[pve-docs.git] / ha-manager.adoc

[[chapter-ha-manager]]
ifdef::manvolnum[]
PVE(1)
======
include::attributes.txt[]

:pve-toplevel:

NAME
----

ha-manager - Proxmox VE HA Manager

SYNOPSIS
--------

include::ha-manager.1-synopsis.adoc[]

DESCRIPTION
-----------
endif::manvolnum[]

ifndef::manvolnum[]
High Availability
=================
include::attributes.txt[]
endif::manvolnum[]

ifdef::wiki[]
:pve-toplevel:
endif::wiki[]

Our modern society depends heavily on information provided by
computers over the network. Mobile devices amplified that dependency,
because people can access the network any time from anywhere. If you
provide such services, it is very important that they are available
most of the time.

We can mathematically define the availability as the ratio of (A) the
total time a service is capable of being used during a given interval
to (B) the length of the interval. It is normally expressed as a
percentage of uptime in a given year.

.Availability - Downtime per Year
[width="60%",cols="<d,d",options="header"]
|===========================================================
|Availability % |Downtime per year
|99             |3.65 days
|99.9           |8.76 hours
|99.99          |52.56 minutes
|99.999         |5.26 minutes
|99.9999        |31.5 seconds
|99.99999       |3.15 seconds
|===========================================================

There are several ways to increase availability. The most elegant
solution is to rewrite your software, so that you can run it on
several host at the same time. The software itself need to have a way
to detect errors and do failover. This is relatively easy if you just
want to serve read-only web pages. But in general this is complex, and
sometimes impossible because you cannot modify the software
yourself. The following solutions works without modifying the
software:

* Use reliable ``server'' components

NOTE: Computer components with same functionality can have varying
reliability numbers, depending on the component quality. Most vendors
sell components with higher reliability as ``server'' components -
usually at higher price.

* Eliminate single point of failure (redundant components)
** use an uninterruptible power supply (UPS)
** use redundant power supplies on the main boards
** use ECC-RAM
** use redundant network hardware
** use RAID for local storage
** use distributed, redundant storage for VM data

* Reduce downtime
** rapidly accessible administrators (24/7)
** availability of spare parts (other nodes in a {pve} cluster)
** automatic error detection (provided by `ha-manager`)
** automatic failover (provided by `ha-manager`)

Virtualization environments like {pve} make it much easier to reach
high availability because they remove the ``hardware'' dependency. They
also support to setup and use redundant storage and network
devices. So if one host fail, you can simply start those services on
another host within your cluster.

Even better, {pve} provides a software stack called `ha-manager`,
which can do that automatically for you. It is able to automatically
detect errors and do automatic failover.

{pve} `ha-manager` works like an ``automated'' administrator. First, you
configure what resources (VMs, containers, ...) it should
manage. `ha-manager` then observes correct functionality, and handles
service failover to another node in case of errors. `ha-manager` can
also handle normal user requests which may start, stop, relocate and
migrate a service.

But high availability comes at a price. High quality components are
more expensive, and making them redundant duplicates the costs at
least. Additional spare parts increase costs further. So you should
carefully calculate the benefits, and compare with those additional
costs.

TIP: Increasing availability from 99% to 99.9% is relatively
simply. But increasing availability from 99.9999% to 99.99999% is very
hard and costly. `ha-manager` has typical error detection and failover
times of about 2 minutes, so you can get no more than 99.999%
availability.

Requirements
------------

* at least three cluster nodes (to get reliable quorum)

* shared storage for VMs and containers

* hardware redundancy (everywhere)

* hardware watchdog - if not available we fall back to the
  linux kernel software watchdog (`softdog`)

* optional hardware fencing devices


Resources
---------

We call the primary management unit handled by `ha-manager` a
resource. A resource (also called ``service'') is uniquely
identified by a service ID (SID), which consists of the resource type
and an type specific ID, e.g.: `vm:100`. That example would be a
resource of type `vm` (virtual machine) with the ID 100.

For now we have two important resources types - virtual machines and
containers. One basic idea here is that we can bundle related software
into such VM or container, so there is no need to compose one big
service from other services, like it was done with `rgmanager`. In
general, a HA enabled resource should not depend on other resources.


How It Works
------------

This section provides an in detail description of the {PVE} HA-manager
internals. It describes how the CRM and the LRM work together.

To provide High Availability two daemons run on each node:

`pve-ha-lrm`::

The local resource manager (LRM), it controls the services running on
the local node.
It reads the requested states for its services from the current manager
status file and executes the respective commands.

`pve-ha-crm`::

The cluster resource manager (CRM), it controls the cluster wide
actions of the services, processes the LRM results and includes the state
machine which controls the state of each service.

.Locks in the LRM & CRM
[NOTE]
Locks are provided by our distributed configuration file system (pmxcfs).
They are used to guarantee that each LRM is active once and working. As a
LRM only executes actions when it holds its lock we can mark a failed node
as fenced if we can acquire its lock. This lets us then recover any failed
HA services securely without any interference from the now unknown failed node.
This all gets supervised by the CRM which holds currently the manager master
lock.

Local Resource Manager
~~~~~~~~~~~~~~~~~~~~~~

The local resource manager (`pve-ha-lrm`) is started as a daemon on
boot and waits until the HA cluster is quorate and thus cluster wide
locks are working.

It can be in three states:

wait for agent lock::

The LRM waits for our exclusive lock. This is also used as idle state if no
service is configured.

active::

The LRM holds its exclusive lock and has services configured.

lost agent lock::

The LRM lost its lock, this means a failure happened and quorum was lost.

After the LRM gets in the active state it reads the manager status
file in `/etc/pve/ha/manager_status` and determines the commands it
has to execute for the services it owns.
For each command a worker gets started, this workers are running in
parallel and are limited to at most 4 by default. This default setting
may be changed through the datacenter configuration key `max_worker`.
When finished the worker process gets collected and its result saved for
the CRM.

.Maximum Concurrent Worker Adjustment Tips
[NOTE]
The default value of at most 4 concurrent workers may be unsuited for
a specific setup. For example may 4 live migrations happen at the same
time, which can lead to network congestions with slower networks and/or
big (memory wise) services. Ensure that also in the worst case no congestion
happens and lower the `max_worker` value if needed. In the contrary, if you
have a particularly powerful high end setup you may also want to increase it.

Each command requested by the CRM is uniquely identifiable by an UID, when
the worker finished its result will be processed and written in the LRM
status file `/etc/pve/nodes/<nodename>/lrm_status`. There the CRM may collect
it and let its state machine - respective the commands output - act on it.

The actions on each service between CRM and LRM are normally always synced.
This means that the CRM requests a state uniquely marked by an UID, the LRM
then executes this action *one time* and writes back the result, also
identifiable by the same UID. This is needed so that the LRM does not
executes an outdated command.
With the exception of the `stop` and the `error` command,
those two do not depend on the result produced and are executed
always in the case of the stopped state and once in the case of
the error state.

.Read the Logs
[NOTE]
The HA Stack logs every action it makes. This helps to understand what
and also why something happens in the cluster. Here its important to see
what both daemons, the LRM and the CRM, did. You may use
`journalctl -u pve-ha-lrm` on the node(s) where the service is and
the same command for the pve-ha-crm on the node which is the current master.

Cluster Resource Manager
~~~~~~~~~~~~~~~~~~~~~~~~

The cluster resource manager (`pve-ha-crm`) starts on each node and
waits there for the manager lock, which can only be held by one node
at a time.  The node which successfully acquires the manager lock gets
promoted to the CRM master.

It can be in three states:

wait for agent lock::

The CRM waits for our exclusive lock. This is also used as idle state if no
service is configured

active::

The CRM holds its exclusive lock and has services configured

lost agent lock::

The CRM lost its lock, this means a failure happened and quorum was lost.

It main task is to manage the services which are configured to be highly
available and try to always enforce them to the wanted state, e.g.: a
enabled service will be started if its not running, if it crashes it will
be started again. Thus it dictates the LRM the actions it needs to execute.

When an node leaves the cluster quorum, its state changes to unknown.
If the current CRM then can secure the failed nodes lock, the services
will be 'stolen' and restarted on another node.

When a cluster member determines that it is no longer in the cluster
quorum, the LRM waits for a new quorum to form. As long as there is no
quorum the node cannot reset the watchdog. This will trigger a reboot
after the watchdog then times out, this happens after 60 seconds.

Configuration
-------------

The HA stack is well integrated in the Proxmox VE API2. So, for
example, HA can be configured via `ha-manager` or the PVE web
interface, which both provide an easy to use tool.

The resource configuration file can be located at
`/etc/pve/ha/resources.cfg` and the group configuration file at
`/etc/pve/ha/groups.cfg`. Use the provided tools to make changes,
there shouldn't be any need to edit them manually.

Node Power Status
-----------------

If a node needs maintenance you should migrate and or relocate all
services which are required to run always on another node first.
After that you can stop the LRM and CRM services. But note that the
watchdog triggers if you stop it with active services.

Package Updates
---------------

When updating the ha-manager you should do one node after the other, never
all at once for various reasons. First, while we test our software
thoughtfully, a bug affecting your specific setup cannot totally be ruled out.
Upgrading one node after the other and checking the functionality of each node
after finishing the update helps to recover from an eventual problems, while
updating all could render you in a broken cluster state and is generally not
good practice.

Also, the {pve} HA stack uses a request acknowledge protocol to perform
actions between the cluster and the local resource manager. For restarting,
the LRM makes a request to the CRM to freeze all its services. This prevents
that they get touched by the Cluster during the short time the LRM is restarting.
After that the LRM may safely close the watchdog during a restart.
Such a restart happens on a update and as already stated a active master
CRM is needed to acknowledge the requests from the LRM, if this is not the case
the update process can be too long which, in the worst case, may result in
a watchdog reset.


Fencing
-------

What is Fencing
~~~~~~~~~~~~~~~

Fencing secures that on a node failure the dangerous node gets will be rendered
unable to do any damage and  that no resource runs twice when it gets recovered
from the failed node. This is a really important task and one of the base
principles to make a system Highly Available.

If a node would not get fenced it would be in an unknown state where it may
have still access to shared resources, this is really dangerous!
Imagine that every network but the storage one broke, now while not
reachable from the public network the VM still runs and writes on the shared
storage. If we would not fence the node and just start up this VM on another
Node we would get dangerous race conditions, atomicity violations the whole VM
could be rendered unusable. The recovery could also simply fail if the storage
protects from multiple mounts and thus defeat the purpose of HA.

How {pve} Fences
~~~~~~~~~~~~~~~~~

There are different methods to fence a node, for example fence devices which
cut off the power from the node or disable their communication completely.

Those are often quite expensive and bring additional critical components in
a system, because if they fail you cannot recover any service.

We thus wanted to integrate a simpler method in the HA Manager first, namely
self fencing with watchdogs.

Watchdogs are widely used in critical and dependable systems since the
beginning of micro controllers, they are often independent and simple
integrated circuit which  programs can use to watch them. After opening they need to
report periodically. If, for whatever reason, a program becomes unable to do
so the watchdogs triggers a reset of the whole server.

Server motherboards often already include such hardware watchdogs, these need
to be configured. If no watchdog is available or configured we fall back to the
Linux Kernel softdog while still reliable it is not independent of the servers
Hardware and thus has a lower reliability then a hardware watchdog.

Configure Hardware Watchdog
~~~~~~~~~~~~~~~~~~~~~~~~~~~
By default all watchdog modules are blocked for security reasons as they are
like a loaded gun if not correctly initialized.
If you have a hardware watchdog available remove its kernel module from the
blacklist, load it with insmod and restart the `watchdog-mux` service or reboot
the node.

Recover Fenced Services
~~~~~~~~~~~~~~~~~~~~~~~

After a node failed and its fencing was successful we start to recover services
to other available nodes and restart them there so that they can provide service
again.

The selection of the node on which the services gets recovered is influenced
by the users group settings, the currently active nodes and their respective
active service count.
First we build a set out of the intersection between user selected nodes and
available nodes. Then the subset with the highest priority of those nodes
gets chosen as possible nodes for recovery. We select the node with the
currently lowest active service count as a new node for the service.
That minimizes the possibility of an overload, which else could cause an
unresponsive node and as a result a chain reaction of node failures in the
cluster.

Groups
------

A group is a collection of cluster nodes which a service may be bound to.

Group Settings
~~~~~~~~~~~~~~

nodes::

List of group node members where a priority can be given to each node.
A service bound to this group will run on the nodes with the highest priority
available. If more nodes are in the highest priority class the services will
get distributed to those node if not already there. The priorities have a
relative meaning only.
  Example;;
  You want to run all services from a group on `node1` if possible. If this node
  is not available, you want them to run equally splitted on `node2` and `node3`, and
  if those fail it should use `node4`.
  To achieve this you could set the node list to:
[source,bash]
  ha-manager groupset mygroup -nodes "node1:2,node2:1,node3:1,node4"

restricted::

Resources bound to this group may only run on nodes defined by the
group. If no group node member is available the resource will be
placed in the stopped state.
  Example;;
  Lets say a service uses resources only available on `node1` and `node2`,
  so we need to make sure that HA manager does not use other nodes.
  We need to create a 'restricted' group with said nodes:
[source,bash]
  ha-manager groupset mygroup -nodes "node1,node2" -restricted 

nofailback::

The resource won't automatically fail back when a more preferred node
(re)joins the cluster.
  Examples;;
  * You need to migrate a service to a node which hasn't the highest priority
  in the group at the moment, to tell the HA manager to not move this service
  instantly back set the 'nofailback' option and the service will stay on
  the current node.

  * A service was fenced and it got recovered to another node. The admin
  repaired the node and brought it up online again but does not want that the
  recovered services move straight back to the repaired node as he wants to
  first investigate the failure cause and check if it runs stable. He can use
  the 'nofailback' option to achieve this.


Start Failure Policy
---------------------

The start failure policy comes in effect if a service failed to start on a
node once ore more times. It can be used to configure how often a restart
should be triggered on the same node and how often a service should be
relocated so that it gets a try to be started on another node.
The aim of this policy is to circumvent temporary unavailability of shared
resources on a specific node. For example, if a shared storage isn't available
on a quorate node anymore, e.g. network problems, but still on other nodes,
the relocate policy allows then that the service gets started nonetheless.

There are two service start recover policy settings which can be configured
specific for each resource.

max_restart::

Maximum number of tries to restart an failed service on the actual
node.  The default is set to one.

max_relocate::

Maximum number of tries to relocate the service to a different node.
A relocate only happens after the max_restart value is exceeded on the
actual node. The default is set to one.

NOTE: The relocate count state will only reset to zero when the
service had at least one successful start. That means if a service is
re-enabled without fixing the error only the restart policy gets
repeated.

Error Recovery
--------------

If after all tries the service state could not be recovered it gets
placed in an error state. In this state the service won't get touched
by the HA stack anymore.  To recover from this state you should follow
these steps:

* bring the resource back into a safe and consistent state (e.g.,
killing its process)

* disable the ha resource to place it in an stopped state

* fix the error which led to this failures

* *after* you fixed all errors you may enable the service again


Service Operations
------------------

This are how the basic user-initiated service operations (via
`ha-manager`) work.

enable::

The service will be started by the LRM if not already running.

disable::

The service will be stopped by the LRM if running.

migrate/relocate::

The service will be relocated (live) to another node.

remove::

The service will be removed from the HA managed resource list. Its
current state will not be touched.

start/stop::

`start` and `stop` commands can be issued to the resource specific tools
(like `qm` or `pct`), they will forward the request to the
`ha-manager` which then will execute the action and set the resulting
service state (enabled, disabled).


Service States
--------------

stopped::

Service is stopped (confirmed by LRM), if detected running it will get stopped
again.

request_stop::

Service should be stopped. Waiting for confirmation from LRM.

started::

Service is active an LRM should start it ASAP if not already running.
If the Service fails and is detected to be not running the LRM restarts it.

fence::

Wait for node fencing (service node is not inside quorate cluster
partition).
As soon as node gets fenced successfully the service will be recovered to
another node, if possible.

freeze::

Do not touch the service state. We use this state while we reboot a
node, or when we restart the LRM daemon.

migrate::

Migrate service (live) to other node.

error::

Service disabled because of LRM errors. Needs manual intervention.


ifdef::manvolnum[]
include::pve-copyright.adoc[]
endif::manvolnum[]