diff --git a/doc/sphinx/shared/pacemaker-intro.rst b/doc/sphinx/shared/pacemaker-intro.rst
new file mode 100644
index 0000000000..37c39afd56
--- /dev/null
+++ b/doc/sphinx/shared/pacemaker-intro.rst
@@ -0,0 +1,201 @@
+What Is Pacemaker?
+####################
+
+Pacemaker is a high-availability *cluster resource manager* -- software that
+runs on a set of hosts (a *cluster* of *nodes*) in order to preserve integrity
+and minimize downtime of desired services (*resources*). [#]_ It is maintained
+by the `ClusterLabs `_ community.
+
+Pacemaker's key features include:
+
+* Detection of and recovery from node- and service-level failures
+* Ability to ensure data integrity by fencing faulty nodes
+* Support for one or more nodes per cluster
+* Support for multiple resource interface standards (anything that can be
+ scripted can be clustered)
+* Support (but no requirement) for shared storage
+* Support for practically any redundancy configuration (active/passive, N+1,
+ etc.)
+* Automatically replicated configuration that can be updated from any node
+* Ability to specify cluster-wide relationships between services,
+ such as ordering, colocation and anti-colocation
+* Support for advanced service types, such as *clones* (services that need to
+ be active on multiple nodes), *promotable clones* (clones that can run in
+ one of two roles), and containerized services
+* Unified, scriptable cluster management tools
+
+.. note:: Fencing
+
+ *Fencing*, also known as *STONITH* (an acronym for Shoot The Other Node In
+ The Head), is the ability to ensure that it is not possible for a node to be
+ running a service. This is accomplished via *fence devices* such as
+ intelligent power switches that cut power to the target, or intelligent
+ network switches that cut the target's access to the local network.
+
+ Pacemaker represents fence devices as a special class of resource.
+
+ A cluster cannot safely recover from certain failure conditions, such as an
+ unresponsive node, without fencing.
+
+Cluster Architecture
+____________________
+
+At a high level, a cluster can be viewed as having these parts (which together
+are often referred to as the *cluster stack*):
+
+ * **Resources:** These are the reason for the cluster's being -- the services
+ that need to be kept highly available.
+
+ * **Resource agents:** These are scripts or operating system components that
+ start, stop, and monitor resources, given a set of resource parameters.
+ These provide a uniform interface between Pacemaker and the managed
+ services.
+
+ * **Fence agents:** These are scripts that execute node fencing actions,
+ given a target and fence device parameters.
+
+ * **Cluster membership layer:** This component provides reliable messaging,
+ membership, and quorum information about the cluster. Currently, Pacemaker
+ supports `Corosync `_ as this layer.
+
+ * **Cluster resource manager:** Pacemaker provides the brain that processes
+ and reacts to events that occur in the cluster. These events may include
+ nodes joining or leaving the cluster; resource events caused by failures,
+ maintenance, or scheduled activities; and other administrative actions.
+ To achieve the desired availability, Pacemaker may start and stop resources
+ and fence nodes.
+
+ * **Cluster tools:** These provide an interface for users to interact with the
+ cluster. Various command-line and graphical (GUI) interfaces are available.
+
+Most managed services are not, themselves, cluster-aware. However, many popular
+open-source cluster filesystems make use of a common *Distributed Lock
+Manager* (DLM), which makes direct use of Corosync for its messaging and
+membership capabilities and Pacemaker for the ability to fence nodes.
+
+.. image:: ../../shared/en-US/images/pcmk-stack.png
+ :alt: Example cluster stack
+ :scale: 75 %
+ :align: center
+
+Pacemaker Architecture
+______________________
+
+Pacemaker itself is composed of multiple daemons that work together:
+
+* pacemakerd
+* pacemaker-attrd
+* pacemaker-based
+* pacemaker-controld
+* pacemaker-execd
+* pacemaker-fenced
+* pacemaker-schedulerd
+
+.. image:: ../../shared/en-US/images/pcmk-internals.png
+ :alt: Pacemaker software components
+ :scale: 65 %
+ :align: center
+
+The Pacemaker master process (pacemakerd) spawns all the other daemons, and
+respawns them if they unexpectedly exit.
+
+The *Cluster Information Base* (CIB) is an
+`XML `_ representation of the cluster's
+configuration and the state of all nodes and resources. The *CIB manager*
+(pacemaker-based) keeps the CIB synchronized across the cluster, and handles
+requests to modify it.
+
+The *attribute manager* (pacemaker-attrd) maintains a database of attributes
+for all nodes, keeps it synchronized across the cluster, and handles requests
+to modify them. These attributes are usually recorded in the CIB.
+
+Given a snapshot of the CIB as input, the *scheduler* (pacemaker-schedulerd)
+determines what actions are necessary to achieve the desired state of the
+cluster.
+
+The *local executor* (pacemaker-execd) handles requests to execute
+resource agents on the local cluster node, and returns the result.
+
+The *fencer* (pacemaker-fenced) handles requests to fence nodes. Given a target
+node, the fencer decides which cluster node(s) should execute which fencing
+device(s), and calls the necessary fencing agents (either directly, or via
+requests to the fencer peers on other nodes), and returns the result.
+
+The *controller* (pacemaker-controld) is Pacemaker's coordinator, maintaining a
+consistent view of the cluster membership and orchestrating all the other
+components.
+
+Pacemaker centralizes cluster decision-making by electing one of the controller
+instances as the 'Designated Controller' ('DC'). Should the elected DC process
+(or the node it is on) fail, a new one is quickly established. The DC responds
+to cluster events by taking a current snapshot of the CIB, feeding it to the
+scheduler, then asking the executors (either directly on the local node, or via
+requests to controller peers on other nodes) and the fencer to execute any
+necessary actions.
+
+.. note:: **Old daemon names**
+
+ The Pacemaker daemons were renamed in version 2.0. You may still find
+ references to the old names, especially in documentation targeted to
+ version 1.1.
+
+ .. table::
+
+ +-------------------+---------------------+
+ | Old name | New name |
+ +===================+=====================+
+ | attrd | pacemaker-attrd |
+ +-------------------+---------------------+
+ | cib | pacemaker-based |
+ +-------------------+---------------------+
+ | crmd | pacemaker-controld |
+ +-------------------+---------------------+
+ | lrmd | pacemaker-execd |
+ +-------------------+---------------------+
+ | stonithd | pacemaker-fenced |
+ +-------------------+---------------------+
+ | pacemaker_remoted | pacemaker-remoted |
+ +-------------------+---------------------+
+
+Node Redundancy Designs
+_______________________
+
+Pacemaker supports practically any `node redundancy configuration
+`_
+including *Active/Active*, *Active/Passive*, *N+1*, *N+M*, *N-to-1* and
+*N-to-N*.
+
+Active/passive clusters with two (or more) nodes using Pacemaker and
+`DRBD `_ are
+a cost-effective high-availability solution for many situations. One of the
+nodes provides the desired services, and if it fails, the other node takes
+over.
+
+.. image:: ../../shared/en-US/images/pcmk-active-passive.png
+ :alt: Active/Passive Redundancy
+ :align: center
+ :scale: 75 %
+
+Pacemaker also supports multiple nodes in a shared-failover design, reducing
+hardware costs by allowing several active/passive clusters to be combined and
+share a common backup node.
+
+.. image:: ../../shared/en-US/images/pcmk-shared-failover.png
+ :alt: Shared Failover
+ :align: center
+ :scale: 75 %
+
+When shared storage is available, every node can potentially be used for
+failover. Pacemaker can even run multiple copies of services to spread out the
+workload. This is sometimes called N to N Redundancy.
+
+.. image:: ../../shared/en-US/images/pcmk-active-active.png
+ :alt: N to N Redundancy
+ :align: center
+ :scale: 75 %
+
+.. rubric:: Footnotes
+
+.. [#] *Cluster* is sometimes used in other contexts to refer to hosts grouped
+ together for other purposes, such as high-performance computing (HPC),
+ but Pacemaker is not intended for those purposes.