vSphere is a VMware product dedicated to managing a (usually) on-premise infrastructure. From physical machines running VMware ESXi that are called ESXi Hosts, users can spin up or migrate Virtual Machines from one host to another.
vSphere is an integrated solution and provides an easy managing interface over concepts like data storage, or computing resource.
This section details some of vSphere specific elements. This section does not intend to be an extensive list, but rather a place for those unfamiliar with the product to have the basics required to understand how the Datadog integration works.
- vSphere - The complete suite of tools and technologies detailed in this article.
- vCenter server - The main machine which controls ESXi hosts and provides both a web UI and an API to control the vSphere environment.
- vCSA (vCenter Server Appliance) - A specific kind of vCenter where the software runs in a dedicated Linux machine (more recent). By opposition, the legacy vCenter is typically installed on an existing Windows machine.
- ESXi host - The physical machine controlled by vCenter where the ESXi (bare-metal) virtualizer is installed. The host boots a minimal OS that can run Virtual Machines.
- VM - What anyone using vSphere really needs in the end, instances that can run applications and code. Note: Datadog monitors both ESXi hosts and VMs and it calls them both "host" (they are in the host map).
- Attributes/tags - It is possible to add attributes and tags to any vSphere resource, note that those two are now very similar with "attributes" being the deprecated thing to use.
- Datacenter - A set of resources grouped together. A single vCenter server can handle multiple datacenters.
- Datastore - A virtual vSphere concept to represent data storing capabilities. It can be an NFS server that ESXi hosts have read/write access to, it can be a mounted disk on the host and more. Datastores are often shared between multiple hosts. This allows Virtual Machines to be migrated from one host to another.
- Cluster - A logical grouping of computational resources, you can add multiple ESXi hosts in your cluster and then you can create VM in the cluster (and not on a specific host, vSphere will take care of placing your VM in one of the ESXi hosts and migrating it when needed).
- Photon OS - An open-source minimal Linux distribution and used by both ESXi and vCSA as a base.
The Datadog vSphere integration runs from a single agent and pulls all the information from a single vCenter endpoint. Because the agent cannot run directly on Photon OS, it is usually required that the agent runs within a dedicated VM inside the vSphere infrastructure.
Once the agent is running, the minimal configuration (as of version 5.x) is as follows:
init_config: instances: - host: username: password: use_legacy_check_version: false empty_default_hostname: true
hostis the endpoint used to access the vSphere Client from a web browser. The host is either a FQDN or an IP, not an http url.
passwordare the credentials to log in to vCenter.
use_legacy_check_versionis a backward compatibility flag. It should always be set to false and this flag will be removed in a future version of the integration. Setting it to true tells the agent to use an older and deprecated version of the vSphere integration.
empty_default_hostnameis a field used by the agent directly (and not the integration). By default, the agent does not allow submitting metrics without attaching an explicit host tag unless this flag is set to true. The vSphere integration uses that behavior for some metrics and service checks. For example, the
vsphere.vm.countmetric which gives a count of the VMs in the infra is not submitted with a host tag. This is particularly important if the agent runs inside a vSphere VM. If the
vsphere.vm.countwas submitted with a host tag, the Datadog backend would attach all the other host tags to the metric, for example
vsphere_host:<NAME_OF_THE_ESX_HOST>which makes the metric almost impossible to use.
vSphere metrics are documented in their documentation page an each metric has a defined "collection level".
That level determines the amount of data gathered by the integration and especially which metrics are available. More details here.
By default, only the level 1 metrics are collected but this can be increased in the integration configuration file.
Realtime vs historical¶
Each ESXi host collects and stores data for each metric on himself and every VM it hosts every 20 seconds. Those data points are stored for up to one hour and are called realtime. Note: Each metric concerns always either a VM or an ESXi hosts. Metrics that concern datastore for example are not collected in the ESXi hosts.
Additionally, the vCenter server collects data from all the ESXi hosts and stores the datapoint with some aggregation rollup into its own database. Those data points are called "historical".
Finally, the vCenter server also collects metrics for other kinds of resources (like Datastore, ClusterComputeResource, Datacenter...) Those data points are necessarily "historical".
The reason for such an important distinction is that historical metrics are much MUCH slower to collect than realtime metrics. The vSphere integration will always collect the "realtime" data for metrics that concern ESXi hosts and VMs. But the integration also collects metrics for Datastores, ClusterComputeResources, Datacenters, and maybe others in the future.
That's why, in the context of the Datadog vSphere integration, we usually simplify by considering that:
VMs and ESXi hosts are "realtime resources". Metrics for such resources are quick and easy to get by querying vCenter that will in turn query all the ESXi hosts.
Datastores, ClusterComputeResources, and Datacenters are "historical resources" and are much slower to collect.
To collect all metrics (realtime and historical), it is advised to use two "check instances". One with
collection_type: realtime and one with
collection_type: historical . This way all metrics will be collected but because both check instances are on different schedules, the slowness of collecting historical metrics won't affect the rate at which realtime metrics are collected.
vSphere tags and attributes¶
Similarly to how Datadog allows you to add tags to your different hosts (thins like the
os or the
instance-type of your machines), vSphere has "tags" and "attributes".
A lot of details can be found here: https://docs.vmware.com/en/VMware-vSphere/7.0/com.vmware.vsphere.vcenterhost.doc/GUID-E8E854DD-AA97-4E0C-8419-CE84F93C4058.html#:~:text=Tags%20and%20attributes%20allow%20you,that%20tag%20to%20a%20category.
But the overall idea is that both tags and attributes are additional information that you can attach to your vSphere resources and that "tags" are newer and more featureful than "attributes".
A very flexible filtering system has been implemented with the vSphere integration.
This allows fine-tuned configuration so that:
- You only pay for the host and VMs you really want to monitor.
- You reduce the load on your vCenter server by running just the queries that you need.
- You improve the check runtime which otherwise increases linearly with the size of their infrastructure and that was seen to take up to 10min in some large environments.
We provide two types of filtering, one based on metrics, the other based on resources.
The metric filter is fairly simple, for each resource type, you can provide some regexes. If a metric match any of the filter, it will be fetched and submitted. The configuration looks like this:
metric_filters: vm: - cpu\..* - mem\..* host: - WHATEVER # Excludes everything datacenter: - .*
The resource filter on the other hand, allows to exclude some vSphere resources (VM, ESXi host, etc.), based on an "attribute" of that resource. The possible attributes as of today are: -
name, literally the name of the resource (as defined in vCenter) -
inventory_path, a path-like string that represents the location of the resource in the inventory tree as each resource only ever has a single parent and recursively up to the root. For example:
tag, see the
tags and attributes section. Used to filter resources based on the attached tags. -
attribute, see the
tags and attributes section. Used to filter resources based on the attached attributes. -
hostname (only for VMs), the name of the ESXi host where the VM is running. -
guest_hostname (only for VMs), the name of the OS as reported from within the machine. VMware tools have to be installed on the VM otherwise, vCenter is not able to fetch this information.
A possible filtering configuration would look like this:
resource_filters: - resource: vm property: name patterns: - <VM_REGEX_1> - <VM_REGEX_2> - resource: vm property: hostname patterns: - <HOSTNAME_REGEX> - resource: vm property: tag type: blacklist patterns: - '^env:staging$' - resource: vm property: tag type: whitelist # type defaults to whitelist patterns: - '^env:.*$' - resource: vm property: guest_hostname patterns: - <GUEST_HOSTNAME_REGEX> - resource: host property: inventory_path patterns: - <INVENTORY_PATH_REGEX>
In vSphere each metric is defined by three "dimensions".
- The resource on which the metric applies (for example the VM called "abc1")
- The name of the metric (for example cpu.usage).
- An additional available dimension that varies between metrics. (for example the cpu core id)
This is similar to how Datadog represent metrics, except that the context cardinality is limited to two "keys", the name of the resource (usually the "host" tag), and there is space for one additional tag key.
This available tag key is defined as the "instance" property, or "instance tag" in vSphere, and this dimension is not collected by default by the Datadog integration as it can have too big performance implications in large systems when compared to their added value from a monitoring perspective.
Also when fetching metrics with the instance tag, vSphere only provides the value of the instance tag, it doesn't expose a human-readable "key" for that tag. In the
cpu.usage metric with the core_id as the instance tag, the integration has to "know" that the meaning of the instance tag and that's why we rely on a hardcoded list in the integration.
Because this instance tag can provide additional visibility, it is possible to enable it for some metrics from the configuration. For example, if we're really interested in getting the usage of the cpu per core, the setup can look like this:
collect_per_instance_filters: vm: - cpu\.usage\..*