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 host at the same time. The software itself need to have a way
52 to detect errors and do failover. This is relatively easy if you just
53 want to serve read-only web pages. But in general this is complex, and
54 sometimes impossible because you cannot modify the software
55 yourself. The following solutions works without modifying the
58 * Use reliable ``server'' components
60 NOTE: Computer components with same functionality can have varying
61 reliability numbers, depending on the component quality. Most vendors
62 sell components with higher reliability as ``server'' components -
63 usually at higher price.
65 * Eliminate single point of failure (redundant components)
66 ** use an uninterruptible power supply (UPS)
67 ** use redundant power supplies on the main boards
69 ** use redundant network hardware
70 ** use RAID for local storage
71 ** use distributed, redundant storage for VM data
74 ** rapidly accessible administrators (24/7)
75 ** availability of spare parts (other nodes in a {pve} cluster)
76 ** automatic error detection (provided by `ha-manager`)
77 ** automatic failover (provided by `ha-manager`)
79 Virtualization environments like {pve} make it much easier to reach
80 high availability because they remove the ``hardware'' dependency. They
81 also support to setup and use redundant storage and network
82 devices. So if one host fail, you can simply start those services on
83 another host within your cluster.
85 Even better, {pve} provides a software stack called `ha-manager`,
86 which can do that automatically for you. It is able to automatically
87 detect errors and do automatic failover.
89 {pve} `ha-manager` works like an ``automated'' administrator. First, you
90 configure what resources (VMs, containers, ...) it should
91 manage. `ha-manager` then observes correct functionality, and handles
92 service failover to another node in case of errors. `ha-manager` can
93 also handle normal user requests which may start, stop, relocate and
96 But high availability comes at a price. High quality components are
97 more expensive, and making them redundant duplicates the costs at
98 least. Additional spare parts increase costs further. So you should
99 carefully calculate the benefits, and compare with those additional
102 TIP: Increasing availability from 99% to 99.9% is relatively
103 simply. But increasing availability from 99.9999% to 99.99999% is very
104 hard and costly. `ha-manager` has typical error detection and failover
105 times of about 2 minutes, so you can get no more than 99.999%
112 You must meet the following requirements before you start with HA:
114 * at least three cluster nodes (to get reliable quorum)
116 * shared storage for VMs and containers
118 * hardware redundancy (everywhere)
120 * use reliable “server” components
122 * hardware watchdog - if not available we fall back to the
123 linux kernel software watchdog (`softdog`)
125 * optional hardware fencing devices
128 [[ha_manager_resources]]
132 We call the primary management unit handled by `ha-manager` a
133 resource. A resource (also called ``service'') is uniquely
134 identified by a service ID (SID), which consists of the resource type
135 and an type specific ID, e.g.: `vm:100`. That example would be a
136 resource of type `vm` (virtual machine) with the ID 100.
138 For now we have two important resources types - virtual machines and
139 containers. One basic idea here is that we can bundle related software
140 into such VM or container, so there is no need to compose one big
141 service from other services, like it was done with `rgmanager`. In
142 general, a HA enabled resource should not depend on other resources.
148 This section provides an in detail description of the {PVE} HA-manager
149 internals. It describes how the CRM and the LRM work together.
151 To provide High Availability two daemons run on each node:
155 The local resource manager (LRM), which controls the services running on
156 the local node. It reads the requested states for its services from
157 the current manager status file and executes the respective commands.
161 The cluster resource manager (CRM), which makes the cluster wide
162 decisions. It sends commands to the LRM, processes the results,
163 and moves resources to other nodes if something fails. The CRM also
164 handles node fencing.
167 .Locks in the LRM & CRM
169 Locks are provided by our distributed configuration file system (pmxcfs).
170 They are used to guarantee that each LRM is active once and working. As a
171 LRM only executes actions when it holds its lock we can mark a failed node
172 as fenced if we can acquire its lock. This lets us then recover any failed
173 HA services securely without any interference from the now unknown failed node.
174 This all gets supervised by the CRM which holds currently the manager master
177 Local Resource Manager
178 ~~~~~~~~~~~~~~~~~~~~~~
180 The local resource manager (`pve-ha-lrm`) is started as a daemon on
181 boot and waits until the HA cluster is quorate and thus cluster wide
184 It can be in three states:
186 wait for agent lock::
188 The LRM waits for our exclusive lock. This is also used as idle state if no
189 service is configured.
193 The LRM holds its exclusive lock and has services configured.
197 The LRM lost its lock, this means a failure happened and quorum was lost.
199 After the LRM gets in the active state it reads the manager status
200 file in `/etc/pve/ha/manager_status` and determines the commands it
201 has to execute for the services it owns.
202 For each command a worker gets started, this workers are running in
203 parallel and are limited to at most 4 by default. This default setting
204 may be changed through the datacenter configuration key `max_worker`.
205 When finished the worker process gets collected and its result saved for
208 .Maximum Concurrent Worker Adjustment Tips
210 The default value of at most 4 concurrent workers may be unsuited for
211 a specific setup. For example may 4 live migrations happen at the same
212 time, which can lead to network congestions with slower networks and/or
213 big (memory wise) services. Ensure that also in the worst case no congestion
214 happens and lower the `max_worker` value if needed. In the contrary, if you
215 have a particularly powerful high end setup you may also want to increase it.
217 Each command requested by the CRM is uniquely identifiable by an UID, when
218 the worker finished its result will be processed and written in the LRM
219 status file `/etc/pve/nodes/<nodename>/lrm_status`. There the CRM may collect
220 it and let its state machine - respective the commands output - act on it.
222 The actions on each service between CRM and LRM are normally always synced.
223 This means that the CRM requests a state uniquely marked by an UID, the LRM
224 then executes this action *one time* and writes back the result, also
225 identifiable by the same UID. This is needed so that the LRM does not
226 executes an outdated command.
227 With the exception of the `stop` and the `error` command,
228 those two do not depend on the result produced and are executed
229 always in the case of the stopped state and once in the case of
234 The HA Stack logs every action it makes. This helps to understand what
235 and also why something happens in the cluster. Here its important to see
236 what both daemons, the LRM and the CRM, did. You may use
237 `journalctl -u pve-ha-lrm` on the node(s) where the service is and
238 the same command for the pve-ha-crm on the node which is the current master.
240 Cluster Resource Manager
241 ~~~~~~~~~~~~~~~~~~~~~~~~
243 The cluster resource manager (`pve-ha-crm`) starts on each node and
244 waits there for the manager lock, which can only be held by one node
245 at a time. The node which successfully acquires the manager lock gets
246 promoted to the CRM master.
248 It can be in three states:
250 wait for agent lock::
252 The CRM waits for our exclusive lock. This is also used as idle state if no
253 service is configured
257 The CRM holds its exclusive lock and has services configured
261 The CRM lost its lock, this means a failure happened and quorum was lost.
263 It main task is to manage the services which are configured to be highly
264 available and try to always enforce them to the wanted state, e.g.: a
265 enabled service will be started if its not running, if it crashes it will
266 be started again. Thus it dictates the LRM the actions it needs to execute.
268 When an node leaves the cluster quorum, its state changes to unknown.
269 If the current CRM then can secure the failed nodes lock, the services
270 will be 'stolen' and restarted on another node.
272 When a cluster member determines that it is no longer in the cluster
273 quorum, the LRM waits for a new quorum to form. As long as there is no
274 quorum the node cannot reset the watchdog. This will trigger a reboot
275 after the watchdog then times out, this happens after 60 seconds.
281 The HA stack is well integrated into the {pve} API. So, for example,
282 HA can be configured via the `ha-manager` command line interface, or
283 the {pve} web interface - both interfaces provide an easy way to
284 manage HA. Automation tools can use the API directly.
286 All HA configuration files are within `/etc/pve/ha/`, so they get
287 automatically distributed to the cluster nodes, and all nodes share
288 the same HA configuration.
294 The resource configuration file `/etc/pve/ha/resources.cfg` stores
295 the list of resources managed by `ha-manager`. A resource configuration
296 inside that list look like this:
304 It starts with a resource type followed by a resource specific name,
305 separated with colon. Together this forms the HA resource ID, which is
306 used by all `ha-manager` commands to uniquely identify a resource
307 (example: `vm:100` or `ct:101`). The next lines contain additional
310 include::ha-resources-opts.adoc[]
312 Here is a real world example with one VM and one container. As you see,
313 the syntax of those files is really simple, so it is even posiible to
314 read or edit those files using your favorite editor:
316 .Configuration Example (`/etc/pve/ha/resources.cfg`)
323 # Note: use default settings for everything
326 Above config was generated using the `ha-manager` command line tool:
329 # ha-manager add vm:501 --state started --max_relocate 2
330 # ha-manager add ct:102
337 The HA group configuration file `/etc/pve/ha/groups.cfg` is used to
338 define groups of cluster nodes. A resource can be restricted to run
339 only on the members of such group. A group configuration look like
349 include::ha-groups-opts.adoc[]
355 If a node needs maintenance you should migrate and or relocate all
356 services which are required to run always on another node first.
357 After that you can stop the LRM and CRM services. But note that the
358 watchdog triggers if you stop it with active services.
363 When updating the ha-manager you should do one node after the other, never
364 all at once for various reasons. First, while we test our software
365 thoughtfully, a bug affecting your specific setup cannot totally be ruled out.
366 Upgrading one node after the other and checking the functionality of each node
367 after finishing the update helps to recover from an eventual problems, while
368 updating all could render you in a broken cluster state and is generally not
371 Also, the {pve} HA stack uses a request acknowledge protocol to perform
372 actions between the cluster and the local resource manager. For restarting,
373 the LRM makes a request to the CRM to freeze all its services. This prevents
374 that they get touched by the Cluster during the short time the LRM is restarting.
375 After that the LRM may safely close the watchdog during a restart.
376 Such a restart happens on a update and as already stated a active master
377 CRM is needed to acknowledge the requests from the LRM, if this is not the case
378 the update process can be too long which, in the worst case, may result in
382 [[ha_manager_fencing]]
389 Fencing secures that on a node failure the dangerous node gets will be rendered
390 unable to do any damage and that no resource runs twice when it gets recovered
391 from the failed node. This is a really important task and one of the base
392 principles to make a system Highly Available.
394 If a node would not get fenced it would be in an unknown state where it may
395 have still access to shared resources, this is really dangerous!
396 Imagine that every network but the storage one broke, now while not
397 reachable from the public network the VM still runs and writes on the shared
398 storage. If we would not fence the node and just start up this VM on another
399 Node we would get dangerous race conditions, atomicity violations the whole VM
400 could be rendered unusable. The recovery could also simply fail if the storage
401 protects from multiple mounts and thus defeat the purpose of HA.
406 There are different methods to fence a node, for example fence devices which
407 cut off the power from the node or disable their communication completely.
409 Those are often quite expensive and bring additional critical components in
410 a system, because if they fail you cannot recover any service.
412 We thus wanted to integrate a simpler method in the HA Manager first, namely
413 self fencing with watchdogs.
415 Watchdogs are widely used in critical and dependable systems since the
416 beginning of micro controllers, they are often independent and simple
417 integrated circuit which programs can use to watch them. After opening they need to
418 report periodically. If, for whatever reason, a program becomes unable to do
419 so the watchdogs triggers a reset of the whole server.
421 Server motherboards often already include such hardware watchdogs, these need
422 to be configured. If no watchdog is available or configured we fall back to the
423 Linux Kernel softdog while still reliable it is not independent of the servers
424 Hardware and thus has a lower reliability then a hardware watchdog.
426 Configure Hardware Watchdog
427 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
428 By default all watchdog modules are blocked for security reasons as they are
429 like a loaded gun if not correctly initialized.
430 If you have a hardware watchdog available remove its kernel module from the
431 blacklist, load it with insmod and restart the `watchdog-mux` service or reboot
434 Recover Fenced Services
435 ~~~~~~~~~~~~~~~~~~~~~~~
437 After a node failed and its fencing was successful we start to recover services
438 to other available nodes and restart them there so that they can provide service
441 The selection of the node on which the services gets recovered is influenced
442 by the users group settings, the currently active nodes and their respective
443 active service count.
444 First we build a set out of the intersection between user selected nodes and
445 available nodes. Then the subset with the highest priority of those nodes
446 gets chosen as possible nodes for recovery. We select the node with the
447 currently lowest active service count as a new node for the service.
448 That minimizes the possibility of an overload, which else could cause an
449 unresponsive node and as a result a chain reaction of node failures in the
452 [[ha_manager_groups]]
456 A group is a collection of cluster nodes which a service may be bound to.
463 List of group node members where a priority can be given to each node.
464 A service bound to this group will run on the nodes with the highest priority
465 available. If more nodes are in the highest priority class the services will
466 get distributed to those node if not already there. The priorities have a
467 relative meaning only.
469 You want to run all services from a group on `node1` if possible. If this node
470 is not available, you want them to run equally splitted on `node2` and `node3`, and
471 if those fail it should use `node4`.
472 To achieve this you could set the node list to:
474 ha-manager groupset mygroup -nodes "node1:2,node2:1,node3:1,node4"
478 Resources bound to this group may only run on nodes defined by the
479 group. If no group node member is available the resource will be
480 placed in the stopped state.
482 Lets say a service uses resources only available on `node1` and `node2`,
483 so we need to make sure that HA manager does not use other nodes.
484 We need to create a 'restricted' group with said nodes:
486 ha-manager groupset mygroup -nodes "node1,node2" -restricted
490 The resource won't automatically fail back when a more preferred node
491 (re)joins the cluster.
493 * You need to migrate a service to a node which hasn't the highest priority
494 in the group at the moment, to tell the HA manager to not move this service
495 instantly back set the 'nofailback' option and the service will stay on
498 * A service was fenced and it got recovered to another node. The admin
499 repaired the node and brought it up online again but does not want that the
500 recovered services move straight back to the repaired node as he wants to
501 first investigate the failure cause and check if it runs stable. He can use
502 the 'nofailback' option to achieve this.
506 ---------------------
508 The start failure policy comes in effect if a service failed to start on a
509 node once ore more times. It can be used to configure how often a restart
510 should be triggered on the same node and how often a service should be
511 relocated so that it gets a try to be started on another node.
512 The aim of this policy is to circumvent temporary unavailability of shared
513 resources on a specific node. For example, if a shared storage isn't available
514 on a quorate node anymore, e.g. network problems, but still on other nodes,
515 the relocate policy allows then that the service gets started nonetheless.
517 There are two service start recover policy settings which can be configured
518 specific for each resource.
522 Maximum number of tries to restart an failed service on the actual
523 node. The default is set to one.
527 Maximum number of tries to relocate the service to a different node.
528 A relocate only happens after the max_restart value is exceeded on the
529 actual node. The default is set to one.
531 NOTE: The relocate count state will only reset to zero when the
532 service had at least one successful start. That means if a service is
533 re-enabled without fixing the error only the restart policy gets
539 If after all tries the service state could not be recovered it gets
540 placed in an error state. In this state the service won't get touched
541 by the HA stack anymore. To recover from this state you should follow
544 * bring the resource back into a safe and consistent state (e.g.,
547 * disable the ha resource to place it in an stopped state
549 * fix the error which led to this failures
551 * *after* you fixed all errors you may enable the service again
554 [[ha_manager_service_operations]]
558 This are how the basic user-initiated service operations (via
563 The service will be started by the LRM if not already running.
567 The service will be stopped by the LRM if running.
571 The service will be relocated (live) to another node.
575 The service will be removed from the HA managed resource list. Its
576 current state will not be touched.
580 `start` and `stop` commands can be issued to the resource specific tools
581 (like `qm` or `pct`), they will forward the request to the
582 `ha-manager` which then will execute the action and set the resulting
583 service state (enabled, disabled).
591 Service is stopped (confirmed by LRM), if detected running it will get stopped
596 Service should be stopped. Waiting for confirmation from LRM.
600 Service is active an LRM should start it ASAP if not already running.
601 If the Service fails and is detected to be not running the LRM restarts it.
605 Wait for node fencing (service node is not inside quorate cluster
607 As soon as node gets fenced successfully the service will be recovered to
608 another node, if possible.
612 Do not touch the service state. We use this state while we reboot a
613 node, or when we restart the LRM daemon.
617 Migrate service (live) to other node.
621 Service disabled because of LRM errors. Needs manual intervention.
625 include::pve-copyright.adoc[]