Hadoop/Spark Config

CMPT 732, Fall 2019

Config Objects

There have been config objects around, but we haven't used them much. In MapReduce and Spark:


Configuration conf = this.getConf();
Job job = Job.getInstance(conf, "word count");

conf = SparkConf().setAppName('word count')
sc = SparkContext(conf=conf)

spark = SparkSession.builder.appName('word count').getOrCreate()

The Command Line

We have modified them with command line switches:


yarn jar wordcount.jar WordCount -D mapreduce.job.reduces=3 …

spark-submit --num-executors=4 --executor-memory=4g …
spark-submit --conf spark.executor.memory=4g …

Both of these have the effect of modifying the configuration object (and thus the behaviour of the jobs).

In Code

Config options can also be modified in code (where that makes sense, e.g. not the Spark driver memory):


Configuration conf = this.getConf();
conf.setInt("mapreduce.job.running.map.limit", 5);
conf.setInt("mapreduce.reduce.memory.mb", 4096);
Job job = Job.getInstance(conf, "word count");

conf = SparkConf().setAppName('word count') \
    .set('spark.shuffle.compress', False)
sc = SparkContext(conf=conf)

spark = SparkSession.builder \
    .config('spark.sql.shuffle.partitions', '100') \
    .getOrCreate()

In Code

End result: these have the same effect.


spark-submit --conf spark.io.compression.codec=snappy code.py

spark = SparkSession.builder \
    .config('spark.io.compression.codec', 'snappy') \
    .getOrCreate()

spark = SparkSession.builder.getOrCreate()
spark.conf.set('spark.io.compression.codec', 'snappy')

Config Options

There are options to tune jobs in many, many ways:

Spark Context/Session

We have seen both the SparkContext object (for RDD operations) and SparkSession object (for DataFrame operations).


conf = SparkConf().setAppName('example code') \
        .set('spark.executor.instances', 8)
sc = SparkContext(conf=conf)

spark = SparkSession.builder.appName('example code') \
        .config('spark.executor.instances', 8).getOrCreate()
sc = spark.sparkContext # SparkContext instance already there

Spark Context/Session

Both have similar jobs:

  • Are singletons: exactly one must exist (but SparkSession.getOrCreate will find an existing instance if it's there) to represent the connection to the master.
  • Hold configuration for the application.
  • Give you access to Spark functionality: creating RDDs/​DataFrames, reading files, etc.

Filesystems

In both MapReduce and Spark, we have always accepted the default input filesystem: local files when running locally; HDFS on the cluster.

On our cluster, the default filesystem (Hadoop config fs.defaultFS) is hdfs://nml-cloud-149.cs.sfu.ca:8020, i.e. our HDFS server. This can be overridden with the path URL.

Filesystems

These are equivalent on our cluster (for MapReduce, Spark RDD, Spark DF):


TextInputFormat.addInputPath(job,
    new Path("/user/me/data"));
TextInputFormat.addInputPath(job,
    new Path("hdfs://nml-cloud-149.cs.sfu.ca:8020/user/me/data"));

sc.textFile('/user/me/data')
sc.textFile('hdfs://nml-cloud-149.cs.sfu.ca:8020/user/me/data')

spark.read.csv('/user/me/data')
spark.read.csv('hdfs://nml-cloud-149.cs.sfu.ca:8020/user/me/data')

Filesystems

There are other URL formats you can access (MapReduce or Spark, input or output):


sc.textFile('file:///mnt/share/inputs')
sc.textFile('s3a://s3key:s3secret@bucket/')

Or any Hadoop InputFormat. Or in Spark DataFrames, one of the data source plugins.


df1 = spark.read.format('solr').options(…).load()
df2 = spark.read.format('org.elasticsearch.spark.sql') \
        .options(…).load('index/foo')
⋮
df3.write.format('couchbase').options(…).save(…)

Filesystems

Your home directories are shared on all of our cluster nodes. That means that if you really want to work with a local filesystem (not HDFS) file, you can.


spark.read.text('file://home/me/data/small_input_file.txt')
⋮
df.write.csv('file://home/me/results/small_output')

… as long as you get the permissions set properly on the directories.

Spark 2.4

In Spark 2.4 (released 2018-11-02), some notable new features…

Spark 2.4

Spark 3.0 should release soon.

Delta Lake

A new IO option for Spark: Delta Lake.

In Spark code, it mostly looks like an IO format:

df.write.format('delta').save(output)

Delta Lake

It's mostly Parquet files, but with many features added in the way they're manipulated. Most notably:

  • ACID transactions;
  • versioning (you can read old contents);
  • updating records in-place;
  • unification between batch and streaming;
  • tables have a dual personality: as DeltaTable objects, and as Spark DataFrame objects.

Delta Lake

Let's have a look at the Delta Lake Quickstart.