Hi Mladen,

Interesting read! Spark is not very sensitive to the number of tasks. I believe that for OLAP on HadoopGraph the optimum is for partitions of 256 Mb or so. Larger is difficult to hold in memory for reasonably sized executors. Smaller gives too much overhead. OLAP with janusgraph-hbase is much harder, because the partition size is determined by the HBase regions which need to be large (10GB). Also note that the entire graph needs to fit into the total memory of all executors  because graph traversing is shuffle-heavy and spilling to disk will take endlessly.

Best wishes,    Marc

Op maandag 30 november 2020 om 19:09:15 UTC+1 schreef Mladen Marović:

I know I'm quite late to the party, but for future reference - the number of input partitions in Spark depends on the partitioning of the source. In case of cassandra, partitioning is determined by the number of tokens each node gets (as configured by `num_tokens` in `cassandra.yaml`), which is set to 256 by default. So, if you have a 3-node cassandra cluster, by default each node should get 256 tokens, which would result in 3*256 = 768 tokens total. Since Spark reads directly from cassandra (if you're using `org.janusgraph.hadoop.formats.cql.CqlInputFormat`), that translates to 768 partitions in the input Spark RDD, or 768 tasks during processing. Add to that 1 task that collects results, or something similar, and you end up at 769. At least that was my experience.

The default value of 256 for `num_tokens` made sense in older versions, but in cassandra 3.x a new token allocation algorithm was implemented to improve performance for operations requiring token-range scans, which is precisely what Spark does. I experimented a bit with smaller values (e.g. 16) and managed to drastically reduce the number of tasks when scanning the entire graph. For further, reading, I recommend this article.

On Thursday, December 5, 2019 at 9:28:26 AM UTC+1 s...@... wrote:

Answering my own question - turned out I had had a mixup of keyspaces used between the two instances

Default the conf/hadoop-graph/read-cql.properties reads

janusgraphmr.ioformat.conf.storage.cassandra.keyspace

While for CQL it should read

janusgraphmr.ioformat.conf.storage.cql.keyspace

Also - as I made a 'named' (ve_graph) graph I had to point to that one rather than the janusgraph keyspace.

Problem 1 solved. Now to the next - how can I lower the number of 'partitions' Spark is using (here 796  '... on localhost (executor driver) (769/769)')?  

On Wednesday, December 4, 2019 at 11:46:42 PM UTC+1, Sture Lygren wrote:

Hi,

I'm trying to get JanusGraph 0.4.0 with a Cassandra (CQL) backend setup and running as OLAP while still keeping OLTP active in order to do graph updates. I've been searching high and low for some guidance, but so far without any luck. Hopefully someone here could tune in and help?

Here's where I'm at currently

• local Hadoop running according to https://old-docs.janusgraph.org/0.4.0/hadoop-tp3.html
• gremlin server started as /bin/gremlin-server.sh conf/gremlin-server/gremlin-server-configuration.yaml
• gremlin-server-configuration.yaml points to init.groovy script doing the traversal mappings for OLTP and OLAP

def globals = [:]

ve = ConfiguredGraphFactory.open("ve_graph")

OLAPGraph = GraphFactory.open('conf/hadoop-graph/read-cql.properties')

globals << [g : ve.traversal(), sg: OLAPGraph.traversal().withComputer(org.apache.tinkerpop.gremlin.spark.process.computer.SparkGraphComputer)]

• conf/hadoop-graph/read-cql.properties reads

gremlin.graph=org.apache.tinkerpop.gremlin.hadoop.structure.HadoopGraph

gremlin.hadoop.graphReader=org.janusgraph.hadoop.formats.cql.CqlInputFormat

gremlin.hadoop.graphWriter=org.apache.tinkerpop.gremlin.hadoop.structure.io.gryo.GryoOutputFormat

gremlin.hadoop.jarsInDistributedCache=true

gremlin.hadoop.inputLocation=none

gremlin.hadoop.outputLocation=output

gremlin.spark.persistContext=true

janusgraphmr.ioformat.conf.storage.backend=cql

janusgraphmr.ioformat.conf.storage.hostname=127.0.0.1

janusgraphmr.ioformat.conf.storage.port=9042

janusgraphmr.ioformat.conf.storage.cassandra.keyspace=janusgraph

cassandra.input.partitioner.class=org.apache.cassandra.dht.Murmur3Partitioner

spark.serializer=org.apache.spark.serializer.KryoSerializer

spark.kryo.registrator=org.janusgraph.hadoop.serialize.JanusGraphKryoRegistrator

• Running the gremlin shell I have

         \,,,/

         (o o)

-----oOOo-(3)-oOOo-----

SLF4J: Class path contains multiple SLF4J bindings.

SLF4J: Found binding in [jar:file:/data/sture/Scripts/janusgraph-0.4.0-hadoop2/lib/slf4j-log4j12-1.7.12.jar!/org/slf4j/impl/StaticLoggerBinder.class]

SLF4J: Found binding in [jar:file:/data/sture/Scripts/janusgraph-0.4.0-hadoop2/lib/logback-classic-1.1.3.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]

plugin activated: tinkerpop.server

plugin activated: tinkerpop.tinkergraph

plugin activated: tinkerpop.hadoop

plugin activated: tinkerpop.spark

plugin activated: tinkerpop.utilities

plugin activated: janusgraph.imports

gremlin> :remote connect tinkerpop.server conf/remote.yaml session

==>Configured localhost/127.0.0.1:8182-[655848fc-b46e-40be-8174-f0dc42cdabd4]

gremlin> :remote console

==>All scripts will now be sent to Gremlin Server - [localhost/127.0.0.1:8182]-[655848fc-b46e-40be-8174-f0dc42cdabd4] - type ':remote console' to return to local mode

gremlin> g

==>graphtraversalsource[standardjanusgraph[cql:[127.0.0.1]], standard]

gremlin>

gremlin> sg

==>graphtraversalsource[hadoopgraph[cqlinputformat->gryooutputformat], sparkgraphcomputer]

gremlin> g.V().has('lbl','System').count()

==>68

gremlin> sg.V().has('lbl','System').count()

• The job is running for some time and while finishing the gremlin-server.log reads

253856 [Executor task launch worker for task 768] INFO  org.apache.spark.executor.Executor  - Finished task 768.0 in stage 0.0 (TID 768). 2388 bytes result sent to driver

253858 [task-result-getter-1] INFO  org.apache.spark.scheduler.TaskSetManager  - Finished task 768.0 in stage 0.0 (TID 768) in 6809 ms on localhost (executor driver) (769/769)

253861 [dag-scheduler-event-loop] INFO  org.apache.spark.scheduler.DAGScheduler  - ResultStage 0 (fold at SparkStarBarrierInterceptor.java:101) finished in 161.427 s

253861 [task-result-getter-1] INFO  org.apache.spark.scheduler.TaskSchedulerImpl  - Removed TaskSet 0.0, whose tasks have all completed, from pool

253876 [SparkGraphComputer-boss] INFO  org.apache.spark.scheduler.DAGScheduler  - Job 0 finished: fold at SparkStarBarrierInterceptor.java:101, took 161.598267 s

253888 [SparkGraphComputer-boss] INFO  org.apache.spark.rdd.MapPartitionsRDD  - Removing RDD 1 from persistence list

253901 [block-manager-slave-async-thread-pool-0] INFO  org.apache.spark.storage.BlockManager  - Removing RDD 1

• However - the count (==> ) reads 0 for the sg traversal

I've most likely missed some crucial point here, but I'm not able to spot it. Please help.