×
Community Blog Fluid Supports SubDataset

Fluid Supports SubDataset

This article uses AlluxioRuntime as an example to explain how to use Fluid to support SubDataset.

By Biran

What Is Fluid?

You can use the cloud-native architecture to run tasks on the cloud (such as AI and big data). You can enjoy the advantages of elastic computing resources. However, you encounter the challenges of data access latency and high bandwidth overhead for remote data pull caused by the separation of computing and storage. The iterative remote reading of many training data will slow down the GPU computing efficiency, especially in GPU deep learning training scenarios.

On the other hand, Kubernetes only provides heterogeneous storage service access and management standard interfaces (Container Storage Interface (CSI)). It does not define how applications use and manage data in container clusters. When running training tasks, data scientists need to be able to manage dataset versions, control access permissions, preprocess datasets, accelerate heterogeneous data reads, and more. However, there is no such standard solution in Kubernetes, which is one of the important capabilities missing from the cloud-native container community.

Fluid abstracts the process of using data for computing tasks and proposes the concept of elastic Dataset, which is implemented in Kubernetes as a first class citizen. Fluid creates a data orchestration and acceleration system around the elastic dataset to implement capabilities (such as Dataset management (CRUD operations), permission control, and access acceleration).

1

Fluid has two core concepts: Dataset and Runtime.

  • Dataset is a collection of logically related data used by computing engines (such as Spark, TensorFlow, and PyTorch).
  • Runtime refers to a system that provides distributed cache. Currently, Fluid supports the following types of runtimes: JuiceFS, Alluxio, JindoFS, and GooseFS. Alluxio and JindoFS are typical distributed cache engines. JuiceFS is a distributed file system that provides distributed cache capabilities. These cache systems use the storage resources (such as memory and disk) on the nodes in the Kubernetes cluster as the compute-side cache of the remote storage system.

Why Does Fluid Need to Support SubDataset?

By default, the earliest mode of Fluid supports a Dataset exclusive to one runtime, which can be understood as a dataset with dedicated cache cluster acceleration. It can be customized and optimized for the characteristics of the dataset (such as single file size, file quantity scale, and the number of clients). A separate caching system is provided. It provides the best performance and stability and does not interfere with each other. However, it is a waste of hardware resources, which requires the deployment of cache systems for different datasets. In addition, it is complex to maintain, which requires managing multiple cache runtimes. This mode is essentially a single-tenant architecture and is suitable for scenarios with high requirements for data access throughput and latency.

With the deepening of the use of Fluid, there are different needs.

  1. Dataset cache can be accessed across namespaces.
  2. Allow users to only access one subdirectory of a dataset

Users of JuiceFS tend to use Dataset to point to the root directory of JuiceFS. Different subdirectories are assigned to different data scientist groups as different data sets, and the data sets are hoped to be invisible to each other. At the same time, it supports the tightening of permissions on SubDataset. For example, the root data set supports reading and writing, and the SubDataset can be tightened to read-only.

2

This article uses AlluxioRuntime as an example to explain how to use Fluid to support SubDataset.

Example

When user A in Kubernetes namespace spark creates a dataset spark of three versions and only wants team A to see dataset spark-3.1.3, team B can only see dataset spark-3.3.1:

.
|-- spark-3.1.3
|   |-- SparkR_3.1.3.tar.gz
|   |-- pyspark-3.1.3.tar.gz
|   |-- spark-3.1.3-bin-hadoop2.7.tgz
|   |-- spark-3.1.3-bin-hadoop3.2.tgz
|   |-- spark-3.1.3-bin-without-hadoop.tgz
|   `-- spark-3.1.3.tgz
|-- spark-3.2.3
|   |-- SparkR_3.2.3.tar.gz
|   |-- pyspark-3.2.3.tar.gz
|   |-- spark-3.2.3-bin-hadoop2.7.tgz
|   |-- spark-3.2.3-bin-hadoop3.2-scala2.13.tgz
|   |-- spark-3.2.3-bin-hadoop3.2.tgz
|   |-- spark-3.2.3-bin-without-hadoop.tgz
|   `-- spark-3.2.3.tgz
`-- spark-3.3.1
    |-- SparkR_3.3.1.tar.gz
    |-- pyspark-3.3.1.tar.gz
    |-- spark-3.3.1-bin-hadoop2.tgz
    |-- spark-3.3.1-bin-hadoop3-scala2.13.tgz
    |-- spark-3.3.1-bin-hadoop3.tgz
    |-- spark-3.3.1-bin-without-hadoop.tgz
    `-- spark-3.3.1.tgz

This enables different teams to access different subDatasets, and their data is not visible to each other.

Phase 1: The Administrator Creates a Dataset to Point to the Root Directory

1.  Before you run the sample code, refer to the installation document and check whether the components of Fluid are running properly.

$ kubectl get po -n fluid-system
NAME                                         READY   STATUS      RESTARTS   AGE
alluxioruntime-controller-5c6d5f44b4-p69mt   1/1     Running     0          149m
csi-nodeplugin-fluid-bx5d2                   2/2     Running     0          149m
csi-nodeplugin-fluid-n6vbv                   2/2     Running     0          149m
csi-nodeplugin-fluid-t5p8c                   2/2     Running     0          148m
dataset-controller-797868bd7f-g4slx          1/1     Running     0          149m
fluid-webhook-6f6b7dfd74-hmsgw               1/1     Running     0          149m
fluidapp-controller-79c7b89c9-rtdg7          1/1     Running     0          149m
thinruntime-controller-5fdbfd54d4-l9jcz      1/1     Running     0          149m

2.  Create a namespace spark:

$ kubectl create ns spark

3.  Create a Dataset and AlluxioRuntime in the namespace development, where the dataset is read-write:

$ cat<<EOF >dataset.yaml
apiVersion: data.fluid.io/v1alpha1
kind: Dataset
metadata:
  name: spark
  namespace: spark
spec:
  mounts:
    - mountPoint: https://mirrors.bit.edu.cn/apache/spark/
      name: spark
      path: "/"
  accessModes:
    - ReadWriteMany
---
apiVersion: data.fluid.io/v1alpha1
kind: AlluxioRuntime
metadata:
  name: spark
  namespace: spark
spec:
  replicas: 1
  tieredstore:
    levels:
      - mediumtype: MEM
        path: /dev/shm
        quota: 4Gi
        high: "0.95"
        low: "0.7"
EOF
$ kubectl create -f dataset.yaml

4.  View the status of a dataset:

$ kubectl get dataset -n spark
NAME    UFS TOTAL SIZE   CACHED   CACHE CAPACITY   CACHED PERCENTAGE   PHASE   AGE
spark   3.41GiB          0.00B    4.00GiB          0.0%                Bound   61s

5.  Create Pod in the namespace spark to access datasets:

$ cat<<EOF >app.yaml
apiVersion: v1
kind: Pod
metadata:
  name: nginx
  namespace: spark
spec:
  containers:
    - name: nginx
      image: nginx
      volumeMounts:
        - mountPath: /data
          name: spark
  volumes:
    - name: spark
      persistentVolumeClaim:
        claimName: spark
EOF
$ kubectl create -f app.yaml

6.  View the data that the application can access through the dataset. You can see three folders: spark-3.1.3, spark-3.2.3, and spark-3.3.1. You can also see the mount permission is RW (ReadWriteMany).

$ kubectl exec -it -n spark nginx -- bash
mount | grep /data
alluxio-fuse on /data type fuse.alluxio-fuse (rw,nosuid,nodev,relatime,user_id=0,group_id=0,allow_other,max_read=131072)
root@nginx:/# ls -ltr /data/
total 2
dr--r----- 1 root root 6 Dec 16 07:25 spark-3.1.3
dr--r----- 1 root root 7 Dec 16 07:25 spark-3.2.3
dr--r----- 1 root root 7 Dec 16 07:25 spark-3.3.1
root@nginx:/# ls -ltr /data/spark-3.1.3/
total 842999
-r--r----- 1 root root  25479200 Feb  6  2022 spark-3.1.3.tgz
-r--r----- 1 root root 164080426 Feb  6  2022 spark-3.1.3-bin-without-hadoop.tgz
-r--r----- 1 root root 231842529 Feb  6  2022 spark-3.1.3-bin-hadoop3.2.tgz
-r--r----- 1 root root 227452039 Feb  6  2022 spark-3.1.3-bin-hadoop2.7.tgz
-r--r----- 1 root root 214027643 Feb  6  2022 pyspark-3.1.3.tar.gz
-r--r----- 1 root root    347324 Feb  6  2022 SparkR_3.1.3.tar.gz
root@nginx:/# ls -ltr /data/spark-3.2.3/
total 1368354
-r--r----- 1 root root  28375439 Nov 14 18:47 spark-3.2.3.tgz
-r--r----- 1 root root 209599610 Nov 14 18:47 spark-3.2.3-bin-without-hadoop.tgz
-r--r----- 1 root root 301136158 Nov 14 18:47 spark-3.2.3-bin-hadoop3.2.tgz
-r--r----- 1 root root 307366638 Nov 14 18:47 spark-3.2.3-bin-hadoop3.2-scala2.13.tgz
-r--r----- 1 root root 272866820 Nov 14 18:47 spark-3.2.3-bin-hadoop2.7.tgz
-r--r----- 1 root root 281497207 Nov 14 18:47 pyspark-3.2.3.tar.gz
-r--r----- 1 root root    349762 Nov 14 18:47 SparkR_3.2.3.tar.gz

Phase 2: The Administrator Creates a SubDataset That Points to the Directory Spark 3.1.3

1.  Create a namespace spark-313:

$ kubectl create ns spark-313

2.  The administrator creates the following information in the spark-313 namespace:

Refer to the spark dataset. The mountPoint format is dataset://${namespace of the initial dataset}/${name of the initial dataset} /subdirectory. In this example, the value is dataset://spark/spark/ spark-3.1.3.

Note: The currently referenced dataset only supports one mount, and the format must be dataset://. (This means the dataset fails to be created when dataset:// or other formats occur.) The dataset permission is read/write.

$ cat<<EOF >spark-313.yaml
apiVersion: data.fluid.io/v1alpha1
kind: Dataset
metadata:
  name: spark
  namespace: spark-313
spec:
  mounts:
    - mountPoint: dataset://spark/spark/spark-3.1.3
  accessModes:
    - ReadWriteMany
EOF
$ kubectl create -f spark-313.yaml

3.  View the dataset status:

$ kubectl get dataset -n spark-313
NAME    UFS TOTAL SIZE   CACHED   CACHE CAPACITY   CACHED PERCENTAGE   PHASE   AGE
spark   3.41GiB          0.00B    4.00GiB          0.0%                Bound   108s

4.  The user creates a Pod in the spark-313 namespace.

$ cat<<EOFba >app-spark313.yaml
apiVersion: v1
kind: Pod
metadata:
  name: nginx
  namespace: spark-313
spec:
  containers:
    - name: nginx
      image: nginx
      volumeMounts:
        - mountPath: /data
          name: spark
  volumes:
    - name: spark
      persistentVolumeClaim:
        claimName: spark
EOF
$ kubectl create -f app-spark313.yaml

5.  If you access data in the spark-313 namespace, you can only see the contents of the spark-3.1.3 folder and see the pvc permission RWX.

$ kubectl get pvc -n spark-313
NAME    STATUS   VOLUME            CAPACITY   ACCESS MODES   STORAGECLASS   AGE
spark   Bound    spark-313-spark   100Gi      RWX            fluid          21h
$ kubectl exec -it -n spark-313 nginx -- bash
root@nginx:/# mount | grep /data
alluxio-fuse on /data type fuse.alluxio-fuse (rw,nosuid,nodev,relatime,user_id=0,group_id=0,allow_other,max_read=131072)
root@nginx:/# ls -ltr /data/
total 842999
-r--r----- 1 root root  25479200 Feb  6  2022 spark-3.1.3.tgz
-r--r----- 1 root root 164080426 Feb  6  2022 spark-3.1.3-bin-without-hadoop.tgz
-r--r----- 1 root root 231842529 Feb  6  2022 spark-3.1.3-bin-hadoop3.2.tgz
-r--r----- 1 root root 227452039 Feb  6  2022 spark-3.1.3-bin-hadoop2.7.tgz
-r--r----- 1 root root 214027643 Feb  6  2022 pyspark-3.1.3.tar.gz
-r--r----- 1 root root    347324 Feb  6  2022 SparkR_3.1.3.tar.gz

Phase 3: The Administrator Creates a SubDataset That Points to the Directory Spark 3.3.1

1.  Create a namespace spark-331:

$ kubectl create ns spark-331

2.  The administrator creates the following information in the spark-331 namespace:

Refer to the spark dataset. The mountPoint format is dataset://${namespace of the initial dataset}/${name of the initial dataset} /subdirectory. In this example, dataset://spark/spark/ spark-3.3.1.

Note: The currently referenced dataset supports only one mount, and the form must be dataset://. (This means the dataset creation fails when dataset:// and other forms occur.) The read and write permissions are specified as ReadOnlyMany.

$ cat<<EOF >spark-331.yaml
apiVersion: data.fluid.io/v1alpha1
kind: Dataset
metadata:
  name: spark
  namespace: spark-331
spec:
  mounts:
    - mountPoint: dataset://spark/spark/spark-3.3.1
  accessModes:
    - ReadOnlyMany
EOF
$ kubectl create -f spark-331.yaml

3.  The user creates a Pod in the spark-331 namespace.

$ cat<<EOF >app-spark331.yaml
apiVersion: v1
kind: Pod
metadata:
  name: nginx
  namespace: spark-331
spec:
  containers:
    - name: nginx
      image: nginx
      volumeMounts:
        - mountPath: /data
          name: spark
  volumes:
    - name: spark
      persistentVolumeClaim:
        claimName: spark
EOF
$ kubectl create -f app-spark331.yaml

4.  If you access data in the spark-331 namespace, you can only view the contents of the spark-3.3.1 folder and view the PVC permission ROX (ReadOnlyMany).

$ kubectl get pvc -n spark-331
NAME    STATUS   VOLUME            CAPACITY   ACCESS MODES   STORAGECLASS   AGE
spark   Bound    spark-331-spark   100Gi      ROX            fluid          14m$ kubectl exec -it -n spark-331 nginx -- bash
mount | grep /data
alluxio-fuse on /data type fuse.alluxio-fuse (rw,nosuid,nodev,relatime,user_id=0,group_id=0,allow_other,max_read=131072)
root@nginx:/# ls -ltr /data/
total 842999
-r--r----- 1 root root  25479200 Feb  6  2022 spark-3.1.3.tgz
-r--r----- 1 root root 164080426 Feb  6  2022 spark-3.1.3-bin-without-hadoop.tgz
-r--r----- 1 root root 231842529 Feb  6  2022 spark-3.1.3-bin-hadoop3.2.tgz
-r--r----- 1 root root 227452039 Feb  6  2022 spark-3.1.3-bin-hadoop2.7.tgz
-r--r----- 1 root root 214027643 Feb  6  2022 pyspark-3.1.3.tar.gz
-r--r----- 1 root root    347324 Feb  6  2022 SparkR_3.1.3.tar.gz

Summary

In this example, we demonstrated the ability of a SubDataset to use a subdirectory of a Dataset as a Dataset. The dataset can be used by different data scientists according to different segmentation strategies.

0 1 0
Share on

Alibaba Container Service

115 posts | 26 followers

You may also like

Comments

Alibaba Container Service

115 posts | 26 followers

Related Products