import Tabs from '@theme/Tabs'; import TabItem from '@theme/TabItem';

JuiceFS provides Hadoop-compatible File System by Hadoop Java SDK. Various applications in the Hadoop ecosystem can smoothly use JuiceFS to store data without changing the code.

Requirements

1. Hadoop and related components

JuiceFS Hadoop Java SDK is compatible with Hadoop 2.x and Hadoop 3.x. As well as variety of components in Hadoop ecosystem.

2. User permissions

JuiceFS uses local "User/UID" and "Group/GID" mappings by default, and when used in a distributed environment, to avoid permission issues, please refer to documentation synchronizes the "User/UID" and "Group/GID" that needs to be used to all Hadoop nodes. It is also possible to define a global user and group file to make all nodes in the cluster share the permission configuration. Please see here for related configurations.

3. File system

You should first create at least one JuiceFS file system to provide storage for components related to the Hadoop ecosystem through the JuiceFS Java SDK. When deploying the Java SDK, specify the metadata engine address of the created file system in the configuration file.

To create a file system, please refer to our quick start.

:::note If you want to use JuiceFS in a distributed environment, when creating a file system, please plan the object storage and database to be used reasonably to ensure that they can be accessed by each node in the cluster. :::

4. Memory

Depending on the read and write load of computing tasks (such as Spark executor), JuiceFS Hadoop Java SDK may require an additional 4 * juicefs.memory-size off-heap memory to speed up read and write performance. By default, it is recommended to configure at least 1.2GB of off-heap memory for compute tasks.

5. Java runtime version

JuiceFS Hadoop Java SDK is compiled with JDK 8 by default. If it needs to be used in a higher version of Java runtime (such as Java 17), the following options need to be added to the JVM parameters to allow the use of reflection API:

--add-exports=java.base/sun.nio.ch=ALL-UNNAMED

For more information on the above option, please refer to official documentation.

Install and compile the client

Install the pre-compiled client

Please refer to the "Installation" document to learn how to download the precompiled JuiceFS Hadoop Java SDK.

Compile the client manually

:::note No matter which system environment the client is compiled for, the compiled JAR file has the same name and can only be deployed in the matching system environment. For example, when compiled in Linux, it can only be used in the Linux environment. In addition, since the compiled package depends on glibc, it is recommended to compile with a lower version system to ensure better compatibility. :::

Compilation depends on the following tools:

• Go 1.20+
• JDK 8+
• Maven 3.3+
• Git
• make
• GCC 5.4+

Linux and macOS

Clone the repository:

git clone https://github.com/juicedata/juicefs.git

Enter the directory and compile:

cd juicefs/sdk/java
make

:::note If Ceph RADOS is used to store data, you need to install librados-dev first and [build libjfs.so]`. :::

cd juicefs/sdk/java
make ceph

After the compilation, you can find the compiled JAR file in the sdk/java/target directory, including two versions:

• Contains third-party dependent packages: juicefs-hadoop-X.Y.Z.jar
• Does not include third-party dependent packages: original-juicefs-hadoop-X.Y.Z.jar

It is recommended to use a version that includes third-party dependencies.

Windows

The client used in the Windows environment needs to be obtained through cross-compilation on Linux or macOS. The compilation depends on mingw-w64, which needs to be installed first.

The steps are the same as compiling on Linux or macOS. For example, on the Ubuntu system, install the mingw-w64 package first to solve the dependency problem:

sudo apt install mingw-w64

Clone and enter the JuiceFS source code directory, execute the following code to compile:

cd juicefs/sdk/java

make win

Deploy the client

To enable each component of the Hadoop ecosystem to correctly identify JuiceFS, the following configurations are required:

1. Place the compiled JAR file and $JAVA_HOME/lib/tools.jar into the classpath of the component. The installation paths of common big data platforms and components are shown in the table below.
2. Put JuiceFS configurations into the configuration file of each Hadoop ecosystem component (usually core-site.xml), see Client Configurations for details.

It is recommended to place the JAR file in a fixed location, and the other locations are called it through symbolic links.

Big Data Platforms

Community Components

Client Configurations

Please refer to the following table to set the relevant parameters of the JuiceFS file system and write it into the configuration file, which is generally core-site.xml.

Core Configurations

Cache Configurations

I/O Configurations

Other Configurations

Multiple file systems configuration

When multiple JuiceFS file systems need to be used at the same time, all the above configuration items can be specified for a specific file system. You only need to put the file system name in the middle of the configuration item, such as jfs1 and jfs2 in the following example:

<property>
  <name>juicefs.jfs1.meta</name>
  <value>redis://jfs1.host:port/1</value>
</property>
<property>
  <name>juicefs.jfs2.meta</name>
  <value>redis://jfs2.host:port/1</value>
</property>

Configuration Example

The following is a commonly used configuration example. Please replace the {HOST}, {PORT} and {DB} variables in the juicefs.meta configuration with actual values.

<property>
  <name>fs.jfs.impl</name>
  <value>io.juicefs.JuiceFileSystem</value>
</property>
<property>
  <name>fs.AbstractFileSystem.jfs.impl</name>
  <value>io.juicefs.JuiceFS</value>
</property>
<property>
  <name>juicefs.meta</name>
  <value>redis://{HOST}:{PORT}/{DB}</value>
</property>
<property>
  <name>juicefs.cache-dir</name>
  <value>/data*/jfs</value>
</property>
<property>
  <name>juicefs.cache-size</name>
  <value>1024</value>
</property>
<property>
  <name>juicefs.access-log</name>
  <value>/tmp/juicefs.access.log</value>
</property>

Configuration in Hadoop

Please refer to the aforementioned configuration tables and add configuration parameters to the Hadoop configuration file core-site.xml.

CDH6

If you are using CDH 6, in addition to modifying core-site, you also need to modify mapreduce.application.classpath through the YARN service interface, adding:

$HADOOP_COMMON_HOME/lib/juicefs-hadoop.jar

HDP

In addition to modifying core-site, you also need to modify the configuration mapreduce.application.classpath through the MapReduce2 service interface and add it at the end (variables do not need to be replaced):

/usr/hdp/${hdp.version}/hadoop/lib/juicefs-hadoop.jar

Flink

Add configuration parameters to conf/flink-conf.yaml. If you only use JuiceFS in Flink, you don't need to configure JuiceFS in the Hadoop environment, you only need to configure the Flink client.

Hudi

:::note Hudi supports JuiceFS since v0.10.0, please make sure you are using the correct version. :::

Please refer to "Hudi Official Documentation" to learn how to configure JuiceFS.

Kafka Connect

It is possible to use Kafka Connect and HDFS Sink Connector（HDFS 2 and HDFS 3）to store data on JuiceFS.

First you need to add JuiceFS SDK to classpath in Kafka Connect, e.g., /usr/share/java/confluentinc-kafka-connect-hdfs/lib.

While creating a Connect Sink task, configuration needs to be set up as follows:

• Specify hadoop.conf.dir as the directory that contains the configuration file core-site.xml. If it is not running in Hadoop environment, you can create a separate directory such as /usr/local/juicefs/hadoop, and then add the JuiceFS related configurations to core-site.xml.
• Specify store.url as a path starting with jfs://.

For example:

# Other configuration items are omitted.
hadoop.conf.dir=/path/to/hadoop-conf
store.url=jfs://path/to/store

HBase

JuiceFS can be used by HBase for HFile, but is not fast (low latency) enough for Write Ahead Log (WAL), because it take much longer time to persist data into object storage than memory of DataNode.

It is recommended to deploy a small HDFS cluster to store WAL and HFile files to be stored on JuiceFS.

Create a new HBase cluster

Modify hbase-site.xml:

<property>
  <name>hbase.rootdir</name>
  <value>jfs://{vol_name}/hbase</value>
</property>
<property>
  <name>hbase.wal.dir</name>
  <value>hdfs://{ns}/hbase-wal</value>
</property>

Modify existing HBase cluster

In addition to modifying the above configurations, since the HBase cluster has already stored some data in ZooKeeper, in order to avoid conflicts, there are two solutions:

1. Delete the old cluster

   Delete the znode (default /hbase) configured by zookeeper.znode.parent via the ZooKeeper client.

   :::note This operation will delete all data on this HBase cluster. :::

2. Use a new znode

   Keep the znode of the original HBase cluster so that it can be recovered later. Then configure a new value for zookeeper.znode.parent:

   <property>
     <name>zookeeper.znode.parent</name>
     <value>/hbase-jfs</value>
   </property>

Restart Services

When the following components need to access JuiceFS, they should be restarted.

:::note Before restart, you need to confirm JuiceFS related configuration has been written to the configuration file of each component, usually you can find them in core-site.xml on the machine where the service of the component was deployed. :::

HDFS, Hue, ZooKeeper and other services don't need to be restarted.

When Class io.juicefs.JuiceFileSystem not found or No FilesSystem for scheme: jfs exceptions was occurred after restart, reference FAQ.

Trash

JuiceFS Hadoop Java SDK also has the same trash function as HDFS, which needs to be enabled by setting fs.trash.interval and fs.trash.checkpoint.interval, please refer to HDFS documentation for more information.

Environmental Verification

After the deployment of the JuiceFS Java SDK, the following methods can be used to verify the success of the deployment.

Hadoop CLI

hadoop fs -ls jfs://{JFS_NAME}/

:::info The JFS_NAME is the volume name when you format JuiceFS file system. :::

Hive

CREATE TABLE IF NOT EXISTS person
(
  name STRING,
  age INT
) LOCATION 'jfs://{JFS_NAME}/tmp/person';

Java/Scala project

1. Add Maven or Gradle dependencies:

   <dependency>
       <groupId>org.apache.hadoop</groupId>
       <artifactId>hadoop-common</artifactId>
       <version>{HADOOP_VERSION}</version>
       <scope>provided</scope>
   </dependency>
   <dependency>
       <groupId>io.juicefs</groupId>
       <artifactId>juicefs-hadoop</artifactId>
       <version>{JUICEFS_HADOOP_VERSION}</version>
       <scope>provided</scope>
   </dependency>

   dependencies {
     implementation 'org.apache.hadoop:hadoop-common:${hadoopVersion}'
     implementation 'io.juicefs:juicefs-hadoop:${juicefsHadoopVersion}'
   }

2. Use the following sample code to verify:

package demo;

import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.fs.FileStatus;
import org.apache.hadoop.fs.FileSystem;
import org.apache.hadoop.fs.Path;

public class JuiceFSDemo {
    public static void main(String[] args) throws Exception {
        Configuration conf = new Configuration();
        conf.set("fs.jfs.impl", "io.juicefs.JuiceFileSystem");
        conf.set("juicefs.meta", "redis://127.0.0.1:6379/0");  // JuiceFS metadata engine URL
        Path p = new Path("jfs://{JFS_NAME}/");  // Please replace "{JFS_NAME}" with the correct value
        FileSystem jfs = p.getFileSystem(conf);
        FileStatus[] fileStatuses = jfs.listStatus(p);
        // Traverse JuiceFS file system and print file paths
        for (FileStatus status : fileStatuses) {
            System.out.println(status.getPath());
        }
    }
}

Monitoring metrics collection

Please see the "Monitoring" documentation to learn how to collect and display JuiceFS monitoring metrics.

Benchmark

Here are a series of methods to use the built-in stress testing tool of the JuiceFS client to test the performance of the client environment that has been successfully deployed.

1. Local Benchmark

Metadata

• create

  hadoop jar juicefs-hadoop.jar nnbench create -files 10000 -baseDir jfs://{JFS_NAME}/tmp/benchmarks/NNBench -local

  This command will create 10000 empty files

• open

  hadoop jar juicefs-hadoop.jar nnbench open -files 10000 -baseDir jfs://{JFS_NAME}/tmp/benchmarks/NNBench -local

  This command will open 10000 files without reading data

• rename

  hadoop jar juicefs-hadoop.jar nnbench rename -files 10000 -baseDir jfs://{JFS_NAME}/tmp/benchmarks/NNBench -local

• delete

  hadoop jar juicefs-hadoop.jar nnbench delete -files 10000 -baseDir jfs://{JFS_NAME}/tmp/benchmarks/NNBench -local

• For reference

  Operation	TPS	Latency (ms)
  create	644	1.55
  open	3467	0.29
  rename	483	2.07
  delete	506	1.97

I/O Performance

• sequential write

  hadoop jar juicefs-hadoop.jar dfsio -write -size 20000 -baseDir jfs://{JFS_NAME}/tmp/benchmarks/DFSIO -local

• sequential read

  hadoop jar juicefs-hadoop.jar dfsio -read -size 20000 -baseDir jfs://{JFS_NAME}/tmp/benchmarks/DFSIO -local

  When run the cmd for the second time, the result may be much better than the first run. It's because the data was cached in memory, just clean the local disk cache.

• For reference

  Operation	Throughput (MB/s)
  write	647
  read	111

If the network bandwidth of the machine is relatively low, it can generally reach the network bandwidth bottleneck.

2. Distributed Benchmark

The following command will start the MapReduce distributed task to test the metadata and IO performance. During the test, it is necessary to ensure that the cluster has sufficient resources to start the required map tasks.

Computing resources used in this test:

• Server: 4 cores and 32 GB memory, burst bandwidth 5Gbit/s x 3
• Database: Alibaba Cloud Redis 5.0 Community 4G Master-Slave Edition

Metadata

• create

  hadoop jar juicefs-hadoop.jar nnbench create -maps 10 -threads 10 -files 1000 -baseDir jfs://{JFS_NAME}/tmp/benchmarks/NNBench

  10 map task, each has 10 threads, each thread create 1000 empty file. 100000 files in total

• open

  hadoop jar juicefs-hadoop.jar nnbench open -maps 10 -threads 10 -files 1000 -baseDir jfs://{JFS_NAME}/tmp/benchmarks/NNBench

  10 map task, each has 10 threads, each thread open 1000 file. 100000 files in total

• rename

  hadoop jar juicefs-hadoop.jar nnbench rename -maps 10 -threads 10 -files 1000 -baseDir jfs://{JFS_NAME}/tmp/benchmarks/NNBench

  10 map task, each has 10 threads, each thread rename 1000 file. 100000 files in total

• delete

  hadoop jar juicefs-hadoop.jar nnbench delete -maps 10 -threads 10 -files 1000 -baseDir jfs://{JFS_NAME}/tmp/benchmarks/NNBench

  10 map task, each has 10 threads, each thread delete 1000 file. 100000 files in total

• For reference

  • 10 threads

    Operation	IOPS	Latency (ms)
    create	4178	2.2
    open	9407	0.8
    rename	3197	2.9
    delete	3060	3.0

  • 100 threads

    Operation	IOPS	Latency (ms)
    create	11773	7.9
    open	34083	2.4
    rename	8995	10.8
    delete	7191	13.6

I/O Performance

• sequential write

  hadoop jar juicefs-hadoop.jar dfsio -write -maps 10 -size 10000 -baseDir jfs://{JFS_NAME}/tmp/benchmarks/DFSIO

  10 map task, each task write 10000MB random data sequentially

• sequential read

  hadoop jar juicefs-hadoop.jar dfsio -read -maps 10 -size 10000 -baseDir jfs://{JFS_NAME}/tmp/benchmarks/DFSIO

  10 map task, each task read 10000MB random data sequentially

• For reference

  Operation	Average throughput (MB/s)	Total Throughput (MB/s)
  write	198	1835
  read	124	1234

3. TPC-DS

The test dataset is 100GB in size, and both Parquet and ORC file formats are tested.

This test only tests the first 10 queries.

Spark Thrift JDBC/ODBC Server is used to start the Spark resident process and then submit the task via Beeline connection.

Test Hardware

Software Configuration

Spark Thrift JDBC/ODBC Server

${SPARK_HOME}/sbin/start-thriftserver.sh \
  --master yarn \
  --driver-memory 8g \
  --executor-memory 10g \
  --executor-cores 3 \
  --num-executors 3 \
  --conf spark.locality.wait=100 \
  --conf spark.sql.crossJoin.enabled=true \
  --hiveconf hive.server2.thrift.port=10001

JuiceFS Cache Configurations

The 2 data disks of Core node are mounted in the /data01 and /data02 directories, and core-site.xml is configured as follows:

<property>
  <name>juicefs.cache-size</name>
  <value>200000</value>
</property>
<property>
  <name>juicefs.cache-dir</name>
  <value>/data*/jfscache</value>
</property>
<property>
  <name>juicefs.cache-full-block</name>
  <value>false</value>
</property>
<property>
  <name>juicefs.discover-nodes-url</name>
  <value>yarn</value>
</property>
<property>
  <name>juicefs.attr-cache</name>
  <value>3</value>
</property>
<property>
  <name>juicefs.entry-cache</name>
  <value>3</value>
</property>
<property>
  <name>juicefs.dir-entry-cache</name>
  <value>3</value>
</property>

Test

The task submission command is as follows:

${SPARK_HOME}/bin/beeline -u jdbc:hive2://localhost:10001/${DATABASE} \
  -n hadoop \
  -f query{i}.sql

Results

JuiceFS can use local disk as a cache to accelerate data access, the following data is the result (in seconds) after 4 runs using Redis and TiKV as the metadata engine of JuiceFS respectively.

ORC

Parquet

FAQ

1. Class io.juicefs.JuiceFileSystem not found exception

It means JAR file was not loaded, you can verify it by lsof -p {pid} | grep juicefs.

You should check whether the JAR file was located properly, or other users have the read permission.

Some Hadoop distribution also need to modify mapred-site.xml and put the JAR file location path to the end of the parameter mapreduce.application.classpath.

2. No FilesSystem for scheme: jfs exception

It means JuiceFS Hadoop Java SDK was not configured properly, you need to check whether there is JuiceFS related configuration in the core-site.xml of the component configuration.

3. What are the similarities and differences between user permission management in JuiceFS and HDFS?

JuiceFS also uses the "User/Group" method to manage file permissions, using local users and groups by default. In order to ensure the unified permissions of different nodes during distributed computing, you can configure global "User/UID" and "Group/GID" mappings through juicefs.users and juicefs.groups configurations.

4. After the data is deleted, it is directly stored in the .trash directory of JuiceFS. Although the files are all there, it is difficult to restore the data through the mv command as easily as HDFS. Is there any way to achieve a similar effect of HDFS trash?

In the Hadoop application scenario, the functions similar to the HDFS trash are still retained. It needs to be explicitly enabled by fs.trash.interval and fs.trash.checkpoint.interval configurations, please refer to document for more information.

5. What are the benefits of setting the juicefs.discover-nodes-url configuration?

In HDFS, each data block will have BlockLocation information, which the computing engine uses to schedule the computing tasks as much as possible to the nodes where the data is stored. JuiceFS will calculate the corresponding BlockLocation for each data block through the consistent hashing algorithm, so that when the same data is read for the second time, the computing engine may schedule the computing task to the same node, and the data cached on the local disk during the first computing can be used to accelerate data access.

This algorithm needs to know all the computing node information in advance. The juicefs.discover-nodes-url configuration is used to obtain these computing node information.

6. Does the community version of JuiceFS currently support a Kerberos-authenticated CDH cluster?

Not supported. JuiceFS does not verify the validity of Kerberos users, but can use Kerberos-authenticated username.