Apache Uniffle (Incubating)

Uniffle is a high performance, general purpose remote shuffle service for distributed computing engines. It provides the ability to push shuffle data into centralized storage service, changing the shuffle style from "local file pull-like style" to "remote block push-like style". It brings in several advantages like supporting disaggregated storage deployment, super large shuffle jobs, and high elasticity. Currently it supports Apache Spark, Apache Hadoop MapReduce and Apache Tez.

Architecture

Uniffle cluster consists of three components, a coordinator cluster, a shuffle server cluster and an optional remote storage (e.g., HDFS).

Coordinator will collect the status of shuffle servers and assign jobs based on some strategy.

Shuffle server will receive the shuffle data, merge them and write to storage.

Depending on different situations, Uniffle supports Memory & Local, Memory & Remote Storage(e.g., HDFS), Memory & Local & Remote Storage(recommendation for production environment).

Shuffle Process with Uniffle

• Spark driver ask coordinator to get shuffle server for shuffle process

• Spark task write shuffle data to shuffle server with following step:

  1. Send KV data to buffer
  2. Flush buffer to queue when buffer is full or buffer manager is full
  3. Thread pool get data from queue
  4. Request memory from shuffle server first and send the shuffle data
  5. Shuffle server cache data in memory first and flush to queue when buffer manager is full
  6. Thread pool get data from queue
  7. Write data to storage with index file and data file
  8. After write data, task report all blockId to shuffle server, this step is used for data validation later
  9. Store taskAttemptId in MapStatus to support Spark speculation

• Depending on different storage types, the spark task will read shuffle data from shuffle server or remote storage or both of them.

Shuffle file format

The shuffle data is stored with index file and data file. Data file has all blocks for a specific partition and the index file has metadata for every block.

Supported Spark Version

Currently supports Spark 2.3.x, Spark 2.4.x, Spark 3.0.x, Spark 3.1.x, Spark 3.2.x, Spark 3.3.x, Spark 3.4.x, Spark 3.5.x

Note: To support dynamic allocation, the patch(which is included in patch/spark folder) should be applied to Spark

Supported MapReduce Version

Currently supports the MapReduce framework of Hadoop 2.8.5, Hadoop 3.2.1

Building Uniffle

note: currently Uniffle requires JDK 1.8 to build, adding later JDK support is on our roadmap.

Uniffle is built using Apache Maven. To build it, run:

./mvnw -DskipTests clean package

To fix code style issues, run:

./mvnw spotless:apply -Pspark3 -Pspark2 -Ptez -Pmr -Phadoop2.8 -Pdashboard

Build against profile Spark 2 (2.4.6)

./mvnw -DskipTests clean package -Pspark2

Build against profile Spark 3 (3.1.2)

./mvnw -DskipTests clean package -Pspark3

Build against Spark 3.2.x, Except 3.2.0

./mvnw -DskipTests clean package -Pspark3.2

Build against Spark 3.2.0

./mvnw -DskipTests clean package -Pspark3.2.0

Build against Hadoop MapReduce 2.8.5

./mvnw -DskipTests clean package -Pmr,hadoop2.8

Build against Hadoop MapReduce 3.2.1

./mvnw -DskipTests clean package -Pmr,hadoop3.2

Build against Tez 0.9.1

./mvnw -DskipTests clean package -Ptez

Build against Tez 0.9.1 and Hadoop 3.2.1

./mvnw -DskipTests clean package -Ptez,hadoop3.2

Build with dashboard

./mvnw -DskipTests clean package -Pdashboard

note: currently Uniffle build the project against Java 8. If you want to compile it against other Java versions, you can build the code with -Dmaven.compiler.release=${release-version}.

To package the Uniffle, run:

./build_distribution.sh

Package against Spark 3.2.x, Except 3.2.0, run:

./build_distribution.sh --spark3-profile 'spark3.2'

Package against Spark 3.2.0, run:

./build_distribution.sh --spark3-profile 'spark3.2.0'

Package will build against Hadoop 2.8.5 in default. If you want to build package against Hadoop 3.2.1, run:

./build_distribution.sh --hadoop-profile 'hadoop3.2'

Package with hadoop jars, If you want to build package against Hadoop 3.2.1, run:

./build_distribution.sh --hadoop-profile 'hadoop3.2' -Phadoop-dependencies-included

rss-xxx.tgz will be generated for deployment

Deploy

If you have packaged tgz with hadoop jars, the env of HADOOP_HOME is needn't specified in rss-env.sh.

Deploy Coordinator

1. unzip package to RSS_HOME
2. update RSS_HOME/conf/rss-env.sh, e.g.,

    JAVA_HOME=<java_home>
    HADOOP_HOME=<hadoop home>
    COORDINATOR_XMX_SIZE="16g"
    # You can set coordinator memory size by `XMX_SIZE` too, but it affects all components.
    # XMX_SIZE="16g"

3. update RSS_HOME/conf/coordinator.conf, e.g.,

    rss.rpc.server.port 19999
    rss.jetty.http.port 19998
    rss.coordinator.server.heartbeat.timeout 30000
    rss.coordinator.app.expired 60000
    rss.coordinator.shuffle.nodes.max 5
    # enable dynamicClientConf, and coordinator will be responsible for most of client conf
    rss.coordinator.dynamicClientConf.enabled true
    # config the path of client conf
    rss.coordinator.dynamicClientConf.path <RSS_HOME>/conf/dynamic_client.conf
    # config the path of excluded shuffle server
    rss.coordinator.exclude.nodes.file.path <RSS_HOME>/conf/exclude_nodes

4. update /conf/dynamic_client.conf, rss client will get default conf from coordinator e.g.,

   # MEMORY_LOCALFILE_HDFS is recommended for production environment
   rss.storage.type MEMORY_LOCALFILE_HDFS
   # multiple remote storages are supported, and client will get assignment from coordinator
   rss.coordinator.remote.storage.path hdfs://cluster1/path,hdfs://cluster2/path
   rss.writer.require.memory.retryMax 1200
   rss.client.retry.max 50
   rss.client.send.check.timeout.ms 600000
   rss.client.read.buffer.size 14m

5. start Coordinator

   bash RSS_HOME/bin/start-coordnator.sh

Deploy Shuffle Server

We recommend to use JDK 11+ if we want to have better performance when we deploy the shuffle server. Some benchmark tests among different JDK is as below: (using spark to write shuffle data with 20 executors. Single executor will total write 1G, and each time write 14M. Shuffle Server use GRPC to transfer data)

Deploy Steps:

1. unzip package to RSS_HOME
2. update RSS_HOME/conf/rss-env.sh, e.g.,

    JAVA_HOME=<java_home>
    HADOOP_HOME=<hadoop home>
    SHUFFLE_SERVER_XMX_SIZE="80g"
    # You can set shuffle server memory size by `XMX_SIZE` too, but it affects all components.
    # XMX_SIZE="80g"

3. update RSS_HOME/conf/server.conf, e.g.,

    rss.rpc.server.port 19999
    rss.jetty.http.port 19998
    rss.rpc.executor.size 2000
    # it should be configured the same as in coordinator
    rss.storage.type MEMORY_LOCALFILE_HDFS
    rss.coordinator.quorum <coordinatorIp1>:19999,<coordinatorIp2>:19999
    # local storage path for shuffle server
    rss.storage.basePath /data1/rssdata,/data2/rssdata....
    # it's better to config thread num according to local disk num
    rss.server.flush.thread.alive 5
    rss.server.flush.localfile.threadPool.size 10
    rss.server.flush.hadoop.threadPool.size 60
    rss.server.buffer.capacity 40g
    rss.server.read.buffer.capacity 20g
    rss.server.heartbeat.interval 10000
    rss.rpc.message.max.size 1073741824
    rss.server.preAllocation.expired 120000
    rss.server.commit.timeout 600000
    rss.server.app.expired.withoutHeartbeat 120000
    # note: the default value of rss.server.flush.cold.storage.threshold.size is 64m
    # there will be no data written to DFS if set it as 100g even rss.storage.type=MEMORY_LOCALFILE_HDFS
    # please set a proper value if DFS is used, e.g., 64m, 128m.
    rss.server.flush.cold.storage.threshold.size 100g

4. start Shuffle Server

   bash RSS_HOME/bin/start-shuffle-server.sh

Deploy Spark Client

1. Add client jar to Spark classpath, e.g., SPARK_HOME/jars/

   The jar for Spark2 is located in /jars/client/spark2/rss-client-spark2-shaded-${version}.jar

   The jar for Spark3 is located in /jars/client/spark3/rss-client-spark3-shaded-${version}.jar

2. Update Spark conf to enable Uniffle, e.g.,

   # Uniffle transmits serialized shuffle data over network, therefore a serializer that supports relocation of
   # serialized object should be used. 
   spark.serializer org.apache.spark.serializer.KryoSerializer # this could also be in the spark-defaults.conf
   spark.shuffle.manager org.apache.spark.shuffle.RssShuffleManager
   spark.rss.coordinator.quorum <coordinatorIp1>:19999,<coordinatorIp2>:19999
   # Note: For Spark2, spark.sql.adaptive.enabled should be false because Spark2 doesn't support AQE.

Support Spark dynamic allocation

To support spark dynamic allocation with Uniffle, spark code should be updated. There are 7 patches for spark (2.3.4/2.4.6/3.0.1/3.1.2/3.2.1/3.3.1/3.4.1) in patch/spark folder for reference.

After apply the patch and rebuild spark, add following configuration in spark conf to enable dynamic allocation:

  spark.shuffle.service.enabled false
  spark.dynamicAllocation.enabled true

For spark3.5 or above just add one more configuration:

  spark.shuffle.sort.io.plugin.class org.apache.spark.shuffle.RssShuffleDataIo

Deploy MapReduce Client

1. Add client jar to the classpath of each NodeManager, e.g., /share/hadoop/mapreduce/

The jar for MapReduce is located in /jars/client/mr/rss-client-mr-XXXXX-shaded.jar

2. Update MapReduce conf to enable Uniffle, e.g.,

   -Dmapreduce.rss.coordinator.quorum=<coordinatorIp1>:19999,<coordinatorIp2>:19999
   -Dyarn.app.mapreduce.am.command-opts=org.apache.hadoop.mapreduce.v2.app.RssMRAppMaster
   -Dmapreduce.job.map.output.collector.class=org.apache.hadoop.mapred.RssMapOutputCollector
   -Dmapreduce.job.reduce.shuffle.consumer.plugin.class=org.apache.hadoop.mapreduce.task.reduce.RssShuffle

   Note that the RssMRAppMaster will automatically disable slow start (i.e., mapreduce.job.reduce.slowstart.completedmaps=1) and job recovery (i.e., yarn.app.mapreduce.am.job.recovery.enable=false)

Deploy Tez Client

1. Append client jar to pacakge which is set by 'tez.lib.uris'.

In production mode, you can append client jar (rss-client-tez-XXXXX-shaded.jar) to package which is set by 'tez.lib.uris'.

In development mode, you can append client jar (rss-client-tez-XXXXX-shaded.jar) to HADOOP_CLASSPATH.

2. Update tez-site.xml to enable Uniffle.

Note that the RssDAGAppMaster will automatically disable slow start (i.e., tez.shuffle-vertex-manager.min-src-fraction=1, tez.shuffle-vertex-manager.max-src-fraction=1).

Deploy In Kubernetes

We have provided an operator for deploying uniffle in kubernetes environments.

For details, see the following document:

operator docs

Configuration

The important configuration is listed as follows.

Security: Hadoop kerberos authentication

The primary goals of the Uniffle Kerberos security are:

1. to enable secure data access for coordinator/shuffle-servers, like dynamic conf/exclude-node files stored in secured dfs cluster
2. to write shuffle data to kerberos secured dfs cluster for shuffle-servers.

The following security configurations are introduced.

• The proxy user mechanism is used to keep the data isolation in uniffle, which means the shuffle-data written by shuffle-servers is owned by spark app's user. To achieve the this, the login user specified by above config should be as the superuser for HDFS. For more details of related sections, please see Proxy user - Superusers Acting On Behalf Of Other Users

Benchmark

We provide some benchmark tests for Uniffle. For details, you can see Uniffle 0.2.0 Benchmark, Uniffle 0.9.0 Benchmark.

LICENSE

Uniffle is under the Apache License Version 2.0. See the LICENSE file for details.

Contributing

For more information about contributing issues or pull requests, see Uniffle Contributing Guide.