JanusGraph批量导入数据优化

批量导入工具：

https://github.com/dengziming/janusgraph-util

批量导入配置项

storage.batch-loading =true

• 导入的数据必须具有一致性并且和已存在的数据必须具有一致性。（比如：name数据是具有唯一索引（a unique composite index），那么导入的数据在name属性上上和已有的数据不能重复）

• 下面是优化配置，优化的目的，就是减少批量导入时间。

ID 分配优化

ID Block Size

ids.block-size

• 配置项，JanusGraph实例通过id池管理对象从id blocks中获取ids值为新加入的vertex、edge分配唯一id，为了保证库唯一性，所以获取id block（id块）是昂贵的（因为存在多个实例竞争），所以增加block-size可以减少获取block的次数，但是值过大会导致多余的id被浪费掉。
• 一般情况下事务的负载，ids.block-size的默认值是满足要求的。但是对于批量导入时，需要调节值为每个JanusGraph实例需要添加节点和边数的10倍。
• 该配置项在集群中所有实例上值必须唯一。

ID Acquisition Process

1) ids.authority.wait-time

• 配置毫秒：id池管理器允许id block获取程序最大允许等待时间，时间到还未获取到就失败。建议值设置为存储后端第95%时读写时间之和。
• 该配置在集群上要设置统一值。

2) ids.renew-timeout

• 在失败获取id block后，id池管理程序等待多少毫秒后再次发起一个新的尝试。尽可能大。

读写优化

Buffer Size

storage.buffer-size  缓存大小

• JanusGraph会缓存写操作，然后批量发送到后端执行。这样可以减少请求次数，从而避免短时间内执行服务器写请求过多导致的失败。
• 如果缓存设置太大，会增加写延迟因此会增加执行失败的可能性。
• 建议：谨慎设置该值。

Read and Write Robustness

• 如果存储后端的写或读操作失败后（storage.buffer-size 太大会增加失败的可能性），将会重试多少次才会放弃。

storage.read-attempts   读尝试的次数
storage.write-attempts  写尝试次数
storage.attempt-wait  ：两次尝试之间的时间间隔，在批量导入情况下，此值可以设置大一些

策略

Parallelizing the Load

• 如果存储后端可以可承受足够的请求，那么可以在多个机器上并行批量导入，可以减少导入时间。
• Chapter 35, JanusGraph with TinkerPop’s Hadoop-Gremlin 通过MapReduce批量导入数据。
• 如果不使用Hadoop，则可以将大图拆分成小图并行导入。
• 如果图不可以被拆分，那么可以将顶点和边分开并行导入。

使用操作：

基本配置

1. 首先从官网下载并解压janusgraph到本地/data/janusgraph/目录。
2. 然后配置图数据库前后端。由于我们用的是es + hbase， 所以直接修改/data/janusgraph/conf/janusgraph-hbase-es.properties ：

#重要
gremlin.graph=org.janusgraph.core.JanusGraphFactory
#hbase配置
storage.batch-loading=true
storage.backend=hbase
storage.hostname=c1-nn1.bdp.idc,c1-nn2.bdp.idc,c1-nn3.bdp.idc
storage.hbase.ext.hbase.zookeeper.property.clientPort=2181
storage.hbase.table = yisou:test_graph
#es配置
index.search.backend=elasticsearch
index.search.hostname=10.120.64.69  #es是只安装在本地，此为本机ip。
index.search.elasticsearch.client-only=true
index.search.index-name=yisou_test_graph
#默认cache配置
cache.db-cache = true
cache.db-cache-clean-wait = 20
cache.db-cache-time = 180000
cache.db-cache-size = 0.5

3.修改/data/janusgraph/lib下的jar包。由于在跑yarn-client批量导入时有guava等jar包冲突，我根据冲突情况对lib下面的jar包作了调整。主要调整了3个jar包：

1. hbase-client-1.2.4.jar ==> yisou-hbase-1.0-SNAPSHOT.jar
   由于lib下的hbase-client-1.2.4.jar用的guava与我们yarn集群的guava版本有冲突，所以我们用了公司内部的去除了guava的hbase-client，即yisou-hbase-1.0-SNAPSHOT.jar 。
   如果不替换，报错 "Caused by: java.lang.IllegalAccessError: tried to access method com.google.common.base.Stopwatch.<init>()V from class org.apache.hadoop.hbase.zookeeper.MetaTableLocator"
2. spark-assembly-1.6.1-hadoop2.6.0.jar ==> spark-assembly-1.6.2-hadoop2.6.0.jar
   lib自带的spark-assembly-1.6.1-hadoop2.6.0.jar也会引起guava冲突，我将其替换成spark-assembly-1.6.2-hadoop2.6.0.jar。
   如果不替换，将会报错"java.lang.NoSuchMethodError: groovy.lang.MetaClassImpl.hasCustomStaticInvokeMethod()Z"
3. 删除 hbase-protocol-1.2.4.jar.
   如果不删除，将会报错 "com.google.protobuf.ServiceException: java.lang.NoSuchMethodError: org.apache.hadoop.hbase.protobuf.generated.RPCProtos$ConnectionHeader$Builder.setVersionInfo(Lorg/apache/hadoop/hbase/protobuf/generated/RPCProtos$VersionInfo;)Lorg/apache/hadoop/hbase/protobuf/generated/RPCProtos$ConnectionHeader$Builder;"

4.配置图中边和节点属性，具体参考官网，本文不展开。

批量导入配置

由于需要与yarn配合，将导入程序放在yarn上执行，所以需要hadoop相关环境配置。需要修改两个配置文件，一个是Janusgraph的启动脚本/data/janusgraph/lib/gremlin.sh, 另一个是hadoop和spark相关的配置/data/janusgraph/conf/hadoop-graph/hadoop-script.properties。

1.复制/data/janusgraph/lib/gremlin.sh, 假定命名为yarn-gremlin.sh。 然后增加hadoop的配置到JAVA_OPTIONS和CLASSPATH中。这样能保证hadoop相关配置能被程序读取到，便于正常启动spark在yarn上的任务。

#!/bin/bash
export HADOOP_CONF_DIR=$HADOOP_HOME/etc/hadoop
export HADOOP_HOME=/usr/local/hadoop-2.7.1
export JAVA_OPTIONS="$JAVA_OPTIONS -Djava.library.path=$HADOOP_HOME/lib/native"
export CLASSPATH=$HADOOP_CONF_DIR
#JANUSGRAPH_HOME为用户安装janusgraph的目录/data/janusgraph/
cd $JANUSGRAPH_HOME
./bin/gremlin.sh

2.修改/data/janusgraph/conf/hadoop-graph/hadoop-script.properties
主要根据要导入文件的格式修改inputFormat、指定要导入的hdfs文件路径、parse函数路径以及spark master指定为yarn-client等。

#
# Hadoop Graph Configuration
#
gremlin.graph=org.apache.tinkerpop.gremlin.hadoop.structure.HadoopGraph
gremlin.hadoop.graphInputFormat=org.apache.tinkerpop.gremlin.hadoop.structure.io.script.ScriptInputFormat
gremlin.hadoop.graphOutputFormat=org.apache.tinkerpop.gremlin.hadoop.structure.io.graphson.GraphSONOutputFormat
gremlin.hadoop.jarsInDistributedCache=true

#导入文件的hdfs路径。也可以在加载该配置文件后指定
gremlin.hadoop.inputLocation=/user/yisou/taotian1/janus/data/fewData.test.dup
#解析hdfs文件的parse函数路径。也可以在加载该配置文件后指定
gremlin.hadoop.scriptInputFormat.script=/user/yisou/taotian1/janus/data/conf/vertex_parse.groovy
#gremlin.hadoop.outputLocation=output

#
# SparkGraphComputer with Yarn Configuration
#
spark.master=yarn-client
spark.executor.memory=6g
spark.executor.instances=10
spark.executor.cores=2
spark.serializer=org.apache.spark.serializer.KryoSerializer
# spark.kryo.registrationRequired=true
# spark.storage.memoryFraction=0.2
# spark.eventLog.enabled=true
# spark.eventLog.dir=/tmp/spark-event-logs
# spark.ui.killEnabled=true

#cache config
gremlin.spark.persistContext=true
gremlin.spark.graphStorageLevel=MEMORY_AND_DISK
#gremlin.spark.persistStorageLevel=DISK_ONLY


#####################################
# GiraphGraphComputer Configuration #
#####################################
giraph.minWorkers=2
giraph.maxWorkers=3
giraph.useOutOfCoreGraph=true
giraph.useOutOfCoreMessages=true
mapred.map.child.java.opts=-Xmx1024m
mapred.reduce.child.java.opts=-Xmx1024m
giraph.numInputThreads=4
giraph.numComputeThreads=4
# giraph.maxPartitionsInMemory=1
# giraph.userPartitionCount=2

执行批量导入

启动命令：

sh /data/janusgraph/lib/yarn-gremlin.sh

批量导入命令：

local_root="/data/janusgraph"
hdfs_root="/user/yisou/taotian1/janus"
social_graph="${local_root}/conf/janusgraph-hbase-es.properties"
graph = GraphFactory.open("${local_root}/conf/hadoop-script.properties")
graph.configuration().setProperty("gremlin.hadoop.inputLocation","/user/yisou/taotian1/janus/data/fewData.test.dup")
graph.configuration().setProperty("gremlin.hadoop.scriptInputFormat.script", "${hdfs_root}/conf/vertex_parse.groovy")
blvp = BulkLoaderVertexProgram.build().writeGraph(social_graph).create(graph)
graph.compute(SparkGraphComputer).program(blvp).submit().get()

运行结果：

sh /data/janusgraph/lib/yarn-gremlin.sh
\,,,/
(o o)
-----oOOo-(3)-oOOo-----
plugin activated: janusgraph.imports
plugin activated: tinkerpop.server
plugin activated: tinkerpop.utilities
SLF4J: Class path contains multiple SLF4J bindings.
SLF4J: Found binding in [jar:file:/data2/janusgraph-0.1.1-hadoop2/lib/slf4j-log4j12-1.7.12.jar!/org/slf4j/impl/StaticLoggerBinder.class]
SLF4J: Found binding in [jar:file:/data2/janusgraph-0.1.1-hadoop2/lib/logback-classic-1.1.2.jar!/org/slf4j/impl/StaticLoggerBinder.class]
SLF4J: Found binding in [jar:file:/data2/janusgraph-0.1.1-hadoop2/lib/spark-assembly-1.6.2-hadoop2.6.0.jar!/org/slf4j/impl/StaticLoggerBinder.class]
SLF4J: Found binding in [jar:file:/data2/janusgraph-0.1.1-hadoop2/lib/yisou-hbase-1.0-SNAPSHOT.jar!/org/slf4j/impl/StaticLoggerBinder.class]
SLF4J: See http://www.slf4j.org/codes.html#multiple_bindings for an explanation.
SLF4J: Actual binding is of type [org.slf4j.impl.Log4jLoggerFactory]
21:22:00,392  INFO HadoopGraph:87 - HADOOP_GREMLIN_LIBS is set to: /data2/janusgraph-0.1.1-hadoop2/lib
plugin activated: tinkerpop.hadoop
plugin activated: tinkerpop.spark
plugin activated: tinkerpop.tinkergraph
gremlin>
gremlin> local_root="/data2/janusgraph-0.1.1-hadoop2/social"
==>/data2/janusgraph-0.1.1-hadoop2/social
gremlin> hdfs_root="/user/yisou/taotian1/janus"
==>/user/yisou/taotian1/janus
gremlin> social_graph="${local_root}/conf/janusgraph-hbase-es-social.properties"
==>/data2/janusgraph-0.1.1-hadoop2/social/conf/janusgraph-hbase-es-social.properties
gremlin> graph = GraphFactory.open("${local_root}/conf/hadoop-yarn.properties")
==>hadoopgraph[scriptinputformat->graphsonoutputformat]
gremlin> graph.configuration().setProperty("gremlin.hadoop.inputLocation","/user/yisou/taotian1/janus/tmp1person/")
==>null
gremlin> graph.configuration().setProperty("gremlin.hadoop.scriptInputFormat.script", "${hdfs_root}/person_parse.groovy")
==>null
gremlin> blvp = BulkLoaderVertexProgram.build().writeGraph(social_graph).create(graph)
==>BulkLoaderVertexProgram[bulkLoader=IncrementalBulkLoader, vertexIdProperty=bulkLoader.vertex.id, userSuppliedIds=false, keepOriginalIds=true, batchSize=0]
gremlin> graph.compute(SparkGraphComputer).program(blvp).submit().get()
21:25:04,666  INFO deprecation:1173 - mapred.reduce.child.java.opts is deprecated. Instead, use mapreduce.reduce.java.opts
21:25:04,667  INFO deprecation:1173 - mapred.map.child.java.opts is deprecated. Instead, use mapreduce.map.java.opts
21:25:04,680  INFO KryoShimServiceLoader:117 - Set KryoShimService provider to org.apache.tinkerpop.gremlin.hadoop.structure.io.HadoopPoolShimService@4cb2918c (class org.apache.tinkerpop.gremlin.hadoop.structure.io.HadoopPoolShimService) because its priority value (0) is the highest available
21:25:04,680  INFO KryoShimServiceLoader:123 - Configuring KryoShimService provider org.apache.tinkerpop.gremlin.hadoop.structure.io.HadoopPoolShimService@4cb2918c with user-provided configuration
  21:25:10,479  WARN SparkConf:70 - The configuration key 'spark.yarn.user.classpath.first' has been deprecated as of Spark 1.3 and may be removed in the future. Please use spark.{driver,executor}.userClassPathFirst instead.
21:25:10,505  INFO SparkContext:58 - Running Spark version 1.6.2
21:25:10,524  WARN SparkConf:70 - The configuration key 'spark.yarn.user.classpath.first' has been deprecated as of Spark 1.3 and may be removed in the future. Please use spark.{driver,executor}.userClassPathFirst instead.
21:25:10,564  INFO SecurityManager:58 - Changing view acls to: yisou
21:25:10,565  INFO SecurityManager:58 - Changing modify acls to: yisou
21:25:10,566  INFO SecurityManager:58 - SecurityManager: authentication disabled; ui acls disabled; users with view permissions: Set(yisou); users with modify permissions: Set(yisou)
21:25:10,833  WARN SparkConf:70 - The configuration key 'spark.yarn.user.classpath.first' has been deprecated as of Spark 1.3 and may be removed in the future. Please use spark.{driver,executor}.userClassPathFirst instead.
21:25:10,835  WARN SparkConf:70 - The configuration key 'spark.yarn.user.classpath.first' has been deprecated as of Spark 1.3 and may be removed in the future. Please use spark.{driver,executor}.userClassPathFirst instead.
21:25:11,035  INFO Utils:58 - Successfully started service 'sparkDriver' on port 36502.
21:25:11,576  INFO Slf4jLogger:80 - Slf4jLogger started
  21:25:11,646  INFO Remoting:74 - Starting remoting
............
21:25:20,736  INFO Client:58 - Submitting application 2727164 to ResourceManager
21:25:20,771  INFO YarnClientImpl:273 - Submitted application application_1466564207556_2727164
21:25:21,780  INFO Client:58 - Application report for application_1466564207556_2727164 (state: ACCEPTED)
21:25:21,785  INFO Client:58 -
client token: N/A
diagnostics: N/A
ApplicationMaster host: N/A
ApplicationMaster RPC port: -1
queue: root.yisou
start time: 1500297920750
final status: UNDEFINED
tracking URL: http://c1-nn3.bdp.idc:8981/proxy/application_1466564207556_2727164/
21:25:22,787  INFO Client:58 - Application report for application_1466564207556_2727164 (state: ACCEPTED)
21:25:23,789  INFO Client:58 - Application report for application_1466564207556_2727164 (state: ACCEPTED)
21:25:24,791  INFO Client:58 - Application report for application_1466564207556_2727164 (state: ACCEPTED)
21:25:25,793  INFO Client:58 - Application report for application_1466564207556_2727164 (state: ACCEPTED)
21:25:39,585  INFO JettyUtils:58 - Adding filter: org.apache.hadoop.yarn.server.webproxy.amfilter.AmIpFilter
21:25:39,823  INFO Client:58 - Application report for application_1466564207556_2727164 (state: RUNNING)
21:25:39,824  INFO Client:58 -
client token: N/A
diagnostics: N/A
ApplicationMaster host: 10.130.1.50
ApplicationMaster RPC port: 0
queue: root.yisou
start time: 1500297920750
final status: UNDEFINED
tracking URL: http://c1-nn3.bdp.idc:8981/proxy/application_1466564207556_2727164/
..........
21:25:42,864  INFO SparkContext:58 - Added JAR /data2/janusgraph-0.1.1-hadoop2/lib/commons-codec-1.7.jar at http://10.130.64.69:38209/jars/commons-codec-1.7.jar with timestamp 1500297942864
21:25:42,866  INFO SparkContext:58 - Added JAR /data2/janusgraph-0.1.1-hadoop2/lib/commons-lang-2.5.jar at http://10.130.64.69:38209/jars/commons-lang-2.5.jar with timestamp 1500297942866
21:25:42,869  INFO SparkContext:58 - Added JAR /data2/janusgraph-0.1.1-hadoop2/lib/commons-collections-3.2.2.jar at http://10.130.64.69:38209/jars/commons-collections-3.2.2.jar with timestamp 1500297942869
21:25:42,872  INFO SparkContext:58 - Added JAR /data2/janusgraph-0.1.1-hadoop2/lib/commons-io-2.3.jar at http://10.130.64.69:38209/jars/commons-io-2.3.jar with timestamp 1500297942872
21:25:42,874  INFO SparkContext:58 - Added JAR /data2/janusgraph-0.1.1-hadoop2/lib/jetty-util-6.1.26.jar at http://10.130.64.69:38209/jars/jetty-util-6.1.26.jar with timestamp 1500297942874
21:25:42,879  INFO SparkContext:58 - Added JAR /data2/janusgraph-0.1.1-hadoop2/lib/htrace-core-3.1.0-incubating.jar at http://10.130.64.69:38209/jars/htrace-core-3.1.0-incubating.jar with timestamp 1
............
21:26:14,751  INFO MapOutputTrackerMaster:58 - Size of output statuses for shuffle 2 is 146 bytes
21:26:14,767  INFO TaskSetManager:58 - Finished task 0.0 in stage 6.0 (TID 4) in 40 ms on c1-dn31.bdp.idc (1/1)
21:26:14,767  INFO YarnScheduler:58 - Removed TaskSet 6.0, whose tasks have all completed, from pool
21:26:14,767  INFO DAGScheduler:58 - ResultStage 6 (foreachPartition at SparkExecutor.java:173) finished in 0.042 s
21:26:14,768  INFO DAGScheduler:58 - Job 1 finished: foreachPartition at SparkExecutor.java:173, took 1.776125 s
21:26:14,775  INFO ShuffledRDD:58 - Removing RDD 2 from persistence list
21:26:14,785  INFO BlockManager:58 - Removing RDD 2
==>result[hadoopgraph[scriptinputformat->graphsonoutputformat],memory[size:0]]
gremlin> 21:26:22,515  INFO YarnClientSchedulerBackend:58 - Registered executor NettyRpcEndpointRef(null) (c1-dn9.bdp.idc:60762) with ID 8

批量导入性能优化

如果不做优化，janusgraph批量导入的速度非常慢，导入4千万条数据大约需要3.5小时。优化后可降低到1小时.
1.加大ids.block-size和storage.buffer-size参数的大小（在janusgraph-hbase-es.properties中配置）。
ids.block-size=100000000
storage.buffer-size=102400

2.指定hbase初始的region数目（在janusgraph-hbase-es.properties中配置）。
storage.hbase.region-count = 50

3.边和顶点同时导入，而不是顶点和边分成不同的文件，分开导入。格式可参考/data/janusgraph/data/grateful-dead.txt。

总结

本文主要讲解了janusgraph中如何配置yarn-client的方式批量导入节点和边。

分为基本配置和批量导入的配置两部分，基本配置中需要注意janusgraph自带jar包与用户yarn环境中jar包的冲突问题，可替换或者删除相关jar包。

批量导入配置中重点是在gremlin.sh中添加hadoop的相关配置，将hadoop环境配置到JAVA_OPTIONS和CLASSPATH中。

参考链接

Titan 数据库使用
图数据库Titan在生产环境中的使用全过程+分析
合并顶点和边，批量导入parse函数样例
Yet Another Analytics & Intelligence Communication Series