跳到主要内容
版本:Latest

How to Run in Knative Environment

This example uses the open source framework Knative as an example to demonstrate how to perform unified data acceleration via Fluid in a Serverless environment. This example uses AlluxioRuntime as an example, and in fact Fluid supports all Runtime running in a Serverless environment.

Installation

1.Install Knative Serving v1.2 according to the Knative documentation, you need to enable the kubernetes.Deploymentspec-persistent-volume-claim option.

Check if Knative's components are working properly

kubectl get Deployments -n knative-serving

Note: This document is just for demonstration purpose, please refer to the best practices of Knative documentation for Knative deployment in production environment. Also, since the container images of Knative are in the gcr.io image repository, please make sure the images are reachable. If you are using AliCloud, you can also use AliCloud ACK hosting service directly to reduce the complexity of configuring Knative.

2.Please refer to the installation documentation to install the latest Fluid, and check that the Fluid components are working properly after installation (this document uses AlluxioRuntime as an example):

$ kubectl get deploy -n fluid-system
NAME READY UP-TO-DATE AVAILABLE AGE
alluxioruntime-controller 1/1 1 1 18m
dataset-controller 1/1 1 1 18m
fluid-webhook 1/1 1 1 18m

Typically, you can see a Deployment named dataset-controller, a Deployment named alluxioruntime-controller, and a Deployment named fluid-webhook.

Configuration

Running

Create dataset and runtime

Create Runtime resources for different types of Runtime, as well as a Dataset with the same name. Here is the example of AlluxioRuntime, the following is the Dataset content:

$ cat<<EOF >dataset.yaml
apiVersion: data.fluid.io/v1alpha1
kind: Dataset
metadata:
name: serverless-data
spec:
mounts:
- mountPoint: https://mirrors.bit.edu.cn/apache/hbase/stable/
name: hbase
path: "/"
accessModes:
- ReadOnlyMany
---
apiVersion: data.fluid.io/v1alpha1
kind: AlluxioRuntime
metadata:
name: serverless-data
spec:
replicas: 2
tieredstore:
levels:
- mediumtype: MEM
path: /dev/shm
quota: 2Gi
high: "0.95"
low: "0.7"
EOF

Execute the Create Dataset operation:

$ kubectl create -f dataset.yaml

Check Dataset Status:

$ kubectl get alluxio
NAME MASTER PHASE WORKER PHASE FUSE PHASE AGE
serverless-data Ready Ready Ready 4m52s
$ kubectl get dataset
NAME UFS TOTAL SIZE CACHED CACHE CAPACITY CACHED PERCENTAGE PHASE AGE
serverless-data 566.22MiB 0.00B 4.00GiB 0.0% Bound 4m52s

Creating Knative Serving Resource Objects

$ cat<<EOF >serving.yaml
apiVersion: serving.knative.dev/v1
kind: Service
metadata:
name: model-serving
spec:
template:
metadata:
labels:
app: model-serving
serverless.fluid.io/inject: "true"
annotations:
autoscaling.knative.dev/target: "10"
autoscaling.knative.dev/scaleDownDelay: "30m"
autoscaling.knative.dev/minScale: "1"
spec:
containers:
- image: fluidcloudnative/serving
ports:
- name: http1
containerPort: 8080
env:
- name: TARGET
value: "World"
volumeMounts:
- mountPath: /data
name: data
readOnly: true
volumes:
- name: data
persistentVolumeClaim:
claimName: serverless-data
readOnly: true
EOF
$ kubectl create -f serving.yaml
service.serving.knative.dev/model-serving created

Please configure serverless.fluid.io/inject: "true" in the label of the podSpec or podTemplateSpec.

Check if Knative Serving is created and check if fuse-container is injected

$ kubectl get po
NAME READY STATUS RESTARTS AGE
model-serving-00001-deployment-64d674d75f-46vvf 3/3 Running 0 76s
serverless-data-master-0 2/2 Running 0 16m
serverless-data-worker-0 2/2 Running 0 16m
serverless-data-worker-1 2/2 Running 0 16m
$ kubectl get po model-serving-00001-deployment-64d674d75f-46vvf -oyaml| grep -i fluid-fuse -B 3
- /opt/alluxio/integration/fuse/bin/alluxio-fuse
- unmount
- /runtime-mnt/alluxio/default/serverless-data/alluxio-fuse
name: fluid-fuse

Checking the Knative Serving startup speed, you can see that the startup data loading time is 92s.

$ kubectl logs model-serving-00001-deployment-64d674d75f-46vvf -c user-container
Begin loading models at 16:29:02

real 1m32.639s
user 0m0.001s
sys 0m1.305s
Finish loading models at 16:29:45
2022-02-15 16:29:45 INFO Hello world sample started.

Clean up Knative serving instances

$ kubectl delete -f serving.yaml

Execute data warm-up

Create the dataload object and check its status:

$ cat<<EOF >dataload.yaml
apiVersion: data.fluid.io/v1alpha1
kind: DataLoad
metadata:
name: serverless-dataload
spec:
dataset:
name: serverless-data
namespace: default
EOF
$ kubectl create -f dataload.yaml
dataload.data.fluid.io/serverless-dataload created
$ kubectl get dataload
NAME DATASET PHASE AGE DURATION
serverless-dataload serverless-data Complete 2m43s 34s

Check the cache status at this point, the data is now fully cached in the cluster.

$ kubectl get dataset
NAME UFS TOTAL SIZE CACHED CACHE CAPACITY CACHED PERCENTAGE PHASE AGE
serverless-data 566.22MiB 566.22MiB 4.00GiB 100.0% Bound 33m

Create Knative service again:

$ kubectl create -f serving.yaml
service.serving.knative.dev/model-serving created

Checking the boot time at this point reveals that the current boot time for loading data is 3.66s, which becomes 1/20 of the performance without warm-up.

kubectl logs model-serving-00001-deployment-6cb54f94d7-dbgxf -c user-container
Begin loading models at 18:38:23

real 0m3.666s
user 0m0.000s
sys 0m1.367s
Finish loading models at 18:38:25
2022-02-15 18:38:25 INFO Hello world sample started.

Note: This example uses Knative serving. If you don't have a Knative environment, you can also experiment with Deployment.

apiVersion: apps/v1
kind: Deployment
metadata:
name: model-serving
spec:
selector:
matchLabels:
app: model-serving
template:
metadata:
labels:
app: model-serving
serverless.fluid.io/inject: "true"
spec:
containers:
- image: fluidcloudnative/serving
name: serving
ports:
- name: http1
containerPort: 8080
env:
- name: TARGET
value: "World"
volumeMounts:
- mountPath: /data
name: data
volumes:
- name: data
persistentVolumeClaim:
claimName: serverless-data

Note: The default sidecar injection mode does not enable cached directory short-circuit reads, if you need to enable this capability, you can configure the parameter cachedir.sidecar.fluid.io/inject to true in the labels.

apiVersion: apps/v1
kind: Deployment
metadata:
name: model-serving
spec:
selector:
matchLabels:
app: model-serving
template:
metadata:
labels:
app: model-serving
serverless.fluid.io/inject: "true"
cachedir.sidecar.fluid.io/inject: "true"
spec:
containers:
- image: fluidcloudnative/serving
name: serving
ports:
- name: http1
containerPort: 8080
env:
- name: TARGET
value: "World"
volumeMounts:
- mountPath: /data
name: data
volumes:
- name: data
persistentVolumeClaim:
claimName: serverless-data