Pipelines on k8s (#142)

* fix dataflow bugs * lint * small updates for docker compose tests * update install doc at top of notebook * Typo fix * k8s pipeline * lint Co-authored-by: Andrei Paleyes <[email protected]>
SeldonIO · Apr 8, 2022 · c89ae5a · c89ae5a
1 parent 3bf3e96
commit c89ae5a
Show file tree

Hide file tree

Showing 18 changed files with 532 additions and 69 deletions.
diff --git a/Makefile b/Makefile
@@ -19,6 +19,7 @@ undeploy-local-triton:
 .PHONY: deploy-k8s
 deploy-k8s:
 	kubectl create -f k8s/seldon-v2-crds.yaml
+	sleep 1
 	kubectl wait --for condition=established --timeout=60s crd/servers.mlops.seldon.io
 	kubectl wait --for condition=established --timeout=60s crd/serverconfigs.mlops.seldon.io
 	kubectl create -f k8s/seldon-v2-components.yaml

diff --git a/k8s/seldon-v2-components.yaml b/k8s/seldon-v2-components.yaml
@@ -413,7 +413,7 @@ subjects:
 ---
 apiVersion: v1
 data:
-  agent.yaml: "rclone: \n  config_secrets: [\"seldon-rclone-gs-public\"]"
+  agent.yaml: "rclone: \n  config_secrets: [\"seldon-rclone-gs-public\"]\nkafka:\n  active: true\n  broker: \"seldon-kafka-plain-bootstrap.kafka:9092\""
 kind: ConfigMap
 metadata:
   name: seldon-agent
@@ -490,6 +490,27 @@ spec:
 ---
 apiVersion: v1
 kind: Service
+metadata:
+  labels:
+    app: seldon-mesh
+  name: seldon-pipelinegateway
+  namespace: seldon-mesh
+spec:
+  clusterIP: None
+  ports:
+  - name: http
+    port: 9010
+    protocol: TCP
+    targetPort: http
+  - name: grpc
+    port: 9011
+    protocol: TCP
+    targetPort: grpc
+  selector:
+    app: pipelinegateway
+---
+apiVersion: v1
+kind: Service
 metadata:
   labels:
     app: seldon-scheduler
@@ -509,6 +530,10 @@ spec:
     port: 9005
     protocol: TCP
     targetPort: agent
+  - name: dataflow
+    port: 9008
+    protocol: TCP
+    targetPort: dataflow
   selector:
     control-plane: seldon-scheduler
   type: LoadBalancer
@@ -592,6 +617,47 @@ spec:
 ---
 apiVersion: apps/v1
 kind: Deployment
+metadata:
+  labels:
+    control-plane: seldon-dataflow-engine
+  name: seldon-dataflow-engine
+  namespace: seldon-mesh
+spec:
+  replicas: 1
+  selector:
+    matchLabels:
+      control-plane: seldon-dataflow-engine
+  template:
+    metadata:
+      labels:
+        control-plane: seldon-dataflow-engine
+    spec:
+      containers:
+      - env:
+        - name: SELDON_KAFKA_BOOTSTRAP_SERVERS
+          value: seldon-kafka-plain-bootstrap.kafka:9092
+        - name: SELDON_UPSTREAM_HOST
+          value: seldon-scheduler
+        - name: SELDON_UPSTREAM_PORT
+          value: "9008"
+        - name: SELDON_CORES_COUNT
+          value: "4"
+        image: seldonio/seldon-dataflow-engine:latest
+        imagePullPolicy: IfNotPresent
+        name: dataflow-engine
+        resources:
+          limits:
+            cpu: 4
+            memory: 2G
+          requests:
+            cpu: 1
+            memory: 1G
+      securityContext:
+        runAsUser: 8888
+      terminationGracePeriodSeconds: 5
+---
+apiVersion: apps/v1
+kind: Deployment
 metadata:
   labels:
     app: seldon-envoy
@@ -625,72 +691,83 @@ apiVersion: apps/v1
 kind: Deployment
 metadata:
   labels:
-    control-plane: seldon-scheduler
-  name: seldon-scheduler
+    control-plane: seldon-modelgateway
+  name: seldon-modelgateway
   namespace: seldon-mesh
 spec:
   replicas: 1
   selector:
     matchLabels:
-      control-plane: seldon-scheduler
+      control-plane: seldon-modelgateway
   template:
     metadata:
       labels:
-        control-plane: seldon-scheduler
+        control-plane: seldon-modelgateway
     spec:
       containers:
-      - command:
-        - /bin/scheduler
-        image: seldonio/seldon-scheduler:latest
+      - args:
+        - --scheduler-host=seldon-scheduler
+        - --scheduler-port=9004
+        - --envoy-host=seldon-mesh
+        - --envoy-port=80
+        command:
+        - /bin/modelgateway
+        image: seldonio/seldon-modelgateway:latest
         imagePullPolicy: IfNotPresent
-        name: scheduler
-        ports:
-        - containerPort: 9002
-          name: xds
-        - containerPort: 9004
-          name: scheduler
-        - containerPort: 9005
-          name: agent
+        name: modelgateway
         resources:
           limits:
             cpu: 500m
             memory: 300Mi
           requests:
             cpu: 100m
             memory: 200Mi
+        volumeMounts:
+        - mountPath: /mnt/config
+          name: config-volume
       securityContext:
         runAsUser: 8888
       serviceAccountName: seldon-scheduler
       terminationGracePeriodSeconds: 5
+      volumes:
+      - configMap:
+          name: seldon-agent
+        name: config-volume
 ---
 apiVersion: apps/v1
 kind: Deployment
 metadata:
   labels:
-    control-plane: seldon-stream
-  name: seldon-stream
+    app: pipelinegateway
+  name: seldon-pipelinegateway
   namespace: seldon-mesh
 spec:
   replicas: 1
   selector:
     matchLabels:
-      control-plane: seldon-stream
+      app: pipelinegateway
   template:
     metadata:
       labels:
-        control-plane: seldon-stream
+        app: pipelinegateway
     spec:
       containers:
       - args:
-        - --scheduler-host=seldon-scheduler
-        - --scheduler-port=9004
-        - --envoy-host=seldon-mesh
-        - --envoy-port=80
+        - --http-port=9010
+        - --grpc-port=9011
+        - --config-path=/mnt/config
         command:
-        - /bin/stream
-        image: seldonio/seldon-stream:latest
+        - /bin/pipelinegateway
+        image: seldonio/seldon-pipelinegateway:latest
         imagePullPolicy: IfNotPresent
-        name: stream
+        name: pipelinegateway
+        ports:
+        - containerPort: 9010
+          name: http
+          protocol: TCP
+        - containerPort: 9011
+          name: grpc
+          protocol: TCP
         resources:
           limits:
             cpu: 500m
@@ -710,6 +787,52 @@ spec:
           name: seldon-agent
         name: config-volume
 ---
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  labels:
+    control-plane: seldon-scheduler
+  name: seldon-scheduler
+  namespace: seldon-mesh
+spec:
+  replicas: 1
+  selector:
+    matchLabels:
+      control-plane: seldon-scheduler
+  template:
+    metadata:
+      labels:
+        control-plane: seldon-scheduler
+    spec:
+      containers:
+      - args:
+        - --pipeline-gateway-host=seldon-pipelinegateway
+        command:
+        - /bin/scheduler
+        image: seldonio/seldon-scheduler:latest
+        imagePullPolicy: IfNotPresent
+        name: scheduler
+        ports:
+        - containerPort: 9002
+          name: xds
+        - containerPort: 9004
+          name: scheduler
+        - containerPort: 9005
+          name: agent
+        - containerPort: 9008
+          name: dataflow
+        resources:
+          limits:
+            cpu: 500m
+            memory: 300Mi
+          requests:
+            cpu: 100m
+            memory: 200Mi
+      securityContext:
+        runAsUser: 8888
+      serviceAccountName: seldon-scheduler
+      terminationGracePeriodSeconds: 5
+---
 apiVersion: mlops.seldon.io/v1alpha1
 kind: ServerConfig
 metadata:
@@ -839,6 +962,8 @@ spec:
         name: mlserver-models
         readOnly: true
     securityContext:
+      fsGroup: 2000
+      runAsNonRoot: true
       runAsUser: 1000
     serviceAccountName: seldon-server
     terminationGracePeriodSeconds: 30
@@ -983,6 +1108,8 @@ spec:
         name: triton-models
         readOnly: true
     securityContext:
+      fsGroup: 2000
+      runAsNonRoot: true
       runAsUser: 1000
     serviceAccountName: seldon-server
     terminationGracePeriodSeconds: 30

diff --git a/k8s/seldon-v2-crds.yaml b/k8s/seldon-v2-crds.yaml
@@ -459,6 +459,11 @@ spec:
                     passEmptyResponses:
                       description: 'Whether empty tensors output should be passed to onwards Default: false'
                       type: boolean
+                    tensorMap:
+                      additionalProperties:
+                        type: string
+                      description: Map of tensor name conversions to use e.g. output1 -> input1
+                      type: object
                   required:
                   - name
                   type: object

diff --git a/operator/controllers/mlops/experiment_controller.go b/operator/controllers/mlops/experiment_controller.go
@@ -19,6 +19,8 @@ package mlops
 import (
 	"context"
 
+	"sigs.k8s.io/controller-runtime/pkg/predicate"
+
 	"github.com/seldonio/seldon-core/operatorv2/pkg/constants"
 	"github.com/seldonio/seldon-core/operatorv2/pkg/utils"
 	"k8s.io/apimachinery/pkg/api/errors"
@@ -110,7 +112,9 @@ func (r *ExperimentReconciler) Reconcile(ctx context.Context, req ctrl.Request)
 
 // SetupWithManager sets up the controller with the Manager.
 func (r *ExperimentReconciler) SetupWithManager(mgr ctrl.Manager) error {
+	pred := predicate.GenerationChangedPredicate{}
 	return ctrl.NewControllerManagedBy(mgr).
 		For(&mlopsv1alpha1.Experiment{}).
+		WithEventFilter(pred).
 		Complete(r)
 }
diff --git a/operator/controllers/mlops/pipeline_controller.go b/operator/controllers/mlops/pipeline_controller.go
@@ -19,6 +19,8 @@ package mlops
 import (
 	"context"
 
+	"sigs.k8s.io/controller-runtime/pkg/predicate"
+
 	"github.com/seldonio/seldon-core/operatorv2/pkg/constants"
 	"github.com/seldonio/seldon-core/operatorv2/pkg/utils"
 	"k8s.io/apimachinery/pkg/api/errors"
@@ -111,7 +113,9 @@ func (r *PipelineReconciler) Reconcile(ctx context.Context, req ctrl.Request) (c
 
 // SetupWithManager sets up the controller with the Manager.
 func (r *PipelineReconciler) SetupWithManager(mgr ctrl.Manager) error {
+	pred := predicate.GenerationChangedPredicate{}
 	return ctrl.NewControllerManagedBy(mgr).
 		For(&mlopsv1alpha1.Pipeline{}).
+		WithEventFilter(pred).
 		Complete(r)
 }