Configurations

This page introduces best practices for setting up values for Vald Helm Chart. Before reading, please read the overview of Vald Helm Chart in its README.

Notable fields in Vald Helm Chart

Basics

Specify image tag

It is highly recommended to specify Vald version. You can specify image version by set image.tag field in each component ([component].image.tag) or defaults section.

defaults:
  image:
    tag: v1.1.1

or you can use the older image only for agent,

agent:
  image:
    tag: v1.1.0

Specify appropriate logging level and format

The default logging levels and formats are configured in defaults.logging.level and defaults.logging.format. You can also specify them in each component section ([component].logging).

defaults:
  logging:
    level: info
    format: raw

you can specify log level debug and JSON format for lb-gateway by the followings:

gateway:
  lb:
    logging:
      level: debug
      format: json

The logging level is defined in the Coding Style Guide.

Servers

Each Vald component has several types of servers. They can be configured by specifying the values in defaults.server_config. They can be overwritten by specifying [component].server_config.

Examples:

defaults:
  server_config:
    servers:
      grpc:
        enabled: true
        host: 0.0.0.0
        port: 8081
        servicePort: 8081
        server:
          mode: GRPC
          ...
gateway:
  lb:
    server_config:
      servers:
        rest:
          enabled: true
          host: 0.0.0.0
          port: 8080
          servicePort: 8080
          server:
            mode: REST
            ...
gRPC server

gRPC server should be enabled, because all Vald components use gRPC to communicate with others. The API specs are placed in apis/docs.

REST server

REST server is optional. The swagger specs are placed in apis/swagger.

Health check servers

There are two types of built-in health check servers, liveness and readiness. They are used as servers for Kubernetes liveness and readiness probe. By default, liveness servers are disabled for agent, because the liveness probes may accidentally kill it.

agent:
  server_config:
    healths:
      liveness:
        enabled: false
Metrics servers

Metrics servers are useful for debugging and monitoring Vald components. There are two types of metrics servers, pprof and Prometheus.

pprof server is implemented using Go’s net/http/pprof package. You can use google’s pprof to analyze the profiling data exported from it.

Prometheus server is a Prometheus exporter. It is required to set the observability section on each Vald component to enable the monitoring using Prometheus. Plese refer to the next section.

Observability

The observability features are useful for monitoring Vald components. They can be enabled by setting the value true on the defaults.observability.enabled field or override it in each component ([component].observability.enabled). And also, enable each feature by setting the value true on its enabled field.

If observability features are enabled, the metrics will be collected periodically. The duration can be set on observability.collector.duration. Please refer to Vald operation guide for more detail.

Agents

NGT

Agent-NGT uses yahoojapan/NGT as a core library for searching vector. The behaviors of NGT can be configured by setting agent.ngt field object.

The important parameters are the followings:

  • agent.ngt.dimension
  • agent.ngt.distance_type
  • agent.ngt.object_type

Users should configure these parameters first to fit to their use case.

For further details, please read NGT wiki.

Agent-NGT has a feature to start indexing automatically. The behavior of this feature can be configured with these parameters:

  • agent.ngt.auto_index_duration_limit
  • agent.ngt.auto_index_check_duration
  • agent.ngt.auto_index_length

Resource requests and limits, Pod priorities

Because agent places indices on memory, termination of agent pods causes loss of indices. It is important to set resource requests and limits appropriately not to terminate agent pods.

It is highly recommended to request a totally 40% of cluster memory for agent pods. And also it is highly recommended not to set resource limits to agent pods.

Pod priorities are also useful for saving agent pods from eviction. By default, very high priority is set to agent pods in the Chart.

Pod scheduling

It is recommended to schedule agent pods on different nodes as much as possible. To achieve this, the following podAntiAffinity is set by default.

agent:
  affinity:
    podAntiAffinity:
      preferredDuringSchedulingIgnoredDuringExecution:
        - weight: 100
          podAffinityTerm:
            topologyKey: kubernetes.io/hostname
            labelSelector:
              matchExpressions:
                - key: app
                  operator: In
                  values:
                    - vald-agent-ngt

It can be also achieved by using pod topology spread constraints.

agent:
  topologySpreadConstraints:
    - topologyKey: node
      maxSkew: 1
      whenUnsatisfiable: ScheduleAnyway
      labelSelector:
        matchLabels:
          app: vald-agent-ngt
  affinity:
    podAntiAffinity:
      preferredDuringSchedulingIgnoredDuringExecution: [] # to disable default settings

Gateways

Ingress

Ingress for gateways can be configured by gateway.{backup,filter,lb,meta}.ingress field object. It is important to set your host to gateway.{backup,filter,lb,meta}.ingress.host field. gateway.{backup,filter,lb,meta}.ingress.servicePort should be grpc or rest.

gateway:
  lb:
    ingress:
      enabled: true
      host: vald.vdaas.org # Set correct hostname here
      servicePort: grpc

Index replica

gateway.lb.gateway_config.index_replica represents how many agent pods that a vector will be inserted into.

gateway:
  lb:
    gateway_config:
      index_replica: 3

Resource requests and limits

Gateway’s resource requests and limits depend on the request traffic and available resources. If the request traffic varies largely, it is recommended to enable HPA for gateway and adjust the resource requests.

Discoverer request duration

gateway.lb.gateway_config.discoverer.duration represents a frequency to send requests to discoverer. If discoverer’s CPU utilization is too high, make this value longer or reduce the number of LB gateway pods.

gateway:
  lb:
    gateway_config:
      discoverer:
        duration: 2s

Discoverer

Resource requests and limits

The number of discoverer pods and resource limits can be estimated by the configurations of your LB gateways and index managers because its APIs are called by them. Discoverer CPU loads almost depend on API request traffic = (the number of LB gateways x its request frequency) + (the number of index managers x its request frequency).

Index Manager

Discoverer request duration

Same as LB gateway, manager.index.indexer.discoverer.duration represents a frequency to send requests to discoverer.

References

For further details, there are references of Helm values in GitHub Vald repository.