services.xml: Vespa Cloud

services.xml is the primary Vespa configuration file. This documents services.xml amendments for Vespa Cloud - see services.xml at docs.vespa.ai for the general reference.

<nodes>

In cloud applications nodes are specified by count and node resources. Example:

<nodes count="4">
    <resources vcpu="8" memory="16Gb" disk="200Gb"/>
</nodes>

Resources must match a node flavor on the cloud(s) you are deploying to, see AWS flavors and GCP flavors.

Subelements: resources

Attributes:

• count: An integer or range. The number of nodes of the cluster.
• exclusive (optional): true or false (default false). If true these nodes will never be placed on shared hosts even when this would otherwise be allowed (which is only for content nodes in some environments). When nodes are allocated exclusively, the resources must match the resources of the host exactly.

In addition there are some attributes for specific cluster types, listed below.

<nodes> for <content>

• groups (optional): Integer or range. Sets the number of groups into which content nodes should be divided. Each group will have an equal share of the nodes and redundancy copies of the corpus, and each query will be routed to just one group. This allows scaling to a higher query load than within a single group.

• group-size (optional): Integer or range where either value can be skipped (replaced by an empty string) to create a one-sided limit. If this is set, the group sizes used will always be within these limits (inclusive).

  If neither groups nor group-size is set, all nodes will always be placed in a single group.

<nodes> for <cluster-controllers> <slobroks> and <logservers>

The nodes element nested in these elements allow specifying whether the nodes used should be dedicated to the service or if it should run on existing nodes. Attribute:

• dedicated (optional): true or false (default false). Whether separate nodes should be allocated for this service or not.

<resources>

Contained in the nodes element, specifies the resources available on each node. The resources must match a node flavor on AWS, GCP, or both, depending on which zones you are deploying to. Exception: If you use remote disk, you can specify any number lower than the max size.

Any element not specified will be assigned a default value.

Subelements: gpu

See index bootstrap for how to set resources in a step-by-step procedure, estimating settings by feeding smaller subsets. Note that autoscaling of content clusters involves data redistribution and cannot speed up bootstrapping.

<gpu>

Declares GPU resources to provision.

Current limitations:

• Available in AWS zones only
• Valid for container clusters only
• Only one resources and gpu configuration is supported, see example. The example configuration will provision a node with a NVIDIA T4 GPU

Subelements: None

Example:

<nodes count="[2, 4]">
    <resources vcpu="4" memory="16Gb" disk="125Gb">
        <gpu count="1" memory="16Gb"/>
    </resources>
</nodes>

Autoscaling ranges

Resources specified as a range will be autoscaled by the system. Ranges are expressed by the syntax [lower-limit, upper-limit]. Both limits are inclusive.

Autoscaling will attempt to keep utilization of all allocated resources close to ideal, and will automatically reconfigure to the cheapest option allowed by the ranges when necessary.

The ideal utilization takes into account that a node may be down or failing, that another region may be down causing doubling of traffic, and that we need headroom for maintenance operations and handling requests with low latency. It acts on what it has observed on your system in the recent past. If you need much more capacity in the near future than you do currently, you may want to set the lower limit to take this into account. Upper limits should be set to the maximum size that makes business sense.

When a new cluster (or application) is deployed it will initially be configured with the minimal resources given by the ranges. When autoscaling is turned on for an existing cluster, it will continue unchanged until autoscaling determines that a change is beneficial.

Examples:

Autoscaling node count:

<nodes count="[4, 8]">
    <resources vcpu="8" memory="16Gb" disk="200Gb"/>
</nodes>

Autoscaling on all resources:

<nodes count="[4, 8]" groups="[1, 2]">
    <resources vcpu="[4, 16]" memory="[32Gb, 150Gb]" disk="[100Gb, 1Tb]"/>
</nodes>

Elements with different meaning on Vespa Cloud

<redundancy>

<redundancy> sets the number of data copies in each group, not in total.

It is usually preferable to set <min-redundancy> instead - especially with autoscaling.

Ignored elements

The admin element is ignored, with one exception: when using metrics - example:

<admin>
    <metrics>
        <consumer id="my-custom-consumer">
            <metric-set id="default" />
            <metric id="vds.idealstate.garbage_collection.documents_removed.count" />
        </consumer>
    </metrics>
</admin>

When migrating from self-hosted to Vespa Cloud, one can safely remove admin, unless one want to configure a metric-set.

Deployment variants

services.xml settings can be made to vary by tags, instance, environment and region, see deployment variants.