Data Cleaning with Apache Spark

Spark Schema

.printSchema()

# Import schema

import pyspark.sql.types 
from pyspark.sql.types import *

peopleSchema = StructType([
# Define the name field
StructField('name', StringType(), True), # Define the name field
StructField('age', IntegerType(), True), # Add the age field
StructField('city', StringType(), True) ]) # Add the city field
#if false, means cannot be nullable.

people_df = spark.read.format('csv').load(name='rawdata.csv', 
schema=peopleSchema)

Immutability and Lazy Processing

A component of functional programming Defined once Unable to be directly modied Re-created if reassigned Able to be shared efficiently

# Immutability Example
# Load the CSV file
aa_dfw_df = spark.read.format('csv').options(Header=True).load('AA_DFW_2018.csv.gz')

# Add the airport column using the F.lower() method
aa_dfw_df = aa_dfw_df.withColumn('airport', F.lower(aa_dfw_df['Destination Airport']))

# Drop the Destination Airport column
aa_dfw_df = aa_dfw_df.drop(aa_dfw_df['Destination Airport'])

# Show the DataFrame
aa_dfw_df.show()

Understanding Parquet

Difficulties with CSV files

  • No defined schema （no data type, no column name beyond a header row）

  • Nested data requires special (content containing a comma requires escaping, using the escape character within content requires even further escaping)

  • handling Encoding format limited

  • for spark: slow to parse, cannot be shared during the import process; if no schema is defined, all data must be read before a schema can be inferred, forcing the code to read the file twice.

  • for spark: files cannot be filtered (no 'predicate pushdown', ordering tasks to do the least amount of work, filtering data prior to processing is one of the primary optimizations of predicate pushdown.)

  • for spark: any intermediate use requires redefining schema. Spark process are often multi-step and may utilize an intermediate file representation. These representations allow data to be used later without regenerating the data from source.

The Parquet Format

  • columnar data format

  • supported in spark and other data processing frameworks

  • supports predicate pushdown

  • automatically stores schema information

  • binary file format

Parquet is a compressed columnar data format developed for use in any Hadoop based system. Include: Spark, Hadoop, Apache Impala... Perfect for intermediary or on-disk representation of processed data.

predicate pushdown This means Spark will only process the data necessary to complete the operations you define versus reading the entire dataset.

Working with Parquet

df = spark.read.format('parquet').load('filename.parquet') 
df = spark.read.parquet('filename.parquet')

df.write.format('parquet').save('filename.parquet') 
df.write.parquet('filename.parquet') #mode='overwrite'

Parquet with SQL

parquet - dataframe - table

flight_df = spark.read.parquet('flights.parquet') 
flight_df.createOrReplaceTempView('flights')
short_flights_df = spark.sql('SELECT * FROM flights WHERE flightduration <100')

# Run a SQL query of the average flight duration
avg_duration = spark.sql('SELECT avg(flight_duration) from flights').collect()[0]
print('The average flight time is: %d' % avg_duration)

DataFrame column operations

DataFrames:

  • Made up of rows & columns

  • Immutable

  • Use various transformation operations to modify data

# Return rows where name starts with "M"
voter_df.filter(voter_df.name.like('M%'))
# Return name and position only
voters = voter_df.select('name', 'position')

Common DataFrame transformations

#Filter / Where interchangeable 
voter_df.filter(voter_df.date > '1/1/2019') # or voter_df.where(...) 

#Select
voter_df.select(voter_df.name)

#withColumn creates new column
voter_df.withColumn('year', voter_df.date.year)

#drop
voter_df.drop('unused_column')

Filtering data

  • Remove nulls

  • Remove odd entries

  • Split data from combined sources

  • Negate with ~

voter_df.filter(voter_df['name'].isNotNull()) 
voter_df.filter(voter_df.date.year > 1800) 
voter_df.where(voter_df['_c0'].contains('VOTE')) 
voter_df.where(~ voter_df._c1.isNull())
# Show the distinct VOTER_NAME entries
voter_df.select('VOTER_NAME').distinct().show(40, truncate=False)

# Filter voter_df where the VOTER_NAME is 1-20 characters in length
voter_df = voter_df.filter('length(VOTER_NAME) > 0 and length(VOTER_NAME) < 20')

# Filter out voter_df where the VOTER_NAME contains an underscore
voter_df = voter_df.filter(~ F.col('VOTER_NAME').contains('_'))

# Show the distinct VOTER_NAME entries again
voter_df.select('VOTER_NAME').distinct().show(40, truncate=False)

Column string transformations

#Contained in pyspark.sql.functions
import pyspark.sql.functions as F

#Applied per column as transformation
voter_df.withColumn('upper', F.upper('name')) 

#Can create intermediary columns
voter_df.withColumn('splits', F.split('name', ' '))

#Can cast to other types
voter_df.withColumn('year',voter_df[' c4'].cast(IntegerType()))

ArrayType() column functions

Various utility functions / transformations to interact with ArrayType() .

.size() - returns length of arrayType() column

.getItem() - used to retrieve a specific item at index of list column.

# Add a new column called splits separated on whitespace
voter_df = voter_df.withColumn('splits', F.split(voter_df.VOTER_NAME, '\s+'))

# Create a new column called first_name based on the first item in splits
voter_df = voter_df.withColumn('first_name', voter_df.splits.getItem(0))

# Get the last entry of the splits list and create a column called last_name
voter_df = voter_df.withColumn('last_name', voter_df.splits.getItem(F.size('splits') - 1))

# Drop the splits column
voter_df = voter_df.drop('splits')

# Show the voter_df DataFrame
voter_df.show()

Conditional DataFrame column operations Conditional clauses

Conditional Clauses are:

  • Inline version of if / then / else

  • .when()

  • .otherwise()

df.select(df.Name, df.Age, 
.when(df.Age >= 18, "Adult")
.when(df.Age < 18, "Minor"))

df.select(df.Name, df.Age, 
.when(df.Age >= 18, "Adult")
.otherwise("Minor"))

# Add a column to voter_df for any voter with the title **Councilmember**
voter_df = voter_df.withColumn('random_val',
          when(voter_df.TITLE == 'Councilmember', F.rand()))

# Add a column to voter_df for a voter based on their position
voter_df = voter_df.withColumn('random_val',
    when(voter_df.TITLE == 'Councilmember', F.rand())
    .when(voter_df.TITLE == 'Mayor', 2)
    .otherwise(0))

User defined functions

  • Python method

  • Wrapped via the pyspark.sql.functions.udf method

  • Stored as a variable

  • Called like a normal Spark function

def getFirstAndMiddle(names):
  # Return a space separated string of names
  return ' '.join(names[:-1])

# Define the method as a UDF
udfFirstAndMiddle = F.udf(getFirstAndMiddle, StringType())

# Create a new column using your UDF
voter_df = voter_df.withColumn('first_and_middle_name', udfFirstAndMiddle(voter_df.splits))

# Drop the unecessary columns then show the DataFrame
voter_df = voter_df.drop('first_name')
voter_df = voter_df.drop('splits')
voter_df.show()

Reverse string UDF

#Define a Python method
def reverseString(mystr): 
    return mystr[::-1]

#Wrap the function and store as a variable
udfReverseString = udf(reverseString, StringType())

#Use with Spark
user_df = user_df.withColumn('ReverseName', udfReverseString())

Argument-less example

def sortingCap():
  return random.choice(['G', 'H', 'R', 'S'])
udfSortingCap = udf(sortingCap, StringType())
user_df = user_df.withColumn('Class', udfSortingCap())

Partitioning and lazy processing

Partitioning

  • DataFrames are broken up into partitions

  • Partition size can vary

  • Each partition is handled independently

To check the number of partitions, use the method .rdd.getNumPartitions() on a DataFrame.

Lazy processing

  • Transformations are lazy * .withColumn(...) * .select(...) * .cache()

  • Nothing is actually done until an action is performed * .count() * .write(...) *. show()

  • Transformations can be re-ordered for best performance

  • Sometimes causes unexpected behavior

Adding IDs

Normal ID fields:

  • Common in relational databases

  • Most usually an integer increasing, sequential and unique

  • Not very parallel

Monotonically increasing IDs

pyspark.sql.functions.monotonically_increasing_id()

  • Integer (64-bit), increases in value, unique

  • Not necessarily sequential (gaps exist)

  • Completely parallel

# Select all the unique council voters
voter_df = df.select(df["VOTER NAME"]).distinct()

# Count the rows in voter_df
print("\nThere are %d rows in the voter_df DataFrame.\n" % voter_df.count())

# Add a ROW_ID
voter_df = voter_df.withColumn('ROW_ID', F.monotonically_increasing_id())

# Show the rows with 10 highest IDs in the set
voter_df.orderBy(voter_df.ROW_ID.desc()).show(10)

More ID Tricks

Depending on your needs, you may want to start your IDs at a certain value so there isn't overlap with previous runs of the Spark task. This behavior is similar to how IDs would behave in a relational database. Make sure that the IDs output from a monthly Spark task start at the highest value from the previous month.

Caching

Caching in Spark:

  • Stores DataFrames in memory or on disk

  • Improves speed on later transformations / actions

  • Reduces resource usage

Disadvantages of caching

  • Very large data sets may not t in memory

  • Local disk based caching may not be a performance improvement

  • Cached objects may not be available

Caching tips When developing Spark tasks:

  • Cache only if you need it

  • Try caching DataFrames at various points and determine if your performance improves

  • Cache in memory and fast SSD / NVMe storage

  • Cache to slow local disk if needed

  • Use intermediate files!

  • Stop caching objects when finished

Eviction Policy Least Recently Used (LRU)

Caching is a lazy operation. It requires an action to trigger it. eg.

  • spark.sql("select count(*) from text").show()

  • partitioned_df.count() 

df.cache()    #df.persist() df.persist(storageLevel=pyspark.StorageLevel.MEMORY_AND_DISK)
df.unpersist()

#Determining whether a dataframe is cached
df.is_cached

#storage level useDisk useMemory useOffHeap deserialized replication
df.storageLevel

#Caching a table
df.createOrReplaceTempView('df') 
spark.catalog.cacheTable('df')
spark.catalog.isCached(tableName='df')
spark.catalog.dropTempView('table1')
# List the tables
print("Tables:\n", spark.catalog.listTables())

# Uncaching a table
spark.catalog.uncacheTable('df')
spark.catalog.clearCache()

Improve import performance

Spark clusters

Spark Clusters are made of two types of processes

  • Driver process

  • Worker processes

Import performance

Important parameters:

  • Number of objects (Files, Network locations, etc)

    • More objects better than larger ones

    • Can import via wildcard

airport_df = spark.read.csv('airports-*.txt.gz')

  • General size of objects

    • Spark performs better if objects are of similar size

It's safe to assume the more import objects available, the better the cluster can divvy up the job.

Schemas

A well-dened schema will drastically improve import performance

  • Avoids reading the data multiple times

  • Provides validation on import

How to split objects

  • Use OS utilities / scripts (split, cut, awk)

split -l 10000 -d largefile chunk-

每个文件100000行，字符，名字叫largefile，生成chunk0000开始

  • Use custom scripts

  • Write out to Parquet

df_csv = spark.read.csv('singlelargefile.csv') df_csv.write.parquet('data.parquet') df = spark.read.parquet('data.parquet')

Explaining the Spark execution plan

voter_df = df.select(df['VOTER NAME']).distinct() 
voter_df.explain()

== Physical Plan == *(2) HashAggregate(keys=[VOTER NAME#15], functions=[]) +- Exchange hashpartitioning(VOTER NAME#15, 200)

+- *(1) HashAggregate(keys=[VOTER NAME#15], functions=[]) +- *(1) FileScan csv [VOTER NAME#15] Batched: false, Format: CSV, Location: InMemoryFileIndex[file:/DallasCouncilVotes.csv.gz], PartitionFilters: [], PushedFilters: [], ReadSchema: struct<VOTER NAME:string>

Shuffling

Shuffling refers to moving data around to various workers to complete a task

  • Hides complexity from the user

  • Can be slow to complete

  • Lowers overall throughput

  • Is often necessary, but try to minimize

How to limit shuffling

  • Limit use of .repartition(num_partitions) Use .coalesce(num_partitions) instead

  • Use care when calling .join() Use .broadcast()

  • May not need to limit it

Broadcasting

  • Provides a copy of an object to each worker

  • Prevents undue / excess communication between nodes

  • Can drastically speed up .join() operations

A couple tips:

  • Broadcast the smaller DataFrame. The larger the DataFrame, the more time required to transfer to the worker nodes.

  • On small DataFrames, it may be better skip broadcasting and let Spark figure out any optimization on its own.

  • If you look at the query execution plan, a broadcastHashJoin indicates you've successfully configured broadcasting.

Use the .broadcast(<DataFrame>) method

from pyspark.sql.functions import broadcast 
combined_df = df_1.join(broadcast(df_2))

Cluster Configurations

Configuration options

  • Spark contains many configuration settings

  • These can be modied to match needs

  • Reading configuration settings: spark.conf.get(<configuration name>)

  • Writing configuration settings spark.conf.set(<configuration name>)

  • Configure Spark to use 500 partitions spark.conf.set('spark.sql.shuffle.partitions', 500)

# Name of the Spark application instance
app_name = spark.conf.get('spark.app.name')

# Driver TCP port
driver_tcp_port = spark.conf.get('spark.driver.port')

# Number of join partitions
num_partitions = spark.conf.get('spark.sql.shuffle.partitions')

# Store the number of partitions in variable
before = departures_df.rdd.getNumPartitions()

# Configure Spark to use 500 partitions
spark.conf.set('spark.sql.shuffle.partitions', 500)

# Recreate the DataFrame using the departures data file
departures_df = spark.read.csv('departures.txt.gz').distinct()

Cluster Types

Spark deployment options:

  • Single node (deploying all components on a single system, can be physical/VM/container)

  • Standalone (dedicated machines as the driver and workers)

  • Managed (cluster components are handled by a third party cluster manager)

    • YARN

    • Mesos

    • Kubernetes

Driver

  • Task assignment

  • Result consolidation

  • Shared data access

Tips:

  • Driver node should have double the memory of the worker

  • Fast local storage helpful

Worker

  • Runs actual tasks

  • Ideally has all code, data, and resources for a given task

Recommendations:

  • More worker nodes is often better than larger workers

  • Test to find the balance

  • Fast local storage extremely useful

Data Pipelines

  • Input(s) CSV, JSON, web services, databases

  • Transformations withColumn() , .filter() , .drop()

  • Output(s) CSV, Parquet, database Validation

  • Analysis

Pipeline details

  • Not formally dened in Spark

  • Typically all normal Spark code required for task

schema = StructType([
 StructField('name', StringType(), False),
 StructField('age', StringType(), False)
 ])
 df = spark.read.format('csv').load('datafile').schema(schema) df = df.withColumn('id', monotonically_increasing_id())
 ...
 df.write.parquet('outdata.parquet') 
 df.write.json('outdata.json')
 
 
 
 # Import the data to a DataFrame
departures_df = spark.read.csv('2015-departures.csv.gz', header=True)

# Remove any duration of 0
departures_df = departures_df.filter(~departures_df[3].contains('0'))

# Add an ID column
departures_df = departures_df.withColumn('id', F.monotonically_increasing_id())

# Write the file out to JSON format
departures_df.write.json('output.json')

Data Handling

Parsing

  • Incorrect data * Empty rows * Commented lines * Headers

  • Nested structures * Multiple delimiters

  • Non-regular data * Differing numbers of columns per row

Removing blank lines, headers, and comments

Spark's CSV parser:

  • Automatically removes blank lines

  • Can remove comments using an optional argument

df1 = spark.read.csv('datafile.csv.gz', comment='#')

Handles header fields

  • Defined via argument

  • Ignored if a schema is defined

df1 = spark.read.csv('datafile.csv.gz', header='True')

Count the number of rows beginning with '#'

comment_count = annotations_df.where(col('_c0').startswith('#')).count()

# Split _c0 on the tab character and store the list in a variable
tmp_fields = F.split(annotations_df['_c0'], '\t')

# Create the colcount column on the DataFrame
annotations_df = annotations_df.withColumn('colcount', F.size(tmp_fields))

# Remove any rows containing fewer than 5 fields
annotations_df_filtered = annotations_df.filter(~ (annotations_df["colcount"] < 5))

# Count the number of rows
final_count = annotations_df_filtered.count()
print("Initial count: %d\nFinal count: %d" % (initial_count, final_count))

Automatic column creation

Spark will:

  • Automatically create columns in a DataFrame based on sep argumentdf1 = spark.read.csv('datafile.csv.gz', sep=',')

Defaults to using ,

  • Can still successfully parse if sep is not in string df1 = spark.read.csv('datafile.csv.gz', sep='*')

  • Stores data in column defaulting to _c0

  • Allows you to properly handle nested separators

# Split the content of _c0 on the tab character (aka, '\t')
split_cols = F.split(annotations_df['_c0'], '\t')

# Add the columns folder, filename, width, and height
split_df = annotations_df.withColumn('folder', split_cols.getItem(0))
split_df = split_df.withColumn('filename', split_cols.getItem(1))
split_df = split_df.withColumn('width', split_cols.getItem(2))
split_df = split_df.withColumn('height', split_cols.getItem(3))
def retriever(cols, colcount):
  # Return a list of dog data
  return cols[4:colcount]

# Define the method as a UDF
udfRetriever = F.udf(retriever, ArrayType(StringType()))

# Create a new column using your UDF
split_df = split_df.withColumn('dog_list', udfRetriever(split_df.split_cols, split_df.colcount))

# Remove the original column, split_cols, and the colcount
split_df = split_df.drop('_c0').drop('colcount').drop('split_cols')

Data Validation

Validation is:

  • Verifying that a dataset complies with the expected format

  • Number of rows / columns

  • Data types

  • Complex validation rules

Validating via joins

  • Compares data against known values

  • Easy to find data in a given set

  • Comparatively fast

parsed_df = spark.read.parquet('parsed_data.parquet')
company_df = spark.read.parquet('companies.parquet')
verified_df = parsed_df.join(company_df, parsed_df.company == company_df.company)
# Rename the column in valid_folders_df
valid_folders_df = valid_folders_df.withColumnRenamed('_c0', 'folder')

# Count the number of rows in split_df
split_count = split_df.count()

# Join the DataFrames
joined_df = split_df.join(F.broadcast(valid_folders_df), "folder")

# Compare the number of rows remaining
joined_count = joined_df.count()
print("Before: %d\nAfter: %d" % (split_count, joined_count))

This automatically removes any rows with a company not in the valid_df !

Complex rule validation

Using Spark components to validate logic:

  • Calculations

  • Verifying against external source

  • Likely uses a UDF to modify / verify the DataFrame

# Determine the row counts for each DataFrame
split_count = split_df.count()
joined_count = joined_df.count()

# Create a DataFrame containing the invalid rows
invalid_df = split_df.join(F.broadcast(joined_df), 'folder', 'left_anti')

# Validate the count of the new DataFrame is as expected
invalid_count = invalid_df.count()
print(" split_df:\t%d\n joined_df:\t%d\n invalid_df: \t%d" % (split_count, joined_count, invalid_count))

# Determine the number of distinct folder columns removed
invalid_folder_count = invalid_df.select('folder').distinct().count()
print("%d distinct invalid folders found" % invalid_folder_count)
# Select the dog details and show 10 untruncated rows
print(joined_df.select('dog_list').show(truncate=False))

# Define a schema type for the details in the dog list
DogType = StructType([
	StructField("breed", StringType(), False),
    StructField("start_x", IntegerType(), False),
    StructField("start_y", IntegerType(), False),
    StructField("end_x", IntegerType(), False),
    StructField("end_y", IntegerType(), False)
])
# Create a function to return the number and type of dogs as a tuple
def dogParse(doglist):
  dogs = []
  for dog in doglist:
    (breed, start_x, start_y, end_x, end_y) = dog.split(',')
    dogs.append((breed, int(start_x), int(start_y), int(end_x), int(end_y)))
  return dogs

# Create a UDF
udfDogParse = F.udf(dogParse, ArrayType(DogType))

# Use the UDF to list of dogs and drop the old column
joined_df = joined_df.withColumn('dogs', udfDogParse('dog_list')).drop('dog_list')

# Show the number of dogs in the first 10 rows
joined_df.select(F.size('dogs')).show(10)
# Define a UDF to determine the number of pixels per image
def dogPixelCount(doglist):
  totalpixels = 0
  for dog in doglist:
    totalpixels += (dog[3] - dog[1]) * (dog[4] - dog[2])
  return totalpixels

# Define a UDF for the pixel count
udfDogPixelCount = F.udf(dogPixelCount, IntegerType())
joined_df = joined_df.withColumn('dog_pixels', udfDogPixelCount('dogs'))

# Create a column representing the percentage of pixels
joined_df = joined_df.withColumn('dog_percent', (joined_df.dog_pixels / (joined_df.width * joined_df.height)) * 100)

# Show the first 10 annotations with more than 60% dog
joined_df.where('dog_percent > 60').show(10)
PreviousIntroduction to PysparkNextFeature Engineering with Pyspark

Last updated