Revisions of kubernetes-salt
Containers Team (containersteam)
committed
(revision 257)
new commit from concourse: Commit 22a3b23 by Florian Bergmann fbergmann@suse.de Install system wide certificates from pillars. `cert`-state will install the certificates as trust anchors.
Containers Team (containersteam)
committed
(revision 256)
new commit from concourse: Commit 6c4ec0c by Maximilian Meister mmeister@suse.de skip removed etcd servers (bsc#1093305) Signed-off-by: Maximilian Meister <mmeister@suse.de>
Containers Team (containersteam)
committed
(revision 255)
new commit from concourse: Commit 03d371f by Rafael Fernández López ereslibre@ereslibre.es Remove default grace period and timeout when draining a node. By default, the grace period is -1, or whatever the pod specifies on its `terminationGracePeriodSeconds` spec. The pod can know better than us what it needs to cleanly stop, and we don't need to apply arbitrary timeouts. If this is not specified, the default `terminationGracePeriodSeconds` value is 30 seconds. After this grace termination period, a SIGKILL will be sent to the process when evicting pods. Aside from this, we should have an "inifinite" timeout. Given that this process doesn't stall, it's safer to perform this operation until it succeeds. If we have proof that this is causing problems we should add a timeout, but in general the draining process should not hang. The alternative is in reality the real problem: if we timeout the draining process, it can happen that certain pods with remote volumes (nfs, rbd...) are never evicted, and when we go to restart the machine it hangs, because systemd fails to kill the processes when there are active mounts. Since there are no sensible defaults for the grace period and for the global timeout is better to let the first one to the pod definition, and the second one to just "infinite" until we really hit an issue because of this. Fixes: bsc#1085980
Jordi Massaguer (jordimassaguerpla)
committed
(revision 254)
fix changelog
Containers Team (containersteam)
committed
(revision 253)
new commit from concourse: Commit 876f7c7 by Rafael Fernández López ereslibre@ereslibre.es Lower the per-request timeout when we are checking for successful query When we are waiting for some service to be up, if the request hangs for some reason, we want to retry at least several times. Without setting this value explicitly, it takes the default (`http_request_timeout` as 3600), what is way over our `wait_for` argument set at 300 seconds. By setting the default `http_request_timeout` to a more reasonable default when doing this kind of checks we can ensure that the request itself will timeout several times before we call it done. Fixes: bsc#1093540 Fixes: bsc#1093685
Containers Team (containersteam)
committed
(revision 252)
new commit from concourse: Commit b13d89a by Rafael Fernández López ereslibre@ereslibre.es Only remove the master grains if there are any masters to be updated. The `salt.function` call will be marked as failed if there were no minions to target. Make sure that we only run this step if we know that we'll have some targets available. Fixes: bsc#1093491
Containers Team (containersteam)
committed
(revision 251)
new commit from concourse: Commit c93d25d by Alvaro Saurin alvaro.saurin@gmail.com Queue the /etc/hosts update when triggered from a reactor. Fixes part of bsc#1093123
Containers Team (containersteam)
committed
(revision 250)
new commit from concourse: Commit bc4b7ae by Alvaro Saurin alvaro.saurin@gmail.com Updated diagrams feature#docs
Containers Team (containersteam)
committed
(revision 249)
new commit from concourse: Commit 442a76c by Rafael Fernández López ereslibre@ereslibre.es Make HAProxy work as an http proxy instead of a tcp proxy. This allows us to add fine-grained timeouts depending on the endpoint being accessed or with what parameters (e.g. /log?follow=true should have no timeout as happens on the apiserver). /exec is another example, but in this case the protocol is upgraded to spdy. Fixes: bsc#1071994
Containers Team (containersteam)
committed
(revision 248)
new commit from concourse: Commit 4b37cb9 by Maximilian Meister mmeister@suse.de fix eviction-hard path feature#compute-resources Signed-off-by: Maximilian Meister <mmeister@suse.de>
Containers Team (containersteam)
committed
(revision 247)
new commit from concourse: Commit 177f774 by Kiall Mac Innes kiall@macinnes.ie Add JUnit output Commit 28e522e by Kiall Mac Innes kiall@macinnes.ie Update README with style check steps Commit 248c228 by Kiall Mac Innes kiall@macinnes.ie Fixup python code style issues Commit 4712a69 by Kiall Mac Innes kiall@macinnes.ie Add flake8 job
Containers Team (containersteam)
committed
(revision 246)
new commit from concourse: Commit 6de5432 by Kiall Mac Innes kiall@macinnes.ie Add Housekeeping Job
buildservice-autocommit
accepted
request 606451
from
Containers Team (containersteam)
(revision 245)
baserev update by copy to link target
Containers Team (containersteam)
committed
(revision 244)
new commit from concourse: Commit 1657de5 by Flavio Castelli fcastelli@suse.com Add missing cri-o removal states This is required to fix node removal on clusters using CRI-O as CRI. Fixes bsc#1092614 Signed-off-by: Flavio Castelli <fcastelli@suse.com>
buildservice-autocommit
accepted
request 606230
from
Containers Team (containersteam)
(revision 243)
baserev update by copy to link target
buildservice-autocommit
accepted
request 605706
from
Containers Team (containersteam)
(revision 242)
baserev update by copy to link target
Containers Team (containersteam)
committed
(revision 241)
new commit from concourse: Commit e286f9b by Flavio Castelli fcastelli@suse.com Make crictl handling more robust Some of our states are now depending on `crictl` tool. All these states have to depend on the `kubelet service.running` one, otherwise the `crictl` socket won't be available and the state will fail. Also, with these changes, the "blame" of a failure should point directly to the guilty (`kubelet` service not running for whatever reason) instead of falling on the `haproxy` one. Finally, the check looking for `crictl` socket has been changed to ensure the socket file exists and the service is actually listening. This will help with bugs like bsc#1091419 Signed-off-by: Flavio Castelli <fcastelli@suse.com>
Containers Team (containersteam)
committed
(revision 240)
new commit from concourse: Commit bcf5415 by Flavio Castelli fcastelli@suse.com kubelet: allow resource reservation Allow kubelet to take into account resource reservation and eviction threshold. == Resource reservation == It's possible to reserve resources for the `kube` and the `system` components. The `kube` component is the one including the kubernetes components: api server, controller manager, scheduler, proxy, kubelet and the container engine components (docker, containerd, cri-o, runc). The `system` component is the `system.slice`, basically all the system services: sshd, cron, logrotate,... By default don't specify any kind of resource reservation. Note well: when the resource reservations are in place kubelet will reduce the amount or resources allocatable by the node. However **no** enforcement will be done neither on the `kube.slice` nor on the `system.slice`. This is not happening because: _service kubernetes-salt.changes kubernetes-salt.spec master.tar.gz Resource enforcement is done using cgroups. _service kubernetes-salt.changes kubernetes-salt.spec master.tar.gz The slices are created by systemd. _service kubernetes-salt.changes kubernetes-salt.spec master.tar.gz systemd doesn't manage all the available cgroups yet. _service kubernetes-salt.changes kubernetes-salt.spec master.tar.gz kubelet tries to manage cgroups that are not handled by systemd, resulting in the kubelet failing at startup. _service kubernetes-salt.changes kubernetes-salt.spec master.tar.gz Changing the cgroup driver to `systemd` doesn't fix the issue. Moreover enforcing limits on the `system` and the `kube` slices can lead to resource starvation of core components of the system. As advised even by the official kubernetes docs, this is something that only expert users should do only after extensive profiling of their nodes. Finally, even if we wanted to enforce the limits, the right place would be systemd (by tuning the slice settings). For more information see the official documentation: https://kubernetes.io/docs/tasks/administer-cluster/reserve-compute-resources/ == Eviction threshold == By default no eviction threshold is set. bsc#1086185 Signed-off-by: Flavio Castelli <fcastelli@suse.com>
Containers Team (containersteam)
committed
(revision 239)
new commit from concourse: Commit 964deee by Maximilian Meister mmeister@suse.de add condition to KUBE_ADMISSION_CONTROL bsc#1092140 Signed-off-by: Maximilian Meister <mmeister@suse.de> Commit eaab500 by Maximilian Meister mmeister@suse.de fix conflicting sls id's they need to be globally unique orch error happened when setting psp to false in params.sls partially fixes https://bugzilla.suse.com/show_bug.cgi?id=1092140 bsc#1092140 Signed-off-by: Maximilian Meister <mmeister@suse.de>
buildservice-autocommit
accepted
request 605055
from
Containers Team (containersteam)
(revision 238)
baserev update by copy to link target
Displaying revisions 121 - 140 of 377