从 Spark GroupedData 对象中选择随机项

Question

I'm new to using Spark in Python and have been unable to solve this problem: After running groupBy on a pyspark.sql.dataframe.DataFrame

df = sqlsc.read.json("data.json")
df.groupBy('teamId')

how can you choose N random samples from each resulting group (grouped by teamId) without replacement?

I'm basically trying to choose N random users from each team, maybe using groupBy is wrong to start with?

Answer 1

Well, it is kind of wrong. GroupedData is not really designed for a data access. It just describes grouping criteria and provides aggregation methods. See my answer to Using groupBy in Spark and getting back to a DataFrame for more details.

Another problem with this idea is selecting N random samples . It is a task which is really hard to achieve in parallel without psychical grouping of data and it is not something that happens when you call groupBy on a DataFrame :

There are at least two ways to handle this:

• convert to RDD, groupBy and perform local sampling

  import random n = 3 def sample(iter, n): rs = random.Random() # We should probably use os.urandom as a seed return rs.sample(list(iter), n) df = sqlContext.createDataFrame( [(x, y, random.random()) for x in (1, 2, 3) for y in "abcdefghi"], ("teamId", "x1", "x2")) grouped = df.rdd.map(lambda row: (row.teamId, row)).groupByKey() sampled = sqlContext.createDataFrame( grouped.flatMap(lambda kv: sample(kv[1], n))) sampled.show() ## +------+---+-------------------+ ## |teamId| x1| x2| ## +------+---+-------------------+ ## | 1| g| 0.81921738561455| ## | 1| f| 0.8563875814036598| ## | 1| a| 0.9010425238735935| ## | 2| c| 0.3864428179837973| ## | 2| g|0.06233470405822805| ## | 2| d|0.37620872770129155| ## | 3| f| 0.7518901502732027| ## | 3| e| 0.5142305439671874| ## | 3| d| 0.6250620479303716| ## +------+---+-------------------+

• use window functions

  from pyspark.sql import Window from pyspark.sql.functions import col, rand, rowNumber w = Window.partitionBy(col("teamId")).orderBy(col("rnd_")) sampled = (df .withColumn("rnd_", rand()) # Add random numbers column .withColumn("rn_", rowNumber().over(w)) # Add rowNumber over windw .where(col("rn_") <= n) # Take n observations .drop("rn_") # drop helper columns .drop("rnd_")) sampled.show() ## +------+---+--------------------+ ## |teamId| x1| x2| ## +------+---+--------------------+ ## | 1| f| 0.8563875814036598| ## | 1| g| 0.81921738561455| ## | 1| i| 0.8173912535268248| ## | 2| h| 0.10862995810038856| ## | 2| c| 0.3864428179837973| ## | 2| a| 0.6695356657072442| ## | 3| b|0.012329360826023095| ## | 3| a| 0.6450777858109182| ## | 3| e| 0.5142305439671874| ## +------+---+--------------------+

but I am afraid both will be rather expensive. If size of the individual groups is balanced and relatively large I would simply use DataFrame.randomSplit .

If number of groups is relatively small it is possible to try something else:

from pyspark.sql.functions import count, udf
from pyspark.sql.types import BooleanType
from operator import truediv

counts = (df
    .groupBy(col("teamId"))
    .agg(count("*").alias("n"))
    .rdd.map(lambda r: (r.teamId, r.n))
    .collectAsMap()) 

# This defines fraction of observations from a group which should
# be taken to get n values 
counts_bd = sc.broadcast({k: truediv(n, v) for (k, v) in counts.items()})

to_take = udf(lambda k, rnd: rnd <= counts_bd.value.get(k), BooleanType())

sampled = (df
    .withColumn("rnd_", rand())
    .where(to_take(col("teamId"), col("rnd_")))
    .drop("rnd_"))

sampled.show()

## +------+---+--------------------+
## |teamId| x1|                  x2|
## +------+---+--------------------+
## |     1|  d| 0.14815204548854788|
## |     1|  f|  0.8563875814036598|
## |     1|  g|    0.81921738561455|
## |     2|  a|  0.6695356657072442|
## |     2|  d| 0.37620872770129155|
## |     2|  g| 0.06233470405822805|
## |     3|  b|0.012329360826023095|
## |     3|  h|  0.9022527556458557|
## +------+---+--------------------+

In Spark 1.5+ you can replace udf with a call to sampleBy method:

df.sampleBy("teamId", counts_bd.value)

It won't give you exact number of observations but should be good enough most of the time as long as a number of observations per group is large enough to get proper samples. You can also use sampleByKey on a RDD in a similar way.

Answer 2

I found this one more dataframey, rather than going into rdd way.

You can use window function to create ranking within a group, where ranking can be random to suit your case. Then, you can filter based on the number of samples (N) you want for each group

window_1 = Window.partitionBy(data['teamId']).orderBy(F.rand())
data_1 = data.select('*', F.rank().over(window_1).alias('rank')).filter(F.col('rank') <= N).drop('rank')

Answer 3

Here's an alternative using Pandas DataFrame.Sample method. This uses the spark applyInPandas method to distribute the groups, available from Spark 3.0.0. This allows you to select an exact number of rows per group.

I've added args and kwargs to the function so you can access the other arguments of DataFrame.Sample .

def sample_n_per_group(n, *args, **kwargs):
    def sample_per_group(pdf):
        return pdf.sample(n, *args, **kwargs)
    return sample_per_group

df = spark.createDataFrame(
    [
        (1, 1.0), 
        (1, 2.0), 
        (2, 3.0), 
        (2, 5.0), 
        (2, 10.0)
    ],
    ("id", "v")
)

(df.groupBy("id")
   .applyInPandas(
        sample_n_per_group(2, random_state=2), 
        schema=df.schema
   )
)

To be aware of the limitations for very large groups, from the documentation :

This function requires a full shuffle. All the data of a group will be loaded into memory, so the user should be aware of the potential OOM risk if data is skewed and certain groups are too large to fit in memory.

See also here: How take a random row from a PySpark DataFrame?