5 include::attributes.txt[]
10 ha-manager - Proxmox VE HA Manager
15 include::ha-manager.1-synopsis.adoc[]
24 include::attributes.txt[]
28 Our modern society depends heavily on information provided by
29 computers over the network. Mobile devices amplified that dependency,
30 because people can access the network any time from anywhere. If you
31 provide such services, it is very important that they are available
34 We can mathematically define the availability as the ratio of (A) the
35 total time a service is capable of being used during a given interval
36 to (B) the length of the interval. It is normally expressed as a
37 percentage of uptime in a given year.
39 .Availability - Downtime per Year
40 [width="60%",cols="<d,d",options="header"]
41 |===========================================================
42 |Availability % |Downtime per year
47 |99.9999 |31.5 seconds
48 |99.99999 |3.15 seconds
49 |===========================================================
51 There are several ways to increase availability. The most elegant
52 solution is to rewrite your software, so that you can run it on
53 several host at the same time. The software itself need to have a way
54 to detect errors and do failover. This is relatively easy if you just
55 want to serve read-only web pages. But in general this is complex, and
56 sometimes impossible because you cannot modify the software
57 yourself. The following solutions works without modifying the
60 * Use reliable ``server'' components
62 NOTE: Computer components with same functionality can have varying
63 reliability numbers, depending on the component quality. Most vendors
64 sell components with higher reliability as ``server'' components -
65 usually at higher price.
67 * Eliminate single point of failure (redundant components)
68 ** use an uninterruptible power supply (UPS)
69 ** use redundant power supplies on the main boards
71 ** use redundant network hardware
72 ** use RAID for local storage
73 ** use distributed, redundant storage for VM data
76 ** rapidly accessible administrators (24/7)
77 ** availability of spare parts (other nodes in a {pve} cluster)
78 ** automatic error detection (provided by `ha-manager`)
79 ** automatic failover (provided by `ha-manager`)
81 Virtualization environments like {pve} make it much easier to reach
82 high availability because they remove the ``hardware'' dependency. They
83 also support to setup and use redundant storage and network
84 devices. So if one host fail, you can simply start those services on
85 another host within your cluster.
87 Even better, {pve} provides a software stack called `ha-manager`,
88 which can do that automatically for you. It is able to automatically
89 detect errors and do automatic failover.
91 {pve} `ha-manager` works like an ``automated'' administrator. First, you
92 configure what resources (VMs, containers, ...) it should
93 manage. `ha-manager` then observes correct functionality, and handles
94 service failover to another node in case of errors. `ha-manager` can
95 also handle normal user requests which may start, stop, relocate and
98 But high availability comes at a price. High quality components are
99 more expensive, and making them redundant duplicates the costs at
100 least. Additional spare parts increase costs further. So you should
101 carefully calculate the benefits, and compare with those additional
104 TIP: Increasing availability from 99% to 99.9% is relatively
105 simply. But increasing availability from 99.9999% to 99.99999% is very
106 hard and costly. `ha-manager` has typical error detection and failover
107 times of about 2 minutes, so you can get no more than 99.999%
113 * at least three cluster nodes (to get reliable quorum)
115 * shared storage for VMs and containers
117 * hardware redundancy (everywhere)
119 * hardware watchdog - if not available we fall back to the
120 linux kernel software watchdog (`softdog`)
122 * optional hardware fencing devices
128 We call the primary management unit handled by `ha-manager` a
129 resource. A resource (also called ``service'') is uniquely
130 identified by a service ID (SID), which consists of the resource type
131 and an type specific ID, e.g.: `vm:100`. That example would be a
132 resource of type `vm` (virtual machine) with the ID 100.
134 For now we have two important resources types - virtual machines and
135 containers. One basic idea here is that we can bundle related software
136 into such VM or container, so there is no need to compose one big
137 service from other services, like it was done with `rgmanager`. In
138 general, a HA enabled resource should not depend on other resources.
144 This section provides an in detail description of the {PVE} HA-manager
145 internals. It describes how the CRM and the LRM work together.
147 To provide High Availability two daemons run on each node:
151 The local resource manager (LRM), it controls the services running on
153 It reads the requested states for its services from the current manager
154 status file and executes the respective commands.
158 The cluster resource manager (CRM), it controls the cluster wide
159 actions of the services, processes the LRM results and includes the state
160 machine which controls the state of each service.
162 .Locks in the LRM & CRM
164 Locks are provided by our distributed configuration file system (pmxcfs).
165 They are used to guarantee that each LRM is active once and working. As a
166 LRM only executes actions when it holds its lock we can mark a failed node
167 as fenced if we can acquire its lock. This lets us then recover any failed
168 HA services securely without any interference from the now unknown failed node.
169 This all gets supervised by the CRM which holds currently the manager master
172 Local Resource Manager
173 ~~~~~~~~~~~~~~~~~~~~~~
175 The local resource manager (`pve-ha-lrm`) is started as a daemon on
176 boot and waits until the HA cluster is quorate and thus cluster wide
179 It can be in three states:
181 * *wait for agent lock*: the LRM waits for our exclusive lock. This is
182 also used as idle sate if no service is configured
183 * *active*: the LRM holds its exclusive lock and has services configured
184 * *lost agent lock*: the LRM lost its lock, this means a failure happened
187 After the LRM gets in the active state it reads the manager status
188 file in `/etc/pve/ha/manager_status` and determines the commands it
189 has to execute for the services it owns.
190 For each command a worker gets started, this workers are running in
191 parallel and are limited to at most 4 by default. This default setting
192 may be changed through the datacenter configuration key `max_worker`.
193 When finished the worker process gets collected and its result saved for
196 .Maximum Concurrent Worker Adjustment Tips
198 The default value of at most 4 concurrent workers may be unsuited for
199 a specific setup. For example may 4 live migrations happen at the same
200 time, which can lead to network congestions with slower networks and/or
201 big (memory wise) services. Ensure that also in the worst case no congestion
202 happens and lower the `max_worker` value if needed. In the contrary, if you
203 have a particularly powerful high end setup you may also want to increase it.
205 Each command requested by the CRM is uniquely identifiable by an UID, when
206 the worker finished its result will be processed and written in the LRM
207 status file `/etc/pve/nodes/<nodename>/lrm_status`. There the CRM may collect
208 it and let its state machine - respective the commands output - act on it.
210 The actions on each service between CRM and LRM are normally always synced.
211 This means that the CRM requests a state uniquely marked by an UID, the LRM
212 then executes this action *one time* and writes back the result, also
213 identifiable by the same UID. This is needed so that the LRM does not
214 executes an outdated command.
215 With the exception of the `stop` and the `error` command,
216 those two do not depend on the result produced and are executed
217 always in the case of the stopped state and once in the case of
222 The HA Stack logs every action it makes. This helps to understand what
223 and also why something happens in the cluster. Here its important to see
224 what both daemons, the LRM and the CRM, did. You may use
225 `journalctl -u pve-ha-lrm` on the node(s) where the service is and
226 the same command for the pve-ha-crm on the node which is the current master.
228 Cluster Resource Manager
229 ~~~~~~~~~~~~~~~~~~~~~~~~
231 The cluster resource manager (`pve-ha-crm`) starts on each node and
232 waits there for the manager lock, which can only be held by one node
233 at a time. The node which successfully acquires the manager lock gets
234 promoted to the CRM master.
236 It can be in three states:
238 * *wait for agent lock*: the LRM waits for our exclusive lock. This is
239 also used as idle sate if no service is configured
240 * *active*: the LRM holds its exclusive lock and has services configured
241 * *lost agent lock*: the LRM lost its lock, this means a failure happened
244 It main task is to manage the services which are configured to be highly
245 available and try to always enforce them to the wanted state, e.g.: a
246 enabled service will be started if its not running, if it crashes it will
247 be started again. Thus it dictates the LRM the actions it needs to execute.
249 When an node leaves the cluster quorum, its state changes to unknown.
250 If the current CRM then can secure the failed nodes lock, the services
251 will be 'stolen' and restarted on another node.
253 When a cluster member determines that it is no longer in the cluster
254 quorum, the LRM waits for a new quorum to form. As long as there is no
255 quorum the node cannot reset the watchdog. This will trigger a reboot
256 after the watchdog then times out, this happens after 60 seconds.
261 The HA stack is well integrated in the Proxmox VE API2. So, for
262 example, HA can be configured via `ha-manager` or the PVE web
263 interface, which both provide an easy to use tool.
265 The resource configuration file can be located at
266 `/etc/pve/ha/resources.cfg` and the group configuration file at
267 `/etc/pve/ha/groups.cfg`. Use the provided tools to make changes,
268 there shouldn't be any need to edit them manually.
273 If a node needs maintenance you should migrate and or relocate all
274 services which are required to run always on another node first.
275 After that you can stop the LRM and CRM services. But note that the
276 watchdog triggers if you stop it with active services.
281 When updating the ha-manager you should do one node after the other, never
282 all at once for various reasons. First, while we test our software
283 thoughtfully, a bug affecting your specific setup cannot totally be ruled out.
284 Upgrading one node after the other and checking the functionality of each node
285 after finishing the update helps to recover from an eventual problems, while
286 updating all could render you in a broken cluster state and is generally not
289 Also, the {pve} HA stack uses a request acknowledge protocol to perform
290 actions between the cluster and the local resource manager. For restarting,
291 the LRM makes a request to the CRM to freeze all its services. This prevents
292 that they get touched by the Cluster during the short time the LRM is restarting.
293 After that the LRM may safely close the watchdog during a restart.
294 Such a restart happens on a update and as already stated a active master
295 CRM is needed to acknowledge the requests from the LRM, if this is not the case
296 the update process can be too long which, in the worst case, may result in
306 Fencing secures that on a node failure the dangerous node gets will be rendered
307 unable to do any damage and that no resource runs twice when it gets recovered
308 from the failed node. This is a really important task and one of the base
309 principles to make a system Highly Available.
311 If a node would not get fenced it would be in an unknown state where it may
312 have still access to shared resources, this is really dangerous!
313 Imagine that every network but the storage one broke, now while not
314 reachable from the public network the VM still runs and writes on the shared
315 storage. If we would not fence the node and just start up this VM on another
316 Node we would get dangerous race conditions, atomicity violations the whole VM
317 could be rendered unusable. The recovery could also simply fail if the storage
318 protects from multiple mounts and thus defeat the purpose of HA.
323 There are different methods to fence a node, for example fence devices which
324 cut off the power from the node or disable their communication completely.
326 Those are often quite expensive and bring additional critical components in
327 a system, because if they fail you cannot recover any service.
329 We thus wanted to integrate a simpler method in the HA Manager first, namely
330 self fencing with watchdogs.
332 Watchdogs are widely used in critical and dependable systems since the
333 beginning of micro controllers, they are often independent and simple
334 integrated circuit which programs can use to watch them. After opening they need to
335 report periodically. If, for whatever reason, a program becomes unable to do
336 so the watchdogs triggers a reset of the whole server.
338 Server motherboards often already include such hardware watchdogs, these need
339 to be configured. If no watchdog is available or configured we fall back to the
340 Linux Kernel softdog while still reliable it is not independent of the servers
341 Hardware and thus has a lower reliability then a hardware watchdog.
343 Configure Hardware Watchdog
344 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
345 By default all watchdog modules are blocked for security reasons as they are
346 like a loaded gun if not correctly initialized.
347 If you have a hardware watchdog available remove its kernel module from the
348 blacklist, load it with insmod and restart the `watchdog-mux` service or reboot
351 Recover Fenced Services
352 ~~~~~~~~~~~~~~~~~~~~~~~
354 After a node failed and its fencing was successful we start to recover services
355 to other available nodes and restart them there so that they can provide service
358 The selection of the node on which the services gets recovered is influenced
359 by the users group settings, the currently active nodes and their respective
360 active service count.
361 First we build a set out of the intersection between user selected nodes and
362 available nodes. Then the subset with the highest priority of those nodes
363 gets chosen as possible nodes for recovery. We select the node with the
364 currently lowest active service count as a new node for the service.
365 That minimizes the possibility of an overload, which else could cause an
366 unresponsive node and as a result a chain reaction of node failures in the
372 A group is a collection of cluster nodes which a service may be bound to.
379 List of group node members where a priority can be given to each node.
380 A service bound to this group will run on the nodes with the highest priority
381 available. If more nodes are in the highest priority class the services will
382 get distributed to those node if not already there. The priorities have a
383 relative meaning only.
387 Resources bound to this group may only run on nodes defined by the
388 group. If no group node member is available the resource will be
389 placed in the stopped state.
393 The resource won't automatically fail back when a more preferred node
394 (re)joins the cluster.
398 ---------------------
400 The start failure policy comes in effect if a service failed to start on a
401 node once ore more times. It can be used to configure how often a restart
402 should be triggered on the same node and how often a service should be
403 relocated so that it gets a try to be started on another node.
404 The aim of this policy is to circumvent temporary unavailability of shared
405 resources on a specific node. For example, if a shared storage isn't available
406 on a quorate node anymore, e.g. network problems, but still on other nodes,
407 the relocate policy allows then that the service gets started nonetheless.
409 There are two service start recover policy settings which can be configured
410 specific for each resource.
414 Maximum number of tries to restart an failed service on the actual
415 node. The default is set to one.
419 Maximum number of tries to relocate the service to a different node.
420 A relocate only happens after the max_restart value is exceeded on the
421 actual node. The default is set to one.
423 NOTE: The relocate count state will only reset to zero when the
424 service had at least one successful start. That means if a service is
425 re-enabled without fixing the error only the restart policy gets
431 If after all tries the service state could not be recovered it gets
432 placed in an error state. In this state the service won't get touched
433 by the HA stack anymore. To recover from this state you should follow
436 * bring the resource back into a safe and consistent state (e.g.,
439 * disable the ha resource to place it in an stopped state
441 * fix the error which led to this failures
443 * *after* you fixed all errors you may enable the service again
449 This are how the basic user-initiated service operations (via
454 The service will be started by the LRM if not already running.
458 The service will be stopped by the LRM if running.
462 The service will be relocated (live) to another node.
466 The service will be removed from the HA managed resource list. Its
467 current state will not be touched.
471 `start` and `stop` commands can be issued to the resource specific tools
472 (like `qm` or `pct`), they will forward the request to the
473 `ha-manager` which then will execute the action and set the resulting
474 service state (enabled, disabled).
482 Service is stopped (confirmed by LRM), if detected running it will get stopped
487 Service should be stopped. Waiting for confirmation from LRM.
491 Service is active an LRM should start it ASAP if not already running.
492 If the Service fails and is detected to be not running the LRM restarts it.
496 Wait for node fencing (service node is not inside quorate cluster
498 As soon as node gets fenced successfully the service will be recovered to
499 another node, if possible.
503 Do not touch the service state. We use this state while we reboot a
504 node, or when we restart the LRM daemon.
508 Migrate service (live) to other node.
512 Service disabled because of LRM errors. Needs manual intervention.
516 include::pve-copyright.adoc[]