icinga2/doc/4-monitoring-remote-systems.md

33 KiB

Monitoring Remote Systems

Agent-less Checks

If the remote service is available using a network protocol and port, and a check plugin is available, you don't necessarily need a local client installed. Rather choose a plugin and configure all parameters and thresholds. The Icinga 2 Template Library already ships various examples.

Agent-based Checks

If the remote services are not directly accessible through the network, a local agent installation exposing the results to check queries can become handy.

SNMP

The SNMP daemon runs on the remote system and answers SNMP queries by plugin binaries. The Monitoring Plugins package ships the check_snmp plugin binary, but there are plenty of existing plugins for specific use cases already around, for example monitoring Cisco routers.

The following example uses the SNMP ITL CheckCommand and just overrides the snmp_oid custom attribute. A service is created for all hosts which have the snmp-community custom attribute.

apply Service "uptime" {
  import "generic-service"

  check_command = "snmp"
  vars.snmp_oid = "1.3.6.1.2.1.1.3.0"

  assign where host.vars.snmp_community != ""
}

SSH

Calling a plugin using the SSH protocol to execute a plugin on the remote server fetching its return code and output. The by_ssh command object is part of the built-in templates and requires the check_by_ssh check plugin which is available in the Monitoring Plugins package.

object CheckCommand "by_ssh_swap" {
  import "by_ssh"

  vars.by_ssh_command = "/usr/lib/nagios/plugins/check_swap -w $by_ssh_swap_warn$ -c $by_ssh_swap_crit$"
  vars.by_ssh_swap_warn = "75%"
  vars.by_ssh_swap_crit = "50%"
}

object Service "swap" {
  import "generic-service"

  host_name = "remote-ssh-host"

  check_command = "by_ssh_swap"

  vars.by_ssh_logname = "icinga"
}

NRPE

NRPE runs as daemon on the remote client including the required plugins and command definitions. Icinga 2 calls the check_nrpe plugin binary in order to query the configured command on the remote client.

The NRPE daemon uses its own configuration format in nrpe.cfg while check_nrpe can be embedded into the Icinga 2 CheckCommand configuration syntax.

You can use the check_nrpe plugin from the NRPE project to query the NRPE daemon. Icinga 2 provides the nrpe check command for this:

Example:

object Service "users" {
  import "generic-service"

  host_name = "remote-nrpe-host"

  check_command = "nrpe"
  vars.nrpe_command = "check_users"
}

nrpe.cfg:

command[check_users]=/usr/local/icinga/libexec/check_users -w 5 -c 10

NSClient++

NSClient++ works on both Windows and Linux platforms and is well known for its magnificent Windows support. There are alternatives like the WMI interface, but using NSClient++ will allow you to run local scripts similar to check plugins fetching the required output and performance counters.

You can use the check_nt plugin from the Monitoring Plugins project to query NSClient++. Icinga 2 provides the nscp check command for this:

Example:

object Service "disk" {
  import "generic-service"

  host_name = "remote-windows-host"

  check_command = "nscp"

  vars.nscp_variable = "USEDDISKSPACE"
  vars.nscp_params = "c"
  vars.nscp_warn = 70
  vars.nscp_crit = 80
}

For details on the NSClient++ configuration please refer to the official documentation.

Icinga 2 Agent

A dedicated Icinga 2 agent supporting all platforms and using the native Icinga 2 communication protocol supported with SSL certificates, IPv4/IPv6 support, etc. is on the development roadmap. Meanwhile remote checkers in a cluster setup could act as immediate replacement, but without any local configuration - or pushing their standalone configuration back to the master node including their check result messages.

Note

Remote checker instances are independent Icinga 2 instances which schedule their checks and just synchronize them back to the defined master zone.

Passive Check Results and SNMP Traps

SNMP Traps can be received and filtered by using SNMPTT and specific trap handlers passing the check results to Icinga 2.

Note

The host and service object configuration must be available on the Icinga 2 server in order to process passive check results.

NSCA-NG

NSCA-ng provides a client-server pair that allows the remote sender to push check results into the Icinga 2 ExternalCommandListener feature.

Note

This addon works in a similar fashion like the Icinga 1.x distributed model. If you are looking for a real distributed architecture with Icinga 2, scroll down.

Distributed Monitoring and High Availability

Building distributed environments with high availability included is fairly easy with Icinga 2. The cluster feature is built-in and allows you to build many scenarios based on your requirements:

  • High Availability. All instances in the Zone elect one active master and run as Active/Active cluster.
  • Distributed Zones. A master zone and one or more satellites in their zones.
  • Load Distribution. A configuration master and multiple checker satellites.

You can combine these scenarios into a global setup fitting your requirements.

Each instance got their own event scheduler, and does not depend on a centralized master coordinating and distributing the events. In case of a cluster failure, all nodes continue to run independently. Be alarmed when your cluster fails and a Split-Brain-scenario is in effect - all alive instances continue to do their job, and history will begin to differ.

** Note **

Before you start, make sure to read the requirements.

Cluster Requirements

Before you start deploying, keep the following things in mind:

  • Your SSL CA and certificates are mandatory for secure communication
  • Get pen and paper or a drawing board and design your nodes and zones!
    • all nodes in a cluster zone are providing high availability functionality and trust each other
    • cluster zones can be built in a Top-Down-design where the child trusts the parent
    • communication between zones happens bi-directional which means that a DMZ-located node can still reach the master node, or vice versa
  • Update firewall rules and ACLs
  • Decide whether to use the built-in configuration syncronization or use an external tool (Puppet, Ansible, Chef, Salt, etc) to manage the configuration deployment

Tip

If you're looking for troubleshooting cluster problems, check the general troubleshooting section.

Cluster Naming Convention

The SSL certificate common name (CN) will be used by the ApiListener object to determine the local authority. This name must match the local Endpoint object name.

Example:

# icinga2-build-key icinga2a
...
Common Name (e.g. server FQDN or YOUR name) [icinga2a]:

# vim cluster.conf

object Endpoint "icinga2a" {
  host = "icinga2a.icinga.org"
}

The Endpoint name is further referenced as endpoints attribute on the Zone object.

object Endpoint "icinga2b" {
  host = "icinga2b.icinga.org"
}

object Zone "config-ha-master" {
  endpoints = [ "icinga2a", "icinga2b" ]
}

Specifying the local node name using the NodeName variable requires the same name as used for the endpoint name and common name above. If not set, the FQDN is used.

const NodeName = "icinga2a"

Certificate Authority and Certificates

Icinga 2 ships two scripts assisting with CA and node certificate creation for your Icinga 2 cluster.

Note

You're free to use your own method to generated a valid ca and signed client certificates.

Please make sure to export the environment variable ICINGA_CA pointing to an empty folder for the newly created CA files:

# export ICINGA_CA="/root/icinga-ca"

The scripts will put all generated data and the required certificates in there.

The first step is the creation of the certificate authority (CA) running the following command:

# icinga2-build-ca

Now create a certificate and key file for each node running the following command (replace icinga2a with the required hostname):

# icinga2-build-key icinga2a

Repeat the step for all nodes in your cluster scenario.

Save the CA key in a secure location in case you want to set up certificates for additional nodes at a later time.

Navigate to the location of your newly generated certificate files, and manually copy/transfer them to /etc/icinga2/pki in your Icinga 2 configuration folder.

Note

The certificate files must be readable by the user Icinga 2 is running as. Also, the private key file must not be world-readable.

Each node requires the following files in /etc/icinga2/pki (replace fqdn-nodename with the host's FQDN):

  • ca.crt
  • <fqdn-nodename>.crt
  • <fqdn-nodename>.key

Cluster Configuration

The following section describe which configuration must be updated/created in order to get your cluster running with basic functionality.

Once you're finished with the basic setup the following section will describe how to use zone configuration synchronisation and configure cluster scenarios.

Configure the Icinga Node Name

Instead of using the default FQDN as node name you can optionally set that value using the NodeName constant.

** Note **

Skip this step if your FQDN already matches the default NodeName set in /etc/icinga2/constants.conf.

This setting must be unique for each node, and must also match the name of the local Endpoint object and the SSL certificate common name as described in the cluster naming convention.

vim /etc/icinga2/constants.conf

/* Our local instance name. By default this is the server's hostname as returned by `hostname --fqdn`.
 * This should be the common name from the API certificate.
 */
const NodeName = "icinga2a"

Read further about additional naming conventions.

Not specifying the node name will make Icinga 2 using the FQDN. Make sure that all configured endpoint names and common names are in sync.

Configure the ApiListener Object

The ApiListener object needs to be configured on every node in the cluster with the following settings:

A sample config looks like:

object ApiListener "api" {
  cert_path = SysconfDir + "/icinga2/pki/" + NodeName + ".crt"
  key_path = SysconfDir + "/icinga2/pki/" + NodeName + ".key"
  ca_path = SysconfDir + "/icinga2/pki/ca.crt"
  accept_config = true
}

You can simply enable the api feature using

# icinga2-enable-feature api

Edit /etc/icinga2/features-enabled/api.conf if you require the configuration synchronisation enabled for this node. Set the accept_config attribute to true.

Note

The certificate files must be readable by the user Icinga 2 is running as. Also, the private key file must not be world-readable.

Configure Cluster Endpoints

Endpoint objects specify the host and port settings for the cluster nodes. This configuration can be the same on all nodes in the cluster only containing connection information.

A sample configuration looks like:

/**
 * Configure config master endpoint
 */

object Endpoint "icinga2a" {
  host = "icinga2a.icinga.org"
}

If this endpoint object is reachable on a different port, you must configure the ApiListener on the local Endpoint object accordingly too.

Configure Cluster Zones

Zone objects specify the endpoints located in a zone. That way your distributed setup can be seen as zones connected together instead of multiple instances in that specific zone.

Zones can be used for high availability, distributed setups and load distribution.

Each Icinga 2 Endpoint must be put into its respective Zone. In this example, you will define the zone config-ha-master where the icinga2a and icinga2b endpoints are located. The check-satellite zone consists of icinga2c only, but more nodes could be added.

The config-ha-master zone acts as High-Availability setup - the Icinga 2 instances elect one active master where all features are running on (for example icinga2a). In case of failure of the icinga2a instance, icinga2b will take over automatically.

object Zone "config-ha-master" {
  endpoints = [ "icinga2a", "icinga2b" ]
}

The check-satellite zone is a separated location and only sends back their checkresults to the defined parent zone config-ha-master.

object Zone "check-satellite" {
  endpoints = [ "icinga2c" ]
  parent = "config-ha-master"
}

Zone Configuration Synchronisation

By default all objects for specific zones should be organized in

/etc/icinga2/zones.d/<zonename>

on the configuration master.

Your child zones and endpoint members must not have their config copied to zones.d. The built-in configuration synchronisation takes care of that if your nodes accept configuration from the parent zone. You can define that in the ApiListener object by configuring the accept_config attribute accordingly.

You should remove the sample config included in conf.d by commenting the recursive_include statement in icinga2.conf:

//include_recursive "conf.d"

Better use a dedicated directory name like cluster or similar, and include that one if your nodes require local configuration not being synced to other nodes. That's useful for local health checks for example.

Note

In a high availability setup only one assigned node can act as configuration master. All other zone member nodes must not have the /etc/icinga2/zones.d directory populated.

These zone packages are then distributed to all nodes in the same zone, and to their respective target zone instances.

Each configured zone must exist with the same directory name. The parent zone syncs the configuration to the child zones, if allowed using the accept_config attribute of the ApiListener object.

Config on node icinga2a:

object Zone "master" {
  endpoints = [ "icinga2a" ]
}

object Zone "checker" {
  endpoints = [ "icinga2b" ]
  parent = "master"
}

/etc/icinga2/zones.d
  master
    health.conf
  checker
    health.conf
    demo.conf

Config on node icinga2b:

object Zone "master" {
  endpoints = [ "icinga2a" ]
}

object Zone "checker" {
  endpoints = [ "icinga2b" ]
  parent = "master"
}

/etc/icinga2/zones.d
  EMPTY_IF_CONFIG_SYNC_ENABLED

If the local configuration is newer than the received update Icinga 2 will skip the synchronisation process.

Note

zones.d must not be included in icinga2.conf. Icinga 2 automatically determines the required include directory. This can be overridden using the global constant ZonesDir.

Global Configuration Zone for Templates

If your zone configuration setup shares the same templates, groups, commands, timeperiods, etc. you would have to duplicate quite a lot of configuration objects making the merged configuration on your configuration master unique.

** Note **

Only put templates, groups, etc into this zone. DO NOT add checkable objects such as hosts or services here. If they are checked by all instances globally, this will lead into duplicated check results and unclear state history. Not easy to troubleshoot too - you've been warned.

That is not necessary by defining a global zone shipping all those templates. By setting global = true you ensure that this zone serving common configuration templates will be synchronized to all involved nodes (only if they accept configuration though).

Config on configuration master:

/etc/icinga2/zones.d
  global-templates/
    templates.conf
    groups.conf
  master
    health.conf
  checker
    health.conf
    demo.conf

In this example, the global zone is called global-templates and must be defined in your zone configuration visible to all nodes.

object Zone "global-templates" {
  global = true
}

Note

If the remote node does not have this zone configured, it will ignore the configuration update, if it accepts synchronized configuration.

If you don't require any global configuration, skip this setting.

Zone Configuration Synchronisation Permissions

Each ApiListener object must have the accept_config attribute set to true to receive configuration from the parent Zone members. Default value is false.

object ApiListener "api" {
  cert_path = SysconfDir + "/icinga2/pki/" + NodeName + ".crt"
  key_path = SysconfDir + "/icinga2/pki/" + NodeName + ".key"
  ca_path = SysconfDir + "/icinga2/pki/ca.crt"
  accept_config = true
}

If accept_config is set to false, this instance won't accept configuration from remote master instances anymore.

** Tip **

Look into the troubleshooting guides for debugging problems with the configuration synchronisation.

Cluster Health Check

The Icinga 2 ITL ships an internal check command checking all configured EndPoints in the cluster setup. The check result will become critical if one or more configured nodes are not connected.

Example:

object Service "cluster" {
    check_command = "cluster"
    check_interval = 5s
    retry_interval = 1s

    host_name = "icinga2a"
}

Each cluster node should execute its own local cluster health check to get an idea about network related connection problems from different points of view.

Additionally you can monitor the connection from the local zone to the remote connected zones.

Example for the checker zone checking the connection to the master zone:

object Service "cluster-zone-master" {
  check_command = "cluster-zone"
  check_interval = 5s
  retry_interval = 1s
  vars.cluster_zone = "master"

  host_name = "icinga2b"
}

Cluster Scenarios

All cluster nodes are full-featured Icinga 2 instances. You only need to enabled the features for their role (for example, a Checker node only requires the checker feature enabled, but not notification or ido-mysql features).

Security in Cluster Scenarios

While there are certain capabilities to ensure the safe communication between all nodes (firewalls, policies, software hardening, etc) the Icinga 2 cluster also provides additional security itself:

  • SSL certificates are mandatory for cluster communication.
  • Child zones only receive event updates (check results, commands, etc) for their configured updates.
  • Zones cannot influence/interfere other zones. Each checked object is assigned to only one zone.
  • All nodes in a zone trust each other.
  • Configuration sync is disabled by default.

Features in Cluster Zones

Each cluster zone may use all available features. If you have multiple locations or departments, they may write to their local database, or populate graphite. Even further all commands are distributed amongst connected nodes. For example, you could re-schedule a check or acknowledge a problem on the master, and it gets replicated to the actual slave checker node.

DB IDO on the left, graphite on the right side - works (if you disable DB IDO HA). Icinga Web 2 on the left, checker and notifications on the right side - works too. Everything on the left and on the right side - make sure to deal with load-balanced notifications and checks in a HA zone. configure-cluster-zones

Distributed Zones

That scenario fits if your instances are spread over the globe and they all report to a master instance. Their network connection only works towards the master master (or the master is able to connect, depending on firewall policies) which means remote instances won't see each/connect to each other.

All events (check results, downtimes, comments, etc) are synced to the master node, but the remote nodes can still run local features such as a web interface, reporting, graphing, etc. in their own specified zone.

Imagine the following example with a master node in Nuremberg, and two remote DMZ based instances in Berlin and Vienna. Additonally you'll specify global templates available in all zones.

The configuration tree on the master instance nuremberg could look like this:

zones.d
  global-templates/
    templates.conf
    groups.conf
  nuremberg/
    local.conf
  berlin/
    hosts.conf
  vienna/
    hosts.conf

The configuration deployment will take care of automatically synchronising the child zone configuration:

  • The master node sends zones.d/berlin to the berlin child zone.
  • The master node sends zones.d/vienna to the vienna child zone.
  • The master node sends zones.d/global-templates to the vienna and berlin child zones.

The endpoint configuration would look like:

object Endpoint "nuremberg-master" {
  host = "nuremberg.icinga.org"
}

object Endpoint "berlin-satellite" {
  host = "berlin.icinga.org"
}

object Endpoint "vienna-satellite" {
  host = "vienna.icinga.org"
}

The zones would look like:

object Zone "nuremberg" {
  endpoints = [ "nuremberg-master" ]
}

object Zone "berlin" {
  endpoints = [ "berlin-satellite" ]
  parent = "nuremberg"
}

object Zone "vienna" {
  endpoints = [ "vienna-satellite" ]
  parent = "nuremberg"
}

object Zone "global-templates" {
  global = true
}

The nuremberg-master zone will only execute local checks, and receive check results from the satellite nodes in the zones berlin and vienna.

Note

The child zones berlin and vienna will get their configuration synchronised from the configuration master 'nuremberg'. The endpoints in the child zones must not have their zones.d directory populated if this endpoint accepts synced configuration.

Load Distribution

If you are planning to off-load the checks to a defined set of remote workers you can achieve that by:

  • Deploying the configuration on all nodes.
  • Let Icinga 2 distribute the load amongst all available nodes.

That way all remote check instances will receive the same configuration but only execute their part. The master instance located in the master zone can also execute checks, but you may also disable the Checker feature.

Configuration on the master node:

zones.d/
  global-templates/
  master/
  checker/

If you are planning to have some checks executed by a specific set of checker nodes you have to define additional zones and define these check objects there.

Endpoints:

object Endpoint "master-node" {
  host = "master.icinga.org"
}

object Endpoint "checker1-node" {
  host = "checker1.icinga.org"
}

object Endpoint "checker2-node" {
  host = "checker2.icinga.org"
}

Zones:

object Zone "master" {
  endpoints = [ "master-node" ]
}

object Zone "checker" {
  endpoints = [ "checker1-node", "checker2-node" ]
  parent = "master"
}

object Zone "global-templates" {
  global = true
}

Note

The child zones checker will get its configuration synchronised from the configuration master 'master'. The endpoints in the child zone must not have their zones.d directory populated if this endpoint accepts synced configuration.

Cluster High Availability

High availability with Icinga 2 is possible by putting multiple nodes into a dedicated zone. All nodes will elect one active master, and retry an election once the current active master is down.

Selected features provide advanced HA functionality. Checks and notifications are load-balanced between nodes in the high availability zone.

Connections from other zones will be accepted by all active and passive nodes but all are forwarded to the current active master dealing with the check results, commands, etc.

object Zone "config-ha-master" {
  endpoints = [ "icinga2a", "icinga2b", "icinga2c" ]
}

Two or more nodes in a high availability setup require an initial cluster sync.

Note

Keep in mind that only one node acts as configuration master having the configuration files in the zones.d directory. All other nodes must not have that directory populated. Instead they are required to accept synced configuration. Details in the Configuration Sync Chapter.

Multiple Hierachies

Your master zone collects all check results for reporting and graphing and also does some sort of additional notifications. The customers got their own instances in their local DMZ zones. They are limited to read/write only their services, but replicate all events back to the master instance. Within each DMZ there are additional check instances also serving interfaces for local departments. The customers instances will collect all results, but also send them back to your master instance. Additionally the customers instance on the second level in the middle prohibits you from sending commands to the subjacent department nodes. You're only allowed to receive the results, and a subset of each customers configuration too.

Your master zone will generate global reports, aggregate alert notifications, and check additional dependencies (for example, the customers internet uplink and bandwidth usage).

The customers zone instances will only check a subset of local services and delegate the rest to each department. Even though it acts as configuration master with a master dashboard for all departments managing their configuration tree which is then deployed to all department instances. Furthermore the master NOC is able to see what's going on.

The instances in the departments will serve a local interface, and allow the administrators to reschedule checks or acknowledge problems for their services.

High Availability for Icinga 2 features

All nodes in the same zone require the same features enabled for High Availability (HA) amongst them.

By default the following features provide advanced HA functionality:

  • Checks (load balanced, automated failover)
  • Notifications (load balanced, automated failover)
  • DB IDO (Run-Once, automated failover)

High Availability with Checks

All nodes in the same zone load-balance the check execution. When one instance fails the other nodes will automatically take over the reamining checks.

Note

If a node should not check anything, disable the checker feature explicitely and reload Icinga 2.

# icinga2-disable-feature checker
# service icinga2 reload

High Availability with Notifications

Notifications are load balanced amongst all nodes in a zone. By default this functionality is enabled. If your nodes should notify independent from any other nodes (this will cause duplicated notifications if not properly handled!), you can set enable_ha = false in the NotificationComponent feature.

High Availability with DB IDO

All instances within the same zone (e.g. the master zone as HA cluster) must have the DB IDO feature enabled.

Example DB IDO MySQL:

# icinga2-enable-feature ido-mysql
The feature 'ido-mysql' is already enabled.

By default the DB IDO feature only runs on the elected zone master. All other passive nodes disable the active IDO database connection at runtime.

Note

The DB IDO HA feature can be disabled by setting the enable_ha attribute to false for the IdoMysqlConnection or IdoPgsqlConnection object on all nodes in the same zone.

All endpoints will enable the DB IDO feature then, connect to the configured database and dump configuration, status and historical data on their own.

If the instance with the active DB IDO connection dies, the HA functionality will re-enable the DB IDO connection on the newly elected zone master.

The DB IDO feature will try to determine which cluster endpoint is currently writing to the database and bail out if another endpoint is active. You can manually verify that by running the following query:

icinga=> SELECT status_update_time, endpoint_name FROM icinga_programstatus;
   status_update_time   | endpoint_name
------------------------+---------------
 2014-08-15 15:52:26+02 | icinga2a
(1 Zeile)

This is useful when the cluster connection between endpoints breaks, and prevents data duplication in split-brain-scenarios. The failover timeout can be set for the failover_timeout attribute, but not lower than 60 seconds.

Add a new cluster endpoint

These steps are required for integrating a new cluster endpoint:

Initial Cluster Sync

In order to make sure that all of your cluster nodes have the same state you will have to pick one of the nodes as your initial "master" and copy its state file to all the other nodes.

You can find the state file in /var/lib/icinga2/icinga2.state. Before copying the state file you should make sure that all your cluster nodes are properly shut down.

Host With Multiple Cluster Nodes

Special scenarios might require multiple cluster nodes running on a single host. By default Icinga 2 and its features will place their runtime data below the prefix LocalStateDir. By default packages will set that path to /var. You can either set that variable as constant configuration definition in icinga2.conf or pass it as runtime variable to the Icinga 2 daemon.

# icinga2 -c /etc/icinga2/node1/icinga2.conf -DLocalStateDir=/opt/node1/var