KServe Multi Model Serving

InferenceDB supports KServe's Multi Model Serving using the filters object in InferenceLogger. See the full example here.

Step 1: Kafka Broker

First, we will need a Kafka broker to collect all KServe inference requests and responses:

apiVersion: eventing.knative.dev/v1
kind: Broker
metadata:
  name: sklearn-mms-broker
  namespace: default
  annotations:
    eventing.knative.dev/broker.class: Kafka
spec:
  config:
    apiVersion: v1
    kind: ConfigMap
    name: inferencedb-kafka-broker-config
    namespace: knative-eventing
---
apiVersion: v1
kind: ConfigMap
metadata:
  name: inferencedb-kafka-broker-config
  namespace: knative-eventing
data:
  # Number of topic partitions
  default.topic.partitions: "8"
  # Replication factor of topic messages.
  default.topic.replication.factor: "1"
  # A comma separated list of bootstrap servers. (It can be in or out the k8s cluster)
  bootstrap.servers: "kafka-cp-kafka.default.svc.cluster.local:9092"

Step 2: InferenceService

Next, we will create an InferenceService that will serve multiple TrainedModel objects:

apiVersion: serving.kserve.io/v1beta1
kind: InferenceService
metadata:
  name: sklearn-mms
spec:
  predictor:
    minReplicas: 1
    logger:
      mode: all
      url: http://kafka-broker-ingress.knative-eventing.svc.cluster.local/default/sklearn-mms-broker
    sklearn:
      name: sklearn-mms
      protocolVersion: v2
      resources:
        limits:
          cpu: 500m
          memory: 1Gi
        requests:
          cpu: 500m
          memory: 1Gi

Note the logger section - you can read more about it in the KServe documentation.

Step 3: TrainedModels

You can now add some models to the new InferenceService:

apiVersion: serving.kserve.io/v1alpha1
kind: TrainedModel
metadata:
  name: model1
spec:
  inferenceService: sklearn-mms
  model:
    storageUri: gs://seldon-models/sklearn/mms/lr_model
    framework: sklearn
    memory: 512Mi
---
apiVersion: serving.kserve.io/v1alpha1
kind: TrainedModel
metadata:
  name: model2
spec:
  inferenceService: sklearn-mms
  model:
    storageUri: gs://seldon-models/sklearn/mms/lr_model
    framework: sklearn
    memory: 512Mi

Step 4: InferenceLoggers

Finally, we can log the predictions of our new models using InferenceDB:

apiVersion: inferencedb.aporia.com/v1alpha1
kind: InferenceLogger
metadata:
  name: sklearn-mms-model-1
  namespace: default
spec:
  # NOTE: The format is knative-broker-<namespace>-<brokerName>
  topic: knative-broker-default-sklearn-mms-broker
  schema:
    type: avro
    config:
      columnNames:
        inputs: [sepal_width, petal_width, sepal_length, petal_length]
        outputs: [flower]
  events:
    type: kserve
    config: {}
  filters:
    modelName: model1
    # modelVersion: v1
  destination:
    type: confluent-s3
    config:
      url: s3://aporia-data/inferencedb
      format: parquet
---
apiVersion: inferencedb.aporia.com/v1alpha1
kind: InferenceLogger
metadata:
  name: sklearn-mms-model-2
  namespace: default
spec:
  # NOTE: The format is knative-broker-<namespace>-<brokerName>
  topic: knative-broker-default-sklearn-mms-broker
  schema:
    type: avro
    config:
      columnNames:
        inputs: [sepal_width, petal_width, sepal_length, petal_length]
        outputs: [flower]
  events:
    type: kserve
    config: {}
  filters:
    modelName: model2
    # modelVersion: v2
  destination:
    type: confluent-s3
    config:
      url: s3://aporia-data/inferencedb
      format: parquet

Note the usage of filters.

Step 4: Send requests

First, we will need to port-forward the Istio service so we can access it from our local machine:

kubectl port-forward --namespace istio-system svc/istio-ingressgateway 8080:80

Prepare a payload in a file called mms-input.json:

{
  "inputs": [
    {
      "name": "input-0",
      "shape": [2, 4],
      "datatype": "FP32",
      "data": [
        [6.8, 2.8, 4.8, 1.4],
        [6.0, 3.4, 4.5, 1.6]
      ]
    }
  ]
}

And finally, you can send some inference requests:

SERVICE_HOSTNAME=$(kubectl get inferenceservice sklearn-mms -o jsonpath='{.status.url}' | cut -d "/" -f 3)

curl -v \
  -H "Host: ${SERVICE_HOSTNAME}" \
  -H "Content-Type: application/json" \
  -d @./iris-input.json \
  http://localhost:8080/v2/models/model1/infer

Step 5: Success!

If everything was configured correctly, these predictions should have been logged to two Parquet files in S3.

import pandas as pd

df = pd.read_parquet("s3://aporia-data/inferencedb/default-sklearn-mms-model1/")
print(df) 

df = pd.read_parquet("s3://aporia-data/inferencedb/default-sklearn-mms-model2/")
print(df)