LogoLogo
v2.9
latestv2.9v2.8
v2.9
latestv2.9v2.8
  • About
    • Production-ready ML Serving Framework
    • Seldon Core Features
    • Concepts
    • Architecture
  • Installation
    • Installation Overview
    • Learning Environment
      • Self-hosted Kafka
    • Production Environment
      • Kafka Integration
      • Managed Kafka
      • Ingress Controller
    • Test the Installation
    • Advanced Configurations
      • Helm Configuration
      • Server Config
      • Server Runtime
      • Seldon Config
      • Pipeline Config
    • Upgrading
  • User Guide
    • Kubernetes Resources
    • Servers
      • Resource allocation
        • Example: Serving models on dedicated GPU nodes
    • Models
      • Multi-Model Serving
      • Inference Artifacts
      • Scheduling
      • rClone
      • Parameterized Models
      • Pandas Query
      • Storage Secrets
    • Inference
      • Inference Server
      • Run Inference
      • OIP
      • Batch
    • Pipelines
    • Scaling
      • Model Autoscaling
      • Server Autoscaling
      • HPA Autoscaling in single-model serving
    • Data Science Monitoring
      • Dataflow with Kafka
      • Model Performance Metrics
      • Drift Detection
      • Outlier Detection
      • Explainability
    • Operational Monitoring
      • Operational Metrics
      • Observability
      • Usage Metrics
      • Local Metrics
      • Tracing
    • Experiments
    • Examples
      • Local examples
      • Kubernetes examples
      • Huggingface models
      • Model zoo
      • Artifact versions
      • Pipeline examples
      • Pipeline to pipeline examples
      • Explainer examples
      • Custom Servers
      • Local experiments
      • Experiment version examples
      • Inference examples
      • Tritonclient examples
      • Batch Inference examples (kubernetes)
      • Batch Inference examples (local)
      • Checking Pipeline readiness
      • Multi-Namespace Kubernetes
      • Huggingface speech to sentiment with explanations pipeline
      • Production image classifier with drift and outlier monitoring
      • Production income classifier with drift, outlier and explanations
      • Conditional pipeline with pandas query model
      • Kubernetes Server with PVC
  • Integrations
    • Service Meshes
      • Ambassador
      • Istio
      • Traefik
      • Secure Model Endpoints
  • Resources
    • Security
    • APIs
      • Internal
        • Chainer
        • Agent
      • Inference
        • Open Inference Protocol
      • Scheduler
      • Seldon CLI
        • CLI
        • Seldon
          • Config
          • Config Activate
          • Config Deactivate
          • Config Add
          • Config List
          • Config Remove
        • Experiment
          • Experiment Start
          • Experiment Status
          • Experiment List
          • Experiment Stop
        • Model
          • Model Status
          • Model Load
          • Model List
          • Model Infer
          • Model Metadata
          • Model Unload
        • Pipeline
          • Pipeline Load
          • Pipeline Status
          • Pipeline List
          • Pipeline Inspect
          • Pipeline Infer
          • Pipeline Unload
        • Server
          • Server List
          • Server Status
    • FAQs
    • Development
      • License
      • Release
Powered by GitBook
On this page
  • Upgrading from 2.7 - 2.8
  • Upgrading from 2.6 - 2.7
  • Upgrading from 2.5 - 2.6

Was this helpful?

Edit on GitHub
Export as PDF
  1. Installation

Upgrading

PreviousPipeline ConfigNextKubernetes Resources

Last updated 10 days ago

Was this helpful?

Upgrading from 2.7 - 2.8

Core 2.8 introduces several new fields in our CRDs:

  • statefulSetPersistentVolumeClaimRetentionPolicy enables users to configure the cleaning of PVC on their servers. This field is set to retain as default.

  • Status.selector was introduced as a mandatory field for models in 2.8.4 and made optional in 2.8.5. This field enables autoscaling with HPA.

  • PodSpec in the OverrideSpec for SeldonRuntimes enables users to customize how Seldon Core 2 pods are created. In particular, this also allows for setting custom taints/tolerations, adding additional containers to our pods, configuring custom security settings.

These added fields do not result in breaking changes, apart from 2.8.4 which required the setting of the Status.selector upon upgrading. This field was however changed to optional in the subsequent 2.8.5 release. Updating the CRDs (e.g. via helm) will enable users to benefit from the associated functionality.

Upgrading from 2.6 - 2.7

All pods provisioned through the operator i.e. SeldonRuntime and Server resources now have the label app.kubernetes.io/name for identifying the pods.

Previously, the labelling has been inconsistent across different versions of Seldon Core 2, with mixture of app and app.kubernetes.io/name used.

If using the Prometheus operator ("Kube Prometheus"), please apply the v2.7.0 manifests for Seldon Core 2 according to the .

Note that these manifests need to be adjusted to discover metrics endpoints based on the existing setup.

If previous pod monitors had namespaceSelector fields set, these should be copied over and applied to the new manifests.

If namespaces do not matter, cluster-wide metrics endpoint discovery can be setup by modifying thenamespaceSelector field in the pod monitors:

spec:
  namespaceSelector:
    any: true

Upgrading from 2.5 - 2.6

Release 2.6 brings with it new custom resources SeldonConfig and SeldonRuntime, which provide a new way to install Seldon Core 2 in Kubernetes. Upgrading in the same namespace will cause downtime while the pods are being recreated. Alternatively users can have an external service mesh or other means to be used over multiple namespaces to bring up the system in a new namespace and redeploy models before switch traffic between them.

If the new 2.6 charts are used to upgrade in an existing namespace models will eventually be redeloyed but there will be service downtime as the core components are redeployed.

metrics documentation