are some code where constants of environment variables
are defined but not used properly. Let's avoid to use string
literals as possible.
Signed-off-by: Satoru Takeuchi <satoru.takeuchi@gmail.com>
Travis Nielsen
binoue
Sébastien Han
UMEZAWA Takeshi
Ali Maredia
Alex Flom
Satoru Takeuchi
Madhu Rajanna