我有以下问题。我有一个数据集,记录了状态的更改。
id valid eventdate
1 False 2020-05-01
1 True 2020-05-06
2 True 2020-05-04
2 False 2020-05-07
2 True 2020-05-09
3 False 2020-05-11
目标:
SELECT valid FROM table WHERE id = 1 AND eventdate = "2020-05-05"
我需要知道在任意给定日期(在起始日期和今天之间),特定一天的状态是什么。例如,对于id
为1,5月5日时valid仍然是False
。
在Pandas中,我有一个解决方案,其中我使用pivot
和ffill
填充空值。我使用melt将其转换回三列数据框。
from datetime import datetime
import pandas as pd
test_data = [
[1,"False","2020-05-01"],
[1,"True","2020-05-06"],
[2,"True","2020-05-04"],
[2,"False","2020-05-07"],
[2,"True","2020-05-09"],
[3,"False","2020-05-11"]
]
# Create inputframe
df = pd.DataFrame(test_data, columns=['id', 'valid', 'eventdate'])
df['id'] = df['id'].astype(str)
df['valid'] = df['valid'] == "True"
df['eventdate'] = pd.to_datetime(df['eventdate'])
print(df.head(6))
# id valid eventdate
# 0 1 False 2020-05-01
# 1 1 True 2020-05-06
# 2 2 True 2020-05-04
# 3 2 False 2020-05-07
# 4 2 True 2020-05-09
# 5 3 False 2020-05-11
# Create full time range as frame
timeframe = pd.date_range(start=min(df['eventdate']),
end=datetime.now().date()).to_frame().reset_index(drop=True).rename(columns={0: 'eventdate'})
print(timeframe.head())
# eventdate
# 0 2020-05-01
# 1 2020-05-02
# 2 2020-05-03
# 3 2020-05-04
# 4 2020-05-05
# Merge timeframe into original frame
df = df.merge(timeframe,
left_on='eventdate',
right_on='eventdate',
how='right')
print(df.sort_values('eventdate').head())
# id valid eventdate
# 0 1 False 2020-05-01
# 6 NaN NaN 2020-05-02
# 7 NaN NaN 2020-05-03
# 2 2 True 2020-05-04
# 8 NaN NaN 2020-05-05
# 1. Pivot to get dates on rows and ids as columns
# 2. Forward fill values per id
# 3. Fill remaining NaNs with False
df = df.pivot(index='eventdate',
columns='id',
values='valid')\
.fillna(method='ffill')\
.fillna(False)
print(df.head())
# id NaN 1 2 3
# eventdate
# 2020-05-01 False False False False
# 2020-05-02 False False False False
# 2020-05-03 False False False False
# 2020-05-04 False False True False
# 2020-05-05 False False True False
# Drop NaN column and reset the index
df = df.loc[:, df.columns.notnull()].reset_index()
# Melt the columns back
out = pd.melt(df,
id_vars='eventdate',
value_name='valid')
print(out.head(10))
# eventdate id valid
# 0 2020-05-01 1 False
# 1 2020-05-02 1 False
# 2 2020-05-03 1 False
# 3 2020-05-04 1 False
# 4 2020-05-05 1 False
# 5 2020-05-06 1 True
# 6 2020-05-07 1 True
# 7 2020-05-08 1 True
# 8 2020-05-09 1 True
# 9 2020-05-10 1 True
我正在尝试在Spark中实现相同的功能,但是前向填充并不存在。我知道如何通过id
实现最新状态:
w = Window().partitionBy("id").orderBy(F.col("eventdate").desc())
df.withColumn("rn", F.row_number().over(w)) \
.where(F.col("rn") == 1) \
.selectExpr("id", "valid", "eventdate AS last_change") \
.dropna() \
.show()
可以使用以下方式进行数据透视:
df\
.select(["id", "valid", "eventdate"])\
.groupBy(["eventdate"])\
.pivot("id")\
.agg(F.min("valid"))\
.drop('null')\
.sort('eventdate')\
.show()
为了进行前向填充,我通过将数据集限制为仅有一个
id
来实现如下:import sys
from datetime import datetime
import pyspark.sql.functions as F
from pyspark.sql import Window
test_data = [
[1,"False","2020-05-01"],
[1,"True","2020-05-06"],
[2,"True","2020-05-04"],
[2,"False","2020-05-07"],
[2,"True","2020-05-09"],
[3,"False","2020-05-11"]
]
# Create dataframe
df = sc\
.parallelize(test_data)\
.toDF(("id", "valid", "eventdate"))\
.withColumn("eventdate", F.to_date(F.to_timestamp("eventdate")))\
.withColumn("valid", F.when(F.col("valid") == "True", 1).otherwise(0))
df.createOrReplaceTempView("df")
# Create event frame
event_dates = spark.sql("SELECT sequence(min(eventdate), CURRENT_DATE(), interval 1 day) as eventdate FROM df")\
.withColumn("eventdate",
F.explode(F.col("eventdate")))
# Join dates and data
df = df.join(event_dates, on='eventdate', how='right')
df2 = df.where(df.id == 1)\
.join(event_dates, on='eventdate', how='right')\
.withColumn('id', F.lit(1))
#df2.sort('eventdate').show()
# +----------+---+-----+
# | eventdate| id|valid|
# +----------+---+-----+
# |2020-05-01| 1| 0|
# |2020-05-02| 1| null|
# |2020-05-03| 1| null|
# |2020-05-04| 1| null|
# |2020-05-05| 1| null|
# |2020-05-06| 1| 1|
# |2020-05-07| 1| null|
# |2020-05-08| 1| null|
# |2020-05-09| 1| null|
# |2020-05-10| 1| null|
# |2020-05-11| 1| null|
# |2020-05-12| 1| null|
# |2020-05-13| 1| null|
# +----------+---+-----+
# Forward fill
window = Window.partitionBy('id')\
.orderBy('eventdate')\
.rowsBetween(-sys.maxsize, 0)
# Set filter
read_last = F.last(df2['valid'], ignorenulls=True).over(window)
df2.withColumn("ffill", read_last).show()
# +----------+---+-----+-----+
# | eventdate| id|valid|ffill|
# +----------+---+-----+-----+
# |2020-05-01| 1| 0| 0|
# |2020-05-02| 1| null| 0|
# |2020-05-03| 1| null| 0|
# |2020-05-04| 1| null| 0|
# |2020-05-05| 1| null| 0|
# |2020-05-06| 1| 1| 1|
# |2020-05-07| 1| null| 1|
# |2020-05-08| 1| null| 1|
# |2020-05-09| 1| null| 1|
# |2020-05-10| 1| null| 1|
# |2020-05-11| 1| null| 1|
# |2020-05-12| 1| null| 1|
# |2020-05-13| 1| null| 1|
# +----------+---+-----+-----+
我认为首先需要确定这种回答问题的方式是否正确。进行 pivot
操作会创建一个具有很少列的长表,同时存储了大量冗余数据。Spark 不是解决这个问题的正确工具,或者更好的说,这个问题并不适合使用 Spark。理想情况下,你需要使用并行处理,并可能将 timeframe
广播到所有节点,并按节点每个 id
计算前向填充。
也许使用一些不同的方法会更好,例如存储事件的 enddate
,然后在查询时使用以下代码:
id valid eventdate enddate
1 False 2020-05-01 2020-05-06
1 True 2020-05-06 2999-12-31
2 True 2020-05-04 2020-05-07
2 False 2020-05-07 2020-05-08
2 True 2020-05-09 2999-12-31
3 False 2020-05-11 2999-12-31
并且。
SELECT valid FROM table WHERE id = 1 AND "2020-05-05" between eventdate and enddate
请告诉我Spark方法是否可行,以及在这样一个稀疏数据集的任何给定日历状态下找到状态的最佳方法是什么?
谢谢。