10 ha-manager - Proxmox VE HA Manager
15 include::ha-manager.1-synopsis.adoc[]
26 Our modern society depends heavily on information provided by
27 computers over the network. Mobile devices amplified that dependency,
28 because people can access the network any time from anywhere. If you
29 provide such services, it is very important that they are available
32 We can mathematically define the availability as the ratio of (A), the
33 total time a service is capable of being used during a given interval
34 to (B), the length of the interval. It is normally expressed as a
35 percentage of uptime in a given year.
37 .Availability - Downtime per Year
38 [width="60%",cols="<d,d",options="header"]
39 |===========================================================
40 |Availability % |Downtime per year
45 |99.9999 |31.5 seconds
46 |99.99999 |3.15 seconds
47 |===========================================================
49 There are several ways to increase availability. The most elegant
50 solution is to rewrite your software, so that you can run it on
51 several hosts at the same time. The software itself needs to have a way
52 to detect errors and do failover. If you only want to serve read-only
53 web pages, then this is relatively simple. However, this is generally complex
54 and sometimes impossible, because you cannot modify the software yourself. The
55 following solutions works without modifying the software:
57 * Use reliable ``server'' components
59 NOTE: Computer components with the same functionality can have varying
60 reliability numbers, depending on the component quality. Most vendors
61 sell components with higher reliability as ``server'' components -
62 usually at higher price.
64 * Eliminate single point of failure (redundant components)
65 ** use an uninterruptible power supply (UPS)
66 ** use redundant power supplies in your servers
68 ** use redundant network hardware
69 ** use RAID for local storage
70 ** use distributed, redundant storage for VM data
73 ** rapidly accessible administrators (24/7)
74 ** availability of spare parts (other nodes in a {pve} cluster)
75 ** automatic error detection (provided by `ha-manager`)
76 ** automatic failover (provided by `ha-manager`)
78 Virtualization environments like {pve} make it much easier to reach
79 high availability because they remove the ``hardware'' dependency. They
80 also support the setup and use of redundant storage and network
81 devices, so if one host fails, you can simply start those services on
82 another host within your cluster.
84 Better still, {pve} provides a software stack called `ha-manager`,
85 which can do that automatically for you. It is able to automatically
86 detect errors and do automatic failover.
88 {pve} `ha-manager` works like an ``automated'' administrator. First, you
89 configure what resources (VMs, containers, ...) it should
90 manage. Then, `ha-manager` observes the correct functionality, and handles
91 service failover to another node in case of errors. `ha-manager` can
92 also handle normal user requests which may start, stop, relocate and
95 But high availability comes at a price. High quality components are
96 more expensive, and making them redundant doubles the costs at
97 least. Additional spare parts increase costs further. So you should
98 carefully calculate the benefits, and compare with those additional
101 TIP: Increasing availability from 99% to 99.9% is relatively
102 simple. But increasing availability from 99.9999% to 99.99999% is very
103 hard and costly. `ha-manager` has typical error detection and failover
104 times of about 2 minutes, so you can get no more than 99.999%
111 You must meet the following requirements before you start with HA:
113 * at least three cluster nodes (to get reliable quorum)
115 * shared storage for VMs and containers
117 * hardware redundancy (everywhere)
119 * use reliable “server” components
121 * hardware watchdog - if not available we fall back to the
122 linux kernel software watchdog (`softdog`)
124 * optional hardware fencing devices
127 [[ha_manager_resources]]
131 We call the primary management unit handled by `ha-manager` a
132 resource. A resource (also called ``service'') is uniquely
133 identified by a service ID (SID), which consists of the resource type
134 and a type specific ID, for example `vm:100`. That example would be a
135 resource of type `vm` (virtual machine) with the ID 100.
137 For now we have two important resources types - virtual machines and
138 containers. One basic idea here is that we can bundle related software
139 into such a VM or container, so there is no need to compose one big
140 service from other services, as was done with `rgmanager`. In
141 general, a HA managed resource should not depend on other resources.
147 This section provides a short overview of common management tasks. The
148 first step is to enable HA for a resource. This is done by adding the
149 resource to the HA resource configuration. You can do this using the
150 GUI, or simply use the command-line tool, for example:
153 # ha-manager add vm:100
156 The HA stack now tries to start the resources and keep them
157 running. Please note that you can configure the ``requested''
158 resources state. For example you may want the HA stack to stop the
162 # ha-manager set vm:100 --state stopped
165 and start it again later:
168 # ha-manager set vm:100 --state started
171 You can also use the normal VM and container management commands. They
172 automatically forward the commands to the HA stack, so
178 simply sets the requested state to `started`. The same applies to `qm
179 stop`, which sets the requested state to `stopped`.
181 NOTE: The HA stack works fully asynchronous and needs to communicate
182 with other cluster members. Therefore, it takes some seconds until you see
183 the result of such actions.
185 To view the current HA resource configuration use:
193 And you can view the actual HA manager and resource state with:
198 master node1 (active, Wed Nov 23 11:07:23 2016)
199 lrm elsa (active, Wed Nov 23 11:07:19 2016)
200 service vm:100 (node1, started)
203 You can also initiate resource migration to other nodes:
206 # ha-manager migrate vm:100 node2
209 This uses online migration and tries to keep the VM running. Online
210 migration needs to transfer all used memory over the network, so it is
211 sometimes faster to stop the VM, then restart it on the new node. This can be
212 done using the `relocate` command:
215 # ha-manager relocate vm:100 node2
218 Finally, you can remove the resource from the HA configuration using
219 the following command:
222 # ha-manager remove vm:100
225 NOTE: This does not start or stop the resource.
227 But all HA related tasks can be done in the GUI, so there is no need to
228 use the command line at all.
234 This section provides a detailed description of the {PVE} HA manager
235 internals. It describes all involved daemons and how they work
236 together. To provide HA, two daemons run on each node:
240 The local resource manager (LRM), which controls the services running on
241 the local node. It reads the requested states for its services from
242 the current manager status file and executes the respective commands.
246 The cluster resource manager (CRM), which makes the cluster-wide
247 decisions. It sends commands to the LRM, processes the results,
248 and moves resources to other nodes if something fails. The CRM also
249 handles node fencing.
252 .Locks in the LRM & CRM
254 Locks are provided by our distributed configuration file system (pmxcfs).
255 They are used to guarantee that each LRM is active once and working. As an
256 LRM only executes actions when it holds its lock, we can mark a failed node
257 as fenced if we can acquire its lock. This then lets us recover any failed
258 HA services securely without any interference from the now unknown failed node.
259 This all gets supervised by the CRM which currently holds the manager master
263 [[ha_manager_service_states]]
267 The CRM uses a service state enumeration to record the current service
268 state. This state is displayed on the GUI and can be queried using
269 the `ha-manager` command-line tool:
274 master elsa (active, Mon Nov 21 07:23:29 2016)
275 lrm elsa (active, Mon Nov 21 07:23:22 2016)
276 service ct:100 (elsa, stopped)
277 service ct:102 (elsa, started)
278 service vm:501 (elsa, started)
281 Here is the list of possible states:
285 Service is stopped (confirmed by LRM). If the LRM detects a stopped
286 service is still running, it will stop it again.
290 Service should be stopped. The CRM waits for confirmation from the
295 Pending stop request. But the CRM did not get the request so far.
299 Service is active an LRM should start it ASAP if not already running.
300 If the Service fails and is detected to be not running the LRM
302 (see xref:ha_manager_start_failure_policy[Start Failure Policy]).
306 Pending start request. But the CRM has not got any confirmation from the
307 LRM that the service is running.
311 Wait for node fencing as the service node is not inside the quorate cluster
312 partition (see xref:ha_manager_fencing[Fencing]).
313 As soon as node gets fenced successfully the service will be placed into the
318 Wait for recovery of the service. The HA manager tries to find a new node where
319 the service can run on. This search depends not only on the list of online and
320 quorate nodes, but also if the service is a group member and how such a group
322 As soon as a new available node is found, the service will be moved there and
323 initially placed into stopped state. If it's configured to run the new node
328 Do not touch the service state. We use this state while we reboot a
329 node, or when we restart the LRM daemon
330 (see xref:ha_manager_package_updates[Package Updates]).
334 Act as if the service were not managed by HA at all.
335 Useful, when full control over the service is desired temporarily, without
336 removing it from the HA configuration.
340 Migrate service (live) to other node.
344 Service is disabled because of LRM errors. Needs manual intervention
345 (see xref:ha_manager_error_recovery[Error Recovery]).
349 Service is newly added, and the CRM has not seen it so far.
353 Service is stopped and marked as `disabled`
357 Local Resource Manager
358 ~~~~~~~~~~~~~~~~~~~~~~
360 The local resource manager (`pve-ha-lrm`) is started as a daemon on
361 boot and waits until the HA cluster is quorate and thus cluster-wide
364 It can be in three states:
366 wait for agent lock::
368 The LRM waits for our exclusive lock. This is also used as idle state if no
369 service is configured.
373 The LRM holds its exclusive lock and has services configured.
377 The LRM lost its lock, this means a failure happened and quorum was lost.
379 After the LRM gets in the active state it reads the manager status
380 file in `/etc/pve/ha/manager_status` and determines the commands it
381 has to execute for the services it owns.
382 For each command a worker gets started, these workers are running in
383 parallel and are limited to at most 4 by default. This default setting
384 may be changed through the datacenter configuration key `max_worker`.
385 When finished the worker process gets collected and its result saved for
388 .Maximum Concurrent Worker Adjustment Tips
390 The default value of at most 4 concurrent workers may be unsuited for
391 a specific setup. For example, 4 live migrations may occur at the same
392 time, which can lead to network congestions with slower networks and/or
393 big (memory wise) services. Also, ensure that in the worst case, congestion is
394 at a minimum, even if this means lowering the `max_worker` value. On the
395 contrary, if you have a particularly powerful, high-end setup you may also want
398 Each command requested by the CRM is uniquely identifiable by a UID. When
399 the worker finishes, its result will be processed and written in the LRM
400 status file `/etc/pve/nodes/<nodename>/lrm_status`. There the CRM may collect
401 it and let its state machine - respective to the commands output - act on it.
403 The actions on each service between CRM and LRM are normally always synced.
404 This means that the CRM requests a state uniquely marked by a UID, the LRM
405 then executes this action *one time* and writes back the result, which is also
406 identifiable by the same UID. This is needed so that the LRM does not
407 execute an outdated command.
408 The only exceptions to this behaviour are the `stop` and `error` commands;
409 these two do not depend on the result produced and are executed
410 always in the case of the stopped state and once in the case of
415 The HA Stack logs every action it makes. This helps to understand what
416 and also why something happens in the cluster. Here its important to see
417 what both daemons, the LRM and the CRM, did. You may use
418 `journalctl -u pve-ha-lrm` on the node(s) where the service is and
419 the same command for the pve-ha-crm on the node which is the current master.
423 Cluster Resource Manager
424 ~~~~~~~~~~~~~~~~~~~~~~~~
426 The cluster resource manager (`pve-ha-crm`) starts on each node and
427 waits there for the manager lock, which can only be held by one node
428 at a time. The node which successfully acquires the manager lock gets
429 promoted to the CRM master.
431 It can be in three states:
433 wait for agent lock::
435 The CRM waits for our exclusive lock. This is also used as idle state if no
436 service is configured
440 The CRM holds its exclusive lock and has services configured
444 The CRM lost its lock, this means a failure happened and quorum was lost.
446 Its main task is to manage the services which are configured to be highly
447 available and try to always enforce the requested state. For example, a
448 service with the requested state 'started' will be started if its not
449 already running. If it crashes it will be automatically started again.
450 Thus the CRM dictates the actions the LRM needs to execute.
452 When a node leaves the cluster quorum, its state changes to unknown.
453 If the current CRM can then secure the failed node's lock, the services
454 will be 'stolen' and restarted on another node.
456 When a cluster member determines that it is no longer in the cluster
457 quorum, the LRM waits for a new quorum to form. As long as there is no
458 quorum the node cannot reset the watchdog. This will trigger a reboot
459 after the watchdog times out (this happens after 60 seconds).
465 [thumbnail="screenshot/gui-ha-manager-status.png"]
467 By using the HA simulator you can test and learn all functionalities of the
468 Proxmox VE HA solutions.
470 By default, the simulator allows you to watch and test the behaviour of a
471 real-world 3 node cluster with 6 VMs. You can also add or remove additional VMs
474 You do not have to setup or configure a real cluster, the HA simulator runs out
480 apt install pve-ha-simulator
483 You can even install the package on any Debian-based system without any
484 other Proxmox VE packages. For that you will need to download the package and
485 copy it to the system you want to run it on for installation. When you install
486 the package with apt from the local file system it will also resolve the
487 required dependencies for you.
490 To start the simulator on a remote machine you must have an X11 redirection to
493 If you are on a Linux machine you can use:
496 ssh root@<IPofPVE> -Y
499 On Windows it works with https://mobaxterm.mobatek.net/[mobaxterm].
501 After connecting to an existing {pve} with the simulator installed or
502 installing it on your local Debian-based system manually, you can try it out as
505 First you need to create a working directory where the simulator saves its
506 current state and writes its default config:
512 Then, simply pass the created directory as a parameter to 'pve-ha-simulator':
515 pve-ha-simulator working/
518 You can then start, stop, migrate the simulated HA services, or even check out
519 what happens on a node failure.
524 The HA stack is well integrated into the {pve} API. So, for example,
525 HA can be configured via the `ha-manager` command-line interface, or
526 the {pve} web interface - both interfaces provide an easy way to
527 manage HA. Automation tools can use the API directly.
529 All HA configuration files are within `/etc/pve/ha/`, so they get
530 automatically distributed to the cluster nodes, and all nodes share
531 the same HA configuration.
534 [[ha_manager_resource_config]]
538 [thumbnail="screenshot/gui-ha-manager-status.png"]
541 The resource configuration file `/etc/pve/ha/resources.cfg` stores
542 the list of resources managed by `ha-manager`. A resource configuration
543 inside that list looks like this:
551 It starts with a resource type followed by a resource specific name,
552 separated with colon. Together this forms the HA resource ID, which is
553 used by all `ha-manager` commands to uniquely identify a resource
554 (example: `vm:100` or `ct:101`). The next lines contain additional
557 include::ha-resources-opts.adoc[]
559 Here is a real world example with one VM and one container. As you see,
560 the syntax of those files is really simple, so it is even possible to
561 read or edit those files using your favorite editor:
563 .Configuration Example (`/etc/pve/ha/resources.cfg`)
570 # Note: use default settings for everything
573 [thumbnail="screenshot/gui-ha-manager-add-resource.png"]
575 The above config was generated using the `ha-manager` command-line tool:
578 # ha-manager add vm:501 --state started --max_relocate 2
579 # ha-manager add ct:102
583 [[ha_manager_groups]]
587 [thumbnail="screenshot/gui-ha-manager-groups-view.png"]
589 The HA group configuration file `/etc/pve/ha/groups.cfg` is used to
590 define groups of cluster nodes. A resource can be restricted to run
591 only on the members of such group. A group configuration look like
601 include::ha-groups-opts.adoc[]
603 [thumbnail="screenshot/gui-ha-manager-add-group.png"]
605 A common requirement is that a resource should run on a specific
606 node. Usually the resource is able to run on other nodes, so you can define
607 an unrestricted group with a single member:
610 # ha-manager groupadd prefer_node1 --nodes node1
613 For bigger clusters, it makes sense to define a more detailed failover
614 behavior. For example, you may want to run a set of services on
615 `node1` if possible. If `node1` is not available, you want to run them
616 equally split on `node2` and `node3`. If those nodes also fail, the
617 services should run on `node4`. To achieve this you could set the node
621 # ha-manager groupadd mygroup1 -nodes "node1:2,node2:1,node3:1,node4"
624 Another use case is if a resource uses other resources only available
625 on specific nodes, lets say `node1` and `node2`. We need to make sure
626 that HA manager does not use other nodes, so we need to create a
627 restricted group with said nodes:
630 # ha-manager groupadd mygroup2 -nodes "node1,node2" -restricted
633 The above commands created the following group configuration file:
635 .Configuration Example (`/etc/pve/ha/groups.cfg`)
641 nodes node2:1,node4,node1:2,node3:1
649 The `nofailback` options is mostly useful to avoid unwanted resource
650 movements during administration tasks. For example, if you need to
651 migrate a service to a node which doesn't have the highest priority in the
652 group, you need to tell the HA manager not to instantly move this service
653 back by setting the `nofailback` option.
655 Another scenario is when a service was fenced and it got recovered to
656 another node. The admin tries to repair the fenced node and brings it
657 up online again to investigate the cause of failure and check if it runs
658 stably again. Setting the `nofailback` flag prevents the recovered services from
659 moving straight back to the fenced node.
662 [[ha_manager_fencing]]
666 On node failures, fencing ensures that the erroneous node is
667 guaranteed to be offline. This is required to make sure that no
668 resource runs twice when it gets recovered on another node. This is a
669 really important task, because without this, it would not be possible to
670 recover a resource on another node.
672 If a node did not get fenced, it would be in an unknown state where
673 it may have still access to shared resources. This is really
674 dangerous! Imagine that every network but the storage one broke. Now,
675 while not reachable from the public network, the VM still runs and
676 writes to the shared storage.
678 If we then simply start up this VM on another node, we would get a
679 dangerous race condition, because we write from both nodes. Such
680 conditions can destroy all VM data and the whole VM could be rendered
681 unusable. The recovery could also fail if the storage protects against
688 There are different methods to fence a node, for example, fence
689 devices which cut off the power from the node or disable their
690 communication completely. Those are often quite expensive and bring
691 additional critical components into a system, because if they fail you
692 cannot recover any service.
694 We thus wanted to integrate a simpler fencing method, which does not
695 require additional external hardware. This can be done using
698 .Possible Fencing Methods
699 - external power switches
700 - isolate nodes by disabling complete network traffic on the switch
701 - self fencing using watchdog timers
703 Watchdog timers have been widely used in critical and dependable systems
704 since the beginning of microcontrollers. They are often simple, independent
705 integrated circuits which are used to detect and recover from computer malfunctions.
707 During normal operation, `ha-manager` regularly resets the watchdog
708 timer to prevent it from elapsing. If, due to a hardware fault or
709 program error, the computer fails to reset the watchdog, the timer
710 will elapse and trigger a reset of the whole server (reboot).
712 Recent server motherboards often include such hardware watchdogs, but
713 these need to be configured. If no watchdog is available or
714 configured, we fall back to the Linux Kernel 'softdog'. While still
715 reliable, it is not independent of the servers hardware, and thus has
716 a lower reliability than a hardware watchdog.
719 Configure Hardware Watchdog
720 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
722 By default, all hardware watchdog modules are blocked for security
723 reasons. They are like a loaded gun if not correctly initialized. To
724 enable a hardware watchdog, you need to specify the module to load in
725 '/etc/default/pve-ha-manager', for example:
728 # select watchdog module (default is softdog)
729 WATCHDOG_MODULE=iTCO_wdt
732 This configuration is read by the 'watchdog-mux' service, which loads
733 the specified module at startup.
736 Recover Fenced Services
737 ~~~~~~~~~~~~~~~~~~~~~~~
739 After a node failed and its fencing was successful, the CRM tries to
740 move services from the failed node to nodes which are still online.
742 The selection of nodes, on which those services gets recovered, is
743 influenced by the resource `group` settings, the list of currently active
744 nodes, and their respective active service count.
746 The CRM first builds a set out of the intersection between user selected
747 nodes (from `group` setting) and available nodes. It then choose the
748 subset of nodes with the highest priority, and finally select the node
749 with the lowest active service count. This minimizes the possibility
750 of an overloaded node.
752 CAUTION: On node failure, the CRM distributes services to the
753 remaining nodes. This increases the service count on those nodes, and
754 can lead to high load, especially on small clusters. Please design
755 your cluster so that it can handle such worst case scenarios.
758 [[ha_manager_start_failure_policy]]
760 ---------------------
762 The start failure policy comes into effect if a service failed to start on a
763 node one or more times. It can be used to configure how often a restart
764 should be triggered on the same node and how often a service should be
765 relocated, so that it has an attempt to be started on another node.
766 The aim of this policy is to circumvent temporary unavailability of shared
767 resources on a specific node. For example, if a shared storage isn't available
768 on a quorate node anymore, for instance due to network problems, but is still
769 available on other nodes, the relocate policy allows the service to start
772 There are two service start recover policy settings which can be configured
773 specific for each resource.
777 Maximum number of attempts to restart a failed service on the actual
778 node. The default is set to one.
782 Maximum number of attempts to relocate the service to a different node.
783 A relocate only happens after the max_restart value is exceeded on the
784 actual node. The default is set to one.
786 NOTE: The relocate count state will only reset to zero when the
787 service had at least one successful start. That means if a service is
788 re-started without fixing the error only the restart policy gets
792 [[ha_manager_error_recovery]]
796 If, after all attempts, the service state could not be recovered, it gets
797 placed in an error state. In this state, the service won't get touched
798 by the HA stack anymore. The only way out is disabling a service:
801 # ha-manager set vm:100 --state disabled
804 This can also be done in the web interface.
806 To recover from the error state you should do the following:
808 * bring the resource back into a safe and consistent state (e.g.:
809 kill its process if the service could not be stopped)
811 * disable the resource to remove the error flag
813 * fix the error which led to this failures
815 * *after* you fixed all errors you may request that the service starts again
818 [[ha_manager_package_updates]]
822 When updating the ha-manager, you should do one node after the other, never
823 all at once for various reasons. First, while we test our software
824 thoroughly, a bug affecting your specific setup cannot totally be ruled out.
825 Updating one node after the other and checking the functionality of each node
826 after finishing the update helps to recover from eventual problems, while
827 updating all at once could result in a broken cluster and is generally not
830 Also, the {pve} HA stack uses a request acknowledge protocol to perform
831 actions between the cluster and the local resource manager. For restarting,
832 the LRM makes a request to the CRM to freeze all its services. This prevents
833 them from getting touched by the Cluster during the short time the LRM is restarting.
834 After that, the LRM may safely close the watchdog during a restart.
835 Such a restart happens normally during a package update and, as already stated,
836 an active master CRM is needed to acknowledge the requests from the LRM. If
837 this is not the case the update process can take too long which, in the worst
838 case, may result in a reset triggered by the watchdog.
841 [[ha_manager_node_maintenance]]
845 Sometimes it is necessary to perform maintenance on a node, such as replacing
846 hardware or simply installing a new kernel image. This also applies while the
849 The HA stack can support you mainly in two types of maintenance:
851 * for general shutdowns or reboots, the behavior can be configured, see
852 xref:ha_manager_shutdown_policy[Shutdown Policy].
853 * for maintenance that does not require a shutdown or reboot, or that should
854 not be switched off automatically after only one reboot, you can enable the
855 manual maintenance mode.
861 You can use the manual maintenance mode to mark the node as unavailable for HA
862 operation, prompting all services managed by HA to migrate to other nodes.
864 The target nodes for these migrations are selected from the other currently
865 available nodes, and determined by the HA group configuration and the configured
866 cluster resource scheduler (CRS) mode.
867 During each migration, the original node will be recorded in the HA managers'
868 state, so that the service can be moved back again automatically once the
869 maintenance mode is disabled and the node is back online.
871 Currently you can enabled or disable the maintenance mode using the ha-manager
874 .Enabling maintenance mode for a node
876 # ha-manager crm-command node-maintenance enable NODENAME
879 This will queue a CRM command, when the manager processes this command it will
880 record the request for maintenance-mode in the manager status. This allows you
881 to submit the command on any node, not just on the one you want to place in, or
882 out of the maintenance mode.
884 Once the LRM on the respective node picks the command up it will mark itself as
885 unavailable, but still process all migration commands. This means that the LRM
886 self-fencing watchdog will stay active until all active services got moved, and
887 all running workers finished.
889 Note that the LRM status will read `maintenance` mode as soon as the LRM
890 picked the requested state up, not only when all services got moved away, this
891 user experience is planned to be improved in the future.
892 For now, you can check for any active HA service left on the node, or watching
893 out for a log line like: `pve-ha-lrm[PID]: watchdog closed (disabled)` to know
894 when the node finished its transition into the maintenance mode.
896 NOTE: The manual maintenance mode is not automatically deleted on node reboot,
897 but only if it is either manually deactivated using the `ha-manager` CLI or if
898 the manager-status is manually cleared.
900 .Disabling maintenance mode for a node
902 # ha-manager crm-command node-maintenance disable NODENAME
905 The process of disabling the manual maintenance mode is similar to enabling it.
906 Using the `ha-manager` CLI command shown above will queue a CRM command that,
907 once processed, marks the respective LRM node as available again.
909 If you deactivate the maintenance mode, all services that were on the node when
910 the maintenance mode was activated will be moved back.
912 [[ha_manager_shutdown_policy]]
916 Below you will find a description of the different HA policies for a node
917 shutdown. Currently 'Conditional' is the default due to backward compatibility.
918 Some users may find that 'Migrate' behaves more as expected.
920 The shutdown policy can be configured in the Web UI (`Datacenter` -> `Options`
921 -> `HA Settings`), or directly in `datacenter.cfg`:
924 ha: shutdown_policy=<value>
930 Once the Local Resource manager (LRM) gets a shutdown request and this policy
931 is enabled, it will mark itself as unavailable for the current HA manager.
932 This triggers a migration of all HA Services currently located on this node.
933 The LRM will try to delay the shutdown process, until all running services get
934 moved away. But, this expects that the running services *can* be migrated to
935 another node. In other words, the service must not be locally bound, for example
936 by using hardware passthrough. As non-group member nodes are considered as
937 runnable target if no group member is available, this policy can still be used
938 when making use of HA groups with only some nodes selected. But, marking a group
939 as 'restricted' tells the HA manager that the service cannot run outside of the
940 chosen set of nodes. If all of those nodes are unavailable, the shutdown will
941 hang until you manually intervene. Once the shut down node comes back online
942 again, the previously displaced services will be moved back, if they were not
943 already manually migrated in-between.
945 NOTE: The watchdog is still active during the migration process on shutdown.
946 If the node loses quorum it will be fenced and the services will be recovered.
948 If you start a (previously stopped) service on a node which is currently being
949 maintained, the node needs to be fenced to ensure that the service can be moved
950 and started on another available node.
955 This mode ensures that all services get stopped, but that they will also be
956 recovered, if the current node is not online soon. It can be useful when doing
957 maintenance on a cluster scale, where live-migrating VMs may not be possible if
958 too many nodes are powered off at a time, but you still want to ensure HA
959 services get recovered and started again as soon as possible.
964 This mode ensures that all services get stopped and frozen, so that they won't
965 get recovered until the current node is online again.
970 The 'Conditional' shutdown policy automatically detects if a shutdown or a
971 reboot is requested, and changes behaviour accordingly.
975 A shutdown ('poweroff') is usually done if it is planned for the node to stay
976 down for some time. The LRM stops all managed services in this case. This means
977 that other nodes will take over those services afterwards.
979 NOTE: Recent hardware has large amounts of memory (RAM). So we stop all
980 resources, then restart them to avoid online migration of all that RAM. If you
981 want to use online migration, you need to invoke that manually before you
987 Node reboots are initiated with the 'reboot' command. This is usually done
988 after installing a new kernel. Please note that this is different from
989 ``shutdown'', because the node immediately starts again.
991 The LRM tells the CRM that it wants to restart, and waits until the CRM puts
992 all resources into the `freeze` state (same mechanism is used for
993 xref:ha_manager_package_updates[Package Updates]). This prevents those resources
994 from being moved to other nodes. Instead, the CRM starts the resources after the
995 reboot on the same node.
998 Manual Resource Movement
999 ^^^^^^^^^^^^^^^^^^^^^^^^
1001 Last but not least, you can also manually move resources to other nodes, before
1002 you shutdown or restart a node. The advantage is that you have full control,
1003 and you can decide if you want to use online migration or not.
1005 NOTE: Please do not 'kill' services like `pve-ha-crm`, `pve-ha-lrm` or
1006 `watchdog-mux`. They manage and use the watchdog, so this can result in an
1007 immediate node reboot or even reset.
1011 Cluster Resource Scheduling
1012 ---------------------------
1014 The cluster resource scheduler (CRS) mode controls how HA selects nodes for the
1015 recovery of a service as well as for migrations that are triggered by a
1016 shutdown policy. The default mode is `basic`, you can change it in the Web UI
1017 (`Datacenter` -> `Options`), or directly in `datacenter.cfg`:
1023 [thumbnail="screenshot/gui-datacenter-options-crs.png"]
1025 The change will be in effect starting with the next manager round (after a few
1028 For each service that needs to be recovered or migrated, the scheduler
1029 iteratively chooses the best node among the nodes with the highest priority in
1030 the service's group.
1032 NOTE: There are plans to add modes for (static and dynamic) load-balancing in
1038 The number of active HA services on each node is used to choose a recovery node.
1039 Non-HA-managed services are currently not counted.
1041 Static-Load Scheduler
1042 ~~~~~~~~~~~~~~~~~~~~~
1044 IMPORTANT: The static mode is still a technology preview.
1046 Static usage information from HA services on each node is used to choose a
1047 recovery node. Usage of non-HA-managed services is currently not considered.
1049 For this selection, each node in turn is considered as if the service was
1050 already running on it, using CPU and memory usage from the associated guest
1051 configuration. Then for each such alternative, CPU and memory usage of all nodes
1052 are considered, with memory being weighted much more, because it's a truly
1053 limited resource. For both, CPU and memory, highest usage among nodes (weighted
1054 more, as ideally no node should be overcommitted) and average usage of all nodes
1055 (to still be able to distinguish in case there already is a more highly
1056 committed node) are considered.
1058 IMPORTANT: The more services the more possible combinations there are, so it's
1059 currently not recommended to use it if you have thousands of HA managed
1063 CRS Scheduling Points
1064 ~~~~~~~~~~~~~~~~~~~~~
1066 The CRS algorithm is not applied for every service in every round, since this
1067 would mean a large number of constant migrations. Depending on the workload,
1068 this could put more strain on the cluster than could be avoided by constant
1070 That's why the {pve} HA manager favors keeping services on their current node.
1072 The CRS is currently used at the following scheduling points:
1074 - Service recovery (always active). When a node with active HA services fails,
1075 all its services need to be recovered to other nodes. The CRS algorithm will
1076 be used here to balance that recovery over the remaining nodes.
1078 - HA group config changes (always active). If a node is removed from a group,
1079 or its priority is reduced, the HA stack will use the CRS algorithm to find a
1080 new target node for the HA services in that group, matching the adapted
1081 priority constraints.
1083 - HA service stopped -> start transtion (opt-in). Requesting that a stopped
1084 service should be started is an good opportunity to check for the best suited
1085 node as per the CRS algorithm, as moving stopped services is cheaper to do
1086 than moving them started, especially if their disk volumes reside on shared
1087 storage. You can enable this by setting the **`ha-rebalance-on-start`**
1088 CRS option in the datacenter config. You can change that option also in the
1089 Web UI, under `Datacenter` -> `Options` -> `Cluster Resource Scheduling`.
1092 include::pve-copyright.adoc[]