Implementation:SeldonIO Seldon core Seldon Pipeline CRD Monitoring

Property	Value
Implementation Name	Seldon Pipeline CRD Monitoring
Type	Pattern Doc
Overview	Concrete pattern for declaring monitoring pipelines with batch processing and multi-output in Seldon Core 2
Domains	MLOps, Data_Flow
Implements Principle	SeldonIO_Seldon_core_Monitoring_Pipeline_Definition
Source	samples/pipelines/income.yaml:L1-18
External Dependencies	Kubernetes API (mlops.seldon.io/v1alpha1), Kafka
Knowledge Sources	Repo (https://github.com/SeldonIO/seldon-core), Doc (https://docs.seldon.io/projects/seldon-core/en/v2/)
Last Updated	2026-02-13 00:00 GMT

Code Reference

Monitoring Pipeline CRD (income.yaml)

apiVersion: mlops.seldon.io/v1alpha1
kind: Pipeline
metadata:
  name: income-production
spec:
  steps:
    - name: income
    - name: income-preprocess
    - name: income-outlier
      inputs:
      - income-preprocess
    - name: income-drift
      batch:
        size: 20
  output:
    steps:
    - income
    - income-outlier.outputs.is_outlier

Key Parameters

Parameter	Path	Description	Example Value
Pipeline name	metadata.name	Unique pipeline identifier	income-production
Steps	spec.steps	List of pipeline model steps	4 steps (income, income-preprocess, income-outlier, income-drift)
Step inputs	spec.steps[].inputs	Dependency chain for a step	income-outlier depends on income-preprocess
Batch size	spec.steps[].batch.size	Batch aggregation count for drift detection	20
Output steps	spec.output.steps	Which step outputs to return to caller	income (predictions), income-outlier.outputs.is_outlier (outlier flags)

Pipeline Data Flow

Input Request
  |
  +---> income (classifier) ---------> output: predictions
  |
  +---> income-preprocess (transform)
  |         |
  |         +---> income-outlier ----> output: is_outlier
  |
  +---> income-drift (batched, size=20) --> async drift results

I/O Contract

Inputs

Input	Type	Description
Deployed model names	String references	All four models must be loaded: income, income-preprocess, income-drift, income-outlier
Pipeline architecture	Design decision	Which steps depend on which, batch sizes, output selection

Outputs

Output	Type	Description
Pipeline CRD YAML	Kubernetes resource	income-production pipeline with 4 steps
Kafka topics	Auto-created	Internal data flow topics for inter-step communication

Usage Examples

Deploying the Monitoring Pipeline

# Load the monitoring pipeline definition
seldon pipeline load -f samples/pipelines/income.yaml

# Wait for the pipeline to become ready
seldon pipeline status income-production -w PipelineReady

Understanding Step Dependencies

Step	Input Source	Output Destination	Batching
income	Pipeline input (raw features)	Pipeline output (predictions)	None
income-preprocess	Pipeline input (raw features)	income-outlier	None
income-outlier	income-preprocess output	Pipeline output (is_outlier)	None
income-drift	Pipeline input (raw features)	Async (not in pipeline output)	Batch size 20

Output Field Selection

The income-outlier.outputs.is_outlier syntax selects a specific field from the outlier detector's output. The full outlier detector response includes multiple fields (instance_score, feature_score, is_outlier), but only the binary outlier flag is included in the pipeline response.

Related Pages

• SeldonIO_Seldon_core_Monitoring_Pipeline_Definition (principle) - Design principles for monitoring pipeline composition
• SeldonIO_Seldon_core_Seldon_Model_Load_For_Monitoring (prerequisite) - Loading component models before pipeline creation
• SeldonIO_Seldon_core_Seldon_Pipeline_Load_And_Status (next step) - Deploying and validating the pipeline
• SeldonIO_Seldon_core_Seldon_Pipeline_Infer_Monitoring (usage) - Sending inference requests through the monitoring pipeline
• SeldonIO_Seldon_core_Seldon_Pipeline_Version_Update (related) - Evolving pipeline versions progressively
• Environment:SeldonIO_Seldon_core_Kafka_Messaging_Environment