Spark Note

 

pyspark-note

Concept

The records/items/elemets are stored in RDD(s).

Each RDD composists of Partitions; each Partition contains equal number of items/elements.

Where code runs

Source: https://spark.apache.org/docs/latest/cluster-overview.html

Most Python code runs in driver (in our local PC), except for code passed to RDD transformations.

  • Transformations run at executors (in workers),
  • actions run at executors and driver.

RDD Basic operation

RDD Programming Guide

==> https://spark.apache.org/docs/latest/rdd-programming-guide.html

Transformations

.map() v.s. .mapPartitions() v.s. .mapPartitionsWithIndex()

Spark的map，mapPartitions，mapPartitionsWithIndex詳解

==> https://blog.csdn.net/QQ1131221088/article/details/104051087

1. .map()

Return a new distributed dataset formed by passing each element of the source through a function func.

rdd_2 = sc.parallelize(range(10), 4)

new_rdd_2 = rdd_2.map(lambda x: str(x))
print('> .map() =\n', new_rdd_2.glom().collect())
print()

#################################################
from pyspark import TaskContext

result = rdd_2.map(lambda x :x+TaskContext().partitionId())
print('> Original RDD, rdd_2.glom().collect() =\n', rdd_2.glom().collect())
print()
print('> .map() with TaskContext().partitionId() =\n', result.glom().collect())

Output:

> .map() =
 [['0', '1'], ['2', '3', '4'], ['5', '6'], ['7', '8', '9']]

> Original RDD, rdd_2.glom().collect() =
 [[0, 1], [2, 3, 4], [5, 6], [7, 8, 9]]

> .map() with TaskContext().partitionId() =
 [[0, 1], [3, 4, 5], [7, 8], [10, 11, 12]]

2. .mapPartitions()

Similar to .map(), but runs separately on each partition (block) of the RDD, so func must be of type Iterator<T> => Iterator<U> when running on an RDD of type T.

==> Divide-and-Conquer algorithm => master node divides the RDD into partitions, and distributes the partitions to workers, workers apply the same function to its partition. Then master node gets back (i.e. conquer) the processed resuts from all workers.

(1)

rdd_2 = sc.parallelize([1,2,3,4,5,'a','b','c','d','e'], 4)

def func(itemsIteratorInPartition):
    # apply this `func` to each partition (=the whole partition) of the RDD
    yield str(itemsIteratorInPartition)
    
rdd_func = rdd_2.mapPartitions(func)
print('rdd_2.mapPartitions(func) =\n', rdd_func.glom().collect())
print()

Output:

rdd_2.mapPartitions(func) =
 [['<itertools.chain object at 0x7ff8094580d0>'], ['<itertools.chain object at 0x7ff8094580d0>'], ['<itertools.chain object at 0x7ff8094580d0>'], ['<itertools.chain object at 0x7ff8094580d0>']]

(2)

def func_2(itemsIteratorInPartition):
    # you loop through each item in each partition of the RDD
    # = just apply this `func_2` to each item in each partition
    for item in itemsIteratorInPartition: 
        yield str(item)

rdd_func_2 = rdd_2.mapPartitions(func_2)
print('rdd_2.mapPartitions(func_2) =\n', rdd_func_2.glom().collect())

Output:

rdd_2.mapPartitions(func_2) =
 [['1', '2'], ['3', '4'], ['5', 'a'], ['b', 'c', 'd', 'e']]

3. .mapPartitionsWithIndex()

Similar to mapPartitions, but also provides func with an integer value representing the index of the partition, so func must be of type (Int, Iterator<T>) => Iterator<U> when running on an RDD of type T.

By using mapParitionsWithIndex you could output new elements which have their partition in it, then when you reduce you will know which partition you are handling the elements from.

==> https://stackoverflow.com/questions/31281225/find-out-the-partition-no-id

rdd_3 = sc.parallelize(range(10), 4)

# mapPartitionsWithIndex
def func(partitionIndex, itemsIteratorInPartition): 
    # apply this `func` to each partition (=the whole partition) of the RDD
    yield (partitionIndex, sum(itemsIteratorInPartition))
new_rdd_3 = rdd_3.mapPartitionsWithIndex(func)

# glom() flattens elements on the same partition
print('> rdd_3.glom().collect() =', rdd_3.glom().collect())
print('> new_rdd_3.glom().collect() =', new_rdd_3.glom().collect())


################################################################################

def func_2(partitionIndex, itemsIteratorInPartition):
    # you loop through each item in each partition of the RDD
    # = just apply this `func_2` to each item in each partition
    for item in itemsIteratorInPartition: 
        yield str(item+partitionIndex)
new_2_rdd_3 = rdd_3.mapPartitionsWithIndex(func_2)

# glom() flattens elements on the same partition
print()
print('>> new_2_rdd_3.glom().collect() =', new_2_rdd_3.glom().collect())

Output:

> rdd_3.glom().collect() = [[0, 1], [2, 3, 4], [5, 6], [7, 8, 9]]
> new_rdd_3.glom().collect() = [[(0, 1)], [(1, 9)], [(2, 11)], [(3, 24)]]

>> new_2_rdd_3.glom().collect() = [['0', '1'], ['3', '4', '5'], ['7', '8'], ['10', '11', '12']]

.map() v.s. .flatmap()

print(sc.version, '\n')

py_list = [str(x) for x in range(5)]
rdd = sc.parallelize(py_list)

# map
new_rdd = rdd.map(lambda item: item+'xx')
print('.map() =\n', new_rdd.collect())
print()

# flatmap
# same as .map(), but flatten the results before returns
# i.e. remove all `list of list`/`nested list`
new_rdd = rdd.flatMap(lambda item: item+'xx')
print('.flatmap() =\n', new_rdd.collect())

Output:

3.1.2

.map() =
['0xx', '1xx', '2xx', '3xx', '4xx']

.flatmap() =
['0', 'x', 'x', '1', 'x', 'x', '2', 'x', 'x', '3', 'x', 'x', '4', 'x', 'x']

.foreach() v.s. .map()

==> See .foreach() / .map() v.s. .foreach()

Actions

Below example shows there are 100 items/elements in this RDD, and this RDD is partitioned into 4 partitions (or items are grouped in 4 partitions).

sc.parallelize()

Store python list [0,1,...,99] as RDD in Spark. This dataset is not loaded in memory. It is merely a pointer to the Python py_list.

# Stores python list `[0,1,...,99]` as RDD in Spark
## This dataset is not loaded in memory
## It is merely a pointer to the Python `py_list` 
py_list = range(100)
rdd = sc.parallelize(py_list)
print(rdd)

Output:

PythonRDD[11] at RDD at PythonRDD.scala:53

.count()

Returns the number of items in this RDD

#.count()
# Shows the number of items in this RDD
print('rdd.count()=', rdd.count())

Output:

100

.collect()

Returns all the items in this RDD as python list

#.collect()
# Returns all the items in this RDD as python list
print('rdd.collect()=', rdd.collect())

Output:

[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99]

.glom().collect()

Returns the content of each partitions as nested list / list of list

# Returns the content of each partitions as `nested list` / `list of list`
rdd.glom().collect()

Output:

[
 [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24], 
 [25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49], 
 [50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74], 
 [75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99]
]

.getNumPartitions()

Returns number of partitions in this RDD

#.getNumPartitions()
# Gets number of partitions in this RDD
print('rdd.getNumPartitions()=', rdd.getNumPartitions())

Output:

4

.foreach()

Just executes inside function for each data element in the RDD, but return NOTHING.

.map() v.s. .foreach()

Short answer

  • .map():

    1. is for transforming one RDD into another, then return the transformed.
    2. Return a new RDD by applying a function to each element of this RDD.

  • .foreach(): is for applying an operation/function on all elements of this RDD.

    Note: RDD = 1 collection of elements

==> is-there-a-difference-between-foreach-and-map

Long answer:

The important difference between them is that map accumulates all of the results into a collection, whereas foreach returns nothing. map is usually used when you want to transform a collection of elements with a function, whereas foreach simply executes an action for each element.

  • In short, foreach is for applying an operation on each element of a collection of elements, whereas map is for transforming one collection into another.

  • foreach works with a single collection of elements. This is the input collection.

  • map works with two collections of elements: the input collection and the output collection.

.reduce()

rdd = sc.parallelize([('a',7),('a',2),('b',2)])
rdd.reduce(lambda a, b: a + b) #Merge the rdd values
('a',7,'a',2,'b',2)

Spark recomputes transformations

Transformed RDD is thrown away from memory after execution. If afterward transformations/actions need it, PySpark recompiles it.

 

Image. RDD Without vs With .cache() / .persist()

P.S. Solution: .cache()/.persist() the transformed RDD

A = sc.parallelize(range(1, 1000)) 
t = 100
B = A.filter(lambda x: x*x < t)
print('B.collect()=', B.collect())  # B.collect()= [1, 2, 3, 4, 5, 6, 7, 8, 9]
## Here: B finishes execution and is thrown away from memory

t = 200
C = B.filter(lambda x: x*x < t) # C needs B again, so recomputes B, but t=200 not =100
                                # So, 
                                # B = A.filter(lambda x: x*x < 200)
                                # C = B.filter(lambda x: x*x < 200)
print('C.collect()=', C.collect())  # C.collect()= [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14]
print('C.count()=', C.count())      # C.count()= 14

.cache()/.persist()

A = sc.parallelize(range(1, 1000)) 
t = 100
B = A.filter(lambda x: x*x < t)
print('B.collect()=', B.collect())  # B.collect()= [1, 2, 3, 4, 5, 6, 7, 8, 9]

# save this RDD in memory/disk
B.cache()
# B.persist()

## Here: B is in memory

t = 200
C = B.filter(lambda x: x*x < t) # C needs B again, memory stores B, NO need to recompute B
                                # So, 
                                # B = previous B
                                # C = B.filter(lambda x: x*x < 200)
print('C.collect()=', C.collect())  # C.collect()= [1, 2, 3, 4, 5, 6, 7, 8, 9]
print('C.count()=', C.count())      # C.count()= 9

RDD - Closure

https://mycupoftea00.medium.com/understanding-closure-in-spark-af6f280eebf9

https://spark.apache.org/docs/latest/rdd-programming-guide.html#understanding-closures-

Closure example

Incorrect way - global variable as counter

Q: Why printed counter is 0?

Ans: Because

  1. each executor (i.e. worker node) just applies increment_counter() func on its own copy of counter.
  2. Also, .foreach() returns nothing
counter = 0
rdd = sc.parallelize(range(10))
print('rdd.collect() =', rdd.collect())

# Wrong: Don't do this!!
def increment_counter(x):
    global counter
    counter += x

rdd.foreach(increment_counter)
print('counter =', counter) # just print out `counter` from your driver program, not from Spark

Output:

rdd.collect() = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
counter = 0

Correct way (1) - rdd.sum()

Correct way to do the above operation:

  1. The .sum() action is executed in Spark executor(s)
  2. .sum() returns the sum value
print('rdd.sum() =', rdd.sum())

Output:

rdd.sum() = 45

Correct way (2) - .accumulator()

Use .accumulator() can also solve the issue.

rdd = sc.parallelize(range(10))
accum = sc.accumulator(0)

def g(x):
    global accum
    accum += x

a = rdd.foreach(g)

print(accum.value)
45

Note

Update in transformations may be applied more than once if tasks or job stages are re-executed.

rdd = sc.parallelize(range(10))
accum = sc.accumulator(0)

def g(x):
    global accum
    accum += x
    return x * x

a = rdd.map(g)
print(type(accum))
print(accum.value) # 0, because no action presents, `accum` is not immediately computed (= laziness/lazy execution)
# print(a.reduce(lambda x, y: x+y))
a.cache()
tmp = a.count()
print(accum.value) # 45
print(rdd.reduce(lambda x, y: x+y)) # 45

tmp = a.count()
print(accum.value) # 45
print(rdd.reduce(lambda x, y: x+y)) # 45

Output:

<class 'pyspark.accumulators.Accumulator'>
0   #why it is 0? because of "lazy execution", if no actions present, "accum" is not compiled immediately
45
45
45
45

Accumulator

https://spark.apache.org/docs/latest/rdd-programming-guide.html#accumulators

Note from lecture note: Suggestion: Avoid using accumulators whenever possible. Use reduce() instead.

Deal with JSON data

** Most useful ==> https://zhuanlan.zhihu.com/p/267353998

Read JSON string to

Read JSON to spark Dataframe first

https://sparkbyexamples.com/pyspark/pyspark-maptype-dict-examples/

Steps,

  1. JSON from API
  2. get list of dict
  3. pySpark dataframe with map type
  4. access PySpark MapType Elements

Details

  1. # The nested json / list of dictionary
data_json = [
    ('James', {'hair': 'black', 'eye': 'brown'}),
    ('Michael', {'hair': 'brown', 'eye': None}),
    ('Robert', {'hair': 'red', 'eye': 'black'}),
    ('Washington', {'hair': 'grey', 'eye': 'grey'}),
    ('Jefferson', {'hair': 'brown', 'eye': ''})
]
df = spark.createDataFrame(data=data_json)
df.printSchema()

    Output:

    root
|-- Name: string (nullable = true)
|-- properties: map (nullable = true)
|    |-- key: string
|    |-- value: string (valueContainsNull = true)

+----------+-----------------------------+
|Name      |properties                   |
+----------+-----------------------------+
|James     |[eye -> brown, hair -> black]|
|Michael   |[eye ->, hair -> brown]      |
|Robert    |[eye -> black, hair -> red]  |
|Washington|[eye -> grey, hair -> grey]  |
|Jefferson |[eye -> , hair -> brown]     |
+----------+-----------------------------+

  2. Access the elements in map datatype

    Method (1):

    df3 = df.rdd.map(lambda x: \
                (x.name, x.properties["hair"], x.properties["eye"])) \
                .toDF(["name", "hair", "eye"])
df3.printSchema()
df3.show()

    OR

    Method (2):

    df.withColumn("hair", df.properties.getItem("hair")) \
    .withColumn("eye", df.properties.getItem("eye")) \
    .drop("properties") \
    .show()

df.withColumn("hair", df.properties["hair"]) \
    .withColumn("eye", df.properties["eye"]) \
    .drop("properties") \
    .show()

    Output:

    root
|-- name: string (nullable = true)
|-- hair: string (nullable = true)
|-- eye: string (nullable = true)

+----------+-----+-----+
|      name| hair|  eye|
+----------+-----+-----+
|     James|black|brown|
|   Michael|brown| null|
|    Robert|  red|black|
|Washington| grey| grey|
| Jefferson|brown|     |
+----------+-----+-----+

Spark Dataframe

Create sparkdf by reading .csv

customer.csv:

CUSTKEY,NAME,ADDRESS,NATIONKEY,PHONE,ACCTBAL,MKTSEGMENT,COMMENT
1,Customer#000000001,IVhzIApeRb ot,c,E,15,25-989-741-2988,711.56,BUILDING,to the even, regular platelets. regular, ironic epitaphs nag e,
2,Customer#000000002,XSTf4,NCwDVaWNe6tEgvwfmRchLXak,13,23-768-687-3665,121.65,AUTOMOBILE,l accounts. blithely ironic theodolites integrate boldly: caref,
3,Customer#000000003,MG9kdTD2WBHm,1,11-719-748-3364,7498.12,AUTOMOBILE, deposits eat slyly ironic, even instructions. express foxes detect slyly. blithely even accounts abov,
4,Customer#000000004,XxVSJsLAGtn,4,14-128-190-5944,2866.83,MACHINERY, requests. final, regular ideas sleep final accou,
...

orders.csv:

ORDERKEY,CUSTKEY,ORDERSTATUS,TOTALPRICE,ORDERDATE,ORDERPRIORITY,CLERK,SHIPPRIORITY,COMMENT
1,370,O,172799.49,1996-01-02,5-LOW,Clerk#000000951,0,nstructions sleep furiously among ,
2,781,O,38426.09,1996-12-01,1-URGENT,Clerk#000000880,0, foxes. pending accounts at the pending, silent asymptot,
3,1234,F,205654.30,1993-10-14,5-LOW,Clerk#000000955,0,sly final accounts boost. carefully regular ideas cajole carefully. depos,
4,1369,O,56000.91,1995-10-11,5-LOW,Clerk#000000124,0,sits. slyly regular warthogs cajole. regular, regular theodolites acro,
5,445,F,105367.67,1994-07-30,5-LOW,Clerk#000000925,0,quickly. bold deposits sleep slyly. packages use slyly,
...
dfCustomer = spark.read.csv('customer.csv', header=True, inferSchema=True)
dfOrders = spark.read.csv('orders.csv', header=True, inferSchema=True)

.printSchema() in df

dfCustomer.printSchema()
print(dfCustomer.count())

dfOrders.printSchema()
print(dfOrders.count())

Output:

root
 |-- CUSTKEY: integer (nullable = true)
 |-- NAME: string (nullable = true)
 |-- ADDRESS: string (nullable = true)
 |-- NATIONKEY: string (nullable = true)
 |-- PHONE: string (nullable = true)
 |-- ACCTBAL: string (nullable = true)
 |-- MKTSEGMENT: string (nullable = true)
 |-- COMMENT: string (nullable = true)
1500

root
 |-- ORDERKEY: integer (nullable = true)
 |-- CUSTKEY: integer (nullable = true)
 |-- ORDERSTATUS: string (nullable = true)
 |-- TOTALPRICE: double (nullable = true)
 |-- ORDERDATE: string (nullable = true)
 |-- ORDERPRIORITY: string (nullable = true)
 |-- CLERK: string (nullable = true)
 |-- SHIPPRIORITY: integer (nullable = true)
 |-- COMMENT: string (nullable = true)
 15000

.groupBy().count()

Find count of orders of each customer CUSTKEY has:

dfOrders_groupby = dfOrders.groupBy('CUSTKEY').count()
dfOrders_groupby.toPandas()

Output:

	CUSTKEY	count
0	463	20
1	1342	20
2	496	18
3	148	15
4	1088	7
...	...	...
995	401	12
996	517	25
997	422	12
998	89	7
999	1138	23
1000 rows × 2 columns

df.createOrReplaceTempView("sql_table"), allows to run SQL queries once register df as temporary tables

dfOrders_groupby.createOrReplaceTempView("sql_table")

# Can run SQL query on it
df = spark.sql("SELECT customer.CUSTKEY, orders.count FROM customer left outer join orders on customer.CUSTKEY = orders.CUSTKEY")
df.toPandas()

Output:

    CUSTKEY	count
0	1	9.0
1	2	10.0
2	3	NaN
3	4	31.0
4	5	9.0
...	...	...
1495	1496	9.0
1496	1497	NaN
1497	1498	20.0
1498	1499	21.0
1499	1500	NaN
1500 rows × 2 columns

.join()/spark.sql() dataframes

.join()

# join 2 df by `CUSTKEY`
joined_df = dfCustomer.join(dfOrders, on='CUSTKEY', how='leftouter')
df2 = joined_df.select('CUSTKEY', 'ORDERKEY').sort(asc('CUSTKEY'),desc('ORDERKEY')) #ascending by 'CUSTKEY', descending by 'ORDERKET'
df2.toPandas() #view

how: str, optional default inner. Must be one of: inner, cross, outer, full, fullouter, full_outer, left, leftouter, left_outer, right, rightouter, right_outer, semi, leftsemi, left_semi, anti, leftanti and left_anti.

spark.sql() + df.createOrReplaceTempView("sql_table")

dfOrders.createOrReplaceTempView("orders")
dfCustomer.createOrReplaceTempView("customer")

# Can run SQL query on it
df2 = spark.sql("SELECT customer.CUSTKEY, orders.ORDERKEY FROM customer left outer join orders on customer.CUSTKEY = orders.CUSTKEY")
df2.toPandas()

Output:

    CUSTKEY	ORDERKEY
0	1	53283.0
1	1	52263.0
2	1	43879.0
3	1	36422.0
4	1	34019.0
...	...	...
15495	1499	7301.0
15496	1499	3523.0
15497	1499	1472.0
15498	1499	1252.0
15499	1500	NaN
15500 rows × 2 columns

df1.union(df2) concat 2 dataframes

The dataframes may need to have identical columns, in which case you can use withColumn() to create normal_1 and normal_2

df_concat = df_1.union(df_2)

UDF df.withColumn, user defined function

Reference

Graph, edge, vertice, Graphframe

Credit to link

GraphFrame(v, e), Create GraphFrame

# Vertics DataFrame
v = spark.createDataFrame([
  ("a", "Alice", 34),
  ("b", "Bob", 36),
  ("c", "Charlie", 37),
  ("d", "David", 29),
  ("e", "Esther", 32),
  ("f", "Fanny", 38),
  ("g", "Gabby", 60)
], ["id", "name", "age"])

# Edges DataFrame
e = spark.createDataFrame([
  ("a", "b", "friend"),
  ("b", "c", "follow"), # b and c follow each other
  ("c", "b", "follow"), #
  ("f", "c", "follow"),
  ("e", "f", "follow"),
  ("e", "d", "friend"),
  ("d", "a", "friend"),
  ("a", "e", "friend"),
  ("g", "e", "follow")
], ["src", "dst", "relationship"])

# Create a GraphFrame
g = GraphFrame(v, e)

g.vertices.show()
g.edges.show()
# Vertics DataFrame
v = spark.createDataFrame([
    ("a", "Alice", 34),
    ("b", "Bob", 36),
    ("c", "Charlie", 37),
    ("d", "David", 29),
    ("e", "Esther", 32),
    ("f", "Fanny", 38),
    ("g", "Gabby", 60)
], ["id", "name", "age"])
# Edges DataFrame
e = spark.createDataFrame([
    ("a", "b", "follow"),
    ("c", "a", "friend"),
    ("b", "c", "follow"),
    ("d", "a", "follow"),
    ("f", "c", "follow"),
    ("f", "d", "follow"),
    ("f", "b", "follow"),
    ("c", "d", "follow"),
    ("g", "a", "friend"),
    ("g", "d", "friend"),
    ("g", "c", "friend"),
    ("e", "a", "follow"),
    ("e", "d", "follow")
], ["src", "dst", "relationship"])

# Create a GraphFrame
g = GraphFrame(v, e)

Explore GraphFrame

Credit to link

g.triplets.show() # display all
g.vertices.show() # display vertices
g.edges.show()    # display edges
g.degrees.show()
g.inDegrees.show()
g.outDegrees.show()

Filter, g.filterVerices() g.filterEdges()

Returns GraphFrame, not DataFrame.

Credit to link

g.filterVerices("columnName > 30")
g.filterEdges("columnName = 30")
g.dropIsolatedVertices() #Drop isolated vertices (users) which are not contained in any edges (relationships).
                         #Vertices without incoming / outgoing edges

.find("(a)-[e]->(b)"), Motif finding

Find the edges e from vertex a to vertex b.

P.S. .find() returns sparkDF DataFrame.

Credit to link

g.find("(a)-[]->(b);(b)-[]->(a)").filter("a.id < b.id") # A and B follow/friend each other; 
                                                        # .filter() out "B follows/friends back A" rows, 
                                                        # just keeps "A follows/friends B" rows
g.find("(a)-[]->(b); !(b)-[]->(a)").filter("a.id < b.id") # jsut A follows B, B not follows A
g.find("!()-[]->(a)") # find vertices without incoming edges
g.find("(a)-[e]->(b)").filter("e.relationship = 'follow'") # find A follows B,

Subgraphs

Credit to msbd5003.

# Build subgraph based on conditions, i.e. subgraph contains (v,e)
# Select subgraph of users older than 30, and relationships of type "friend".
# Drop isolated vertices (users) which are not contained in any edges (relationships).
g1 = g.filterVertices("age > 30").filterEdges("relationship = 'friend'")\
      .dropIsolatedVertices()
g1.vertices.show()
g1.edges.show()

Output:

+---+------+---+
| id|  name|age|
+---+------+---+
|  e|Esther| 32|
|  b|   Bob| 36|
|  a| Alice| 34|
+---+------+---+
+---+---+------------+
|src|dst|relationship|
+---+---+------------+
|  a|  e|      friend|
|  a|  b|      friend|
+---+---+------------+

spark-install-macos

Run the following in macOS terminal,

How to start Jupyter Notebook with Spark + GraphFrames

Start it locally

  1. Modify the PATH variables,

    $ nano ~/.bashrc

  2. Add the following lines in ~/.bashrc, so spark and jupyter notebook can be launched at the same time.

    # Setting PATH for Spark 3.1.2
export SPARK_HOME=/usr/local/Cellar/apache-spark/3.1.2/libexec
export PATH="$SPARK_HOME/bin/:$PATH"
export PYSPARK_DRIVER_PYTHON="jupyter"
export PYSPARK_DRIVER_PYTHON_OPTS="notebook"

    In terminal,

  3. Update $PATH variable,

    $ source ~/.bashrc

  4. Ensure graphframes-0.8.1-spark3.0-s_2.12.jar presents in the /Users/<USER_NAME>/.ivy2/jars folder:

    Start the pyspark with graphframes in terminal,

    ==> Needs 2 flags, --packages and --jars

    ==> Ensure the graphframes package name is same as graphframes-0.8.1-spark3.0-s_2.12.jar in folder.

    => Deal with error: java.lang.ClassNotFoundException: org.graphframes.GraphFramePythonAPI

    => https://blog.csdn.net/qq_42166929/article/details/105983616

    $ pyspark --packages graphframes:graphframes:0.8.1-spark3.0-s_2.12 --jars graphframes-0.8.1-spark3.0-s_2.12.jar

    Terminal output:

    Jupyter Notebook:

Start in Google Colab

[Still need to investigate]

https://github.com/cenzwong/tech/tree/master/Note/Spark#graphframe

Use MongoDB in Spark

Terminal:

pyspark --packages graphframes:graphframes:0.8.1-spark3.0-s_2.12 --jars graphframes-0.8.1-spark3.0-s_2.12.jar \
        --conf "spark.mongodb.input.uri=mongodb://127.0.0.1/test.application_test?readPreference=primaryPreferred" \
        --conf "spark.mongodb.output.uri=mongodb://127.0.0.1/test.application_test" \
        --packages org.mongodb.spark:mongo-spark-connector_2.12:3.0.1

Notebook:

from pymongo import MongoClient 

def _saveDfToMongoDB(sparkDF, mongoCollection):
    sparkDF.cache()
    print(f"Storing {sparkDF.count()} {mongoCollection} to db")

    start = datetime.now()

    sparkDF.write.format("mongo") \
        .mode("append") \
        .option("database", "msbd5003") \
        .option("collection", mongoCollection) \
        .save()

    end = datetime.now()
    spent = (end - start).total_seconds()
    print(f"Stored, time used: {spent} s")

df = spark.read.json("file path")
_saveDfToMongoDB(df, "mongodb collection name")

Test Spark in Jupyter Notebook

Inside Jupyter Notebook:

==> Reference: https://medium.com/@roshinijohri/spark-with-jupyter-notebook-on-macos-2-0-0-and-higher-c61b971b5007

Cell 1:

print(sc.version)

Output 1:

3.1.2

OR

from pyspark import SparkContext
from pyspark.sql import SparkSession
from pyspark.sql.functions import *
from graphframes import *

spark = SparkSession.builder.master("local[*]").getOrCreate()
sc = spark.sparkContext

sc

Output:

SparkSession - hive
SparkContext

Spark UI
Version
v3.1.2
Master
local[*]
AppName
PySparkShell

Cell 2:

path_to_the_jar_file = 'graphframes-0.8.1-spark3.0-s_2.12.jar'
sc.addPyFile(path_to_the_jar_file)

Cell 3:

from pyspark.sql import SQLContext 
sqlContext = SQLContext(sc) 
v = sqlContext.createDataFrame([("a", ),("b", ),], ["id", ]) 
e = sqlContext.createDataFrame([("a", "b"),], ["src", "dst"]) 

from graphframes import * 
g = GraphFrame(v, e) 
g.inDegrees.show()

Output 3:

+---+--------+
| id|inDegree|
+---+--------+
|  b|       1|
+---+--------+

Cell 4:

import pyspark
from pyspark.sql.session import SparkSession
spark = SparkSession.builder.appName("spark test").getOrCreate()
columns = ['id', 'dogs', 'cats']
vals = [
     (1, 2, 0),
     (2, 0, 1)
]
# create DataFrame
df = spark.createDataFrame(vals, columns)
df.show()

Output 4:

+---+----+----+
| id|dogs|cats|
+---+----+----+
|  1|   2|   0|
|  2|   0|   1|
+---+----+----+

Reference

  1. Official pyspark example: https://github.com/spark-examples/pyspark-examples


ntroduction to Spark

Background

Hadoop

First open-source project for distributed storage, resource management and computing. Consist of 3 components:

  1. HDFS -> Distributed Storage
  2. YARN -> Resource Management
  3. Map Reduce -> Computing in parallel process (mostly written in Java or can also written in Python)
  • Hadoop 1.0: consists of HDFS and Map Reduce (taking care of resource management)
  • Hadoop 2.0: inclusive of YARN
  • Hadoop 3.0: Improved storage/HDFS layer

Spark start to emerge and dominate over MapReduce in market, for a few reasons:

  1. Spark is In-Memory(RAM) processing framework, whereas MapReduce process in HDD/SSD, thus Spark has better performance than MapReduce.
  2. Hadoop MapReduce performance is related to HDD/SSD space. Say working on 50 hadoop node clusters (50 servers connected in parallel), if the hard disks space of those servers getting small, the CPU utilization will be affected and computing performance will drop, unless more and more clusters are continuously connected in parallel (horizontal scaling up).
  3. Since Hadoop makes use of servers for both computing and storage, the servers cannot be shut down, otherwise there will be data loss. So the deployment on cloud is expensive, as the node clusters must be running always. A better choice is to separate computing and storage.
  4. Memory card becomes cheaper nowadays. Spark can run in standalong mode, then we can have separate storage, so we don't have to integrate it with Hadoop.

Spark support four langurages:

  • Scala (Most widely used since Spark itself is written in Scala, better integration) -> spark-shell
  • Java (2nd most popular lang for data engineering) -> No shell, no CLI
  • Python (Mostly for data science application) -> pyspark
  • R (less used) -> sparkR

Definition

Spark is an unified computing engine for parallel data processing on clusters. It supports batch processing, streaming processing, ML and SQL queries.

Low Level APIs of Spark

  1. RDD - Resilient Dsitributed Datasets (it's like a row of data/record)
  2. Broadcast (less used)
  3. Accumulator (less used)

High Level APIs of Spark - Structured APIs

  1. Dataframe - no fixed schema
  2. SQL - SQL statement for dataframe/dataset control (data define and manipulation)
  3. Dataset - has fixed schema

On top of High Level APIs, it supports:

  1. Structured Streaming.
  2. Advanced Analytics. (for ML use)
  3. Other Libraries.

Versions

  • 2014 -> 1.0 (1.6) -> RDD
  • 2016 -> 2.0 (2.1, 2.4) -> Dataframe, Dataset
  • 2020 -> 3.0 -> Faster speed and more ML libraries

Driver vs Executor

  • Spark Driver is the central coordinator and it communicates with all the Workers. It controls the flow of program.
  • Each Worker node consists of one or more Executor(s) who are responsible for running the Task. Executors register themselves with Driver. The Driver has all the information about the Executors at all the time.
  • A Spark Application consists of a Driver Program and a group of Executors on the cluster.
  • The Driver is a process that executes the main program of your Spark application and creates the SparkContext that coordinates the execution of jobs. SparkSession is created within driver node.
  • The executors are processes running on the worker nodes of the cluster which are responsible for executing the tasks the driver process has assigned to them.
  • The cluster manager (such as Mesos or YARN) is responsible for the allocation of physical resources to Spark Applications SparkSession is created within driver node.

Start Spark in Command Line

  • "spark-shell" is the command to start scala spark shell
  • "pyspark" is the command to start python spark shell
  • "sparkR" is the command to start R spark shell
  • Java has no shell, no command line interface

Spark Session vs Spark Context

  • Since earlier versions of Spark or Pyspark, SparkContext (JavaSparkContext for Java) is an entry point to Spark programming with RDD and to connect to Spark Cluster
  • Since Spark 2.0 SparkSession has been introduced and became an entry point to start programming with DataFrame and Dataset.

Comments

Post a Comment

Popular posts from this blog

Flutter for Single-Page Scrollable Websites with Navigator 2.0

Image
  Single page scrollable websites are everywhere. In this website design pattern, all the content is presented in one long-scrolling layout that contains multiple sections. Visitors can scroll or jump to a section by clicking buttons of a sticky menu. This pattern is a good fit for small content such as brochure websites, software library documentation, portfolios, and landing pages that are used to convert users [ 1 ]. Designers also love this pattern because it is simple, clean, and enables cool scroll animations. Jetbrains Mono single-page scrollable website Hive docs multi-page scrollable website In this series of articles, we will explore how to build a  single-page scrollable website  using Flutter. We will benefit from the Navigator 2.0 API to provide a good navigation experience to the users. You can find the source code of this series’ sample apps on my  Github page . The implementation details of the sample apps will be explained in the following articles: ...
Read more

A Data Science Portfolio is More Valuable than a Resume

Image
A Data Science Portfolio is More Valuable than a Resume Benjamin Obi Tayo Ph.D. Following Nov 20  · 5 min read D ata Science is a practical field. Hands-on skills are very important especially when you are interested in working outside academia as a practicing data scientist. In academia, you need more theoretical and research skills. While in-depth knowledge in the theoretical foundations of data science is important, as a practicing data scientist, hands-on experience is very crucial and one way to showcase you hands-on skills is via building a very great portfolio. Companies interested in hiring you would definitely be asking you for a portfolio, as it gives evidence of your strengths in fundamental data science concepts. While a data science resume is important, a data science portfolio is more valuable than a resume. This article will discuss 3 important platforms that can be used for portfolio buildi...
Read more

Better File Storage in Oracle Cloud

Image
Use Oracle Application Express with Oracle Cloud Infrastructure REST APIs to upload, download, and manage big files in Oracle Cloud. By Adrian Png November 27, 2019 Oracle Cloud Infrastructure  provides a comprehensive suite of REST APIs for accessing and managing cloud resources. For example, administrators can use REST APIs to create and manage autonomous databases and initiate backups. Oracle Application Express (Oracle APEX)  makes working with REST easy. The  APEX_WEB_SERVICE  package, available since Oracle APEX 4.0, provides developers with a function for consuming REST APIs. Oracle APEX 5.0 introduced the  APEX_JSON  package, allowing easy parsing of JSON responses. Oracle APEX 18.1 introduced the Web Source Module (WSM), which enables developers to access REST services and use the data in APEX components such as reports, interactive reports, and interactive grids. These features, when used together with Oracle Cloud Infrastructure REST ...
Read more