Spark:保存由“虚拟"分区的 DataFrame柱子 [英] Spark: save DataFrame partitioned by "virtual" column

问题描述

我正在使用 PySpark 来执行经典的 ETL 作业(加载数据集、处理它、保存它)，并希望将我的 Dataframe 保存为由虚拟"列分区的文件/目录；我所说的虚拟"是指我有一列时间戳，它是一个包含 ISO 8601 编码日期的字符串，我想按年/月/日进行分区；但我实际上在 DataFrame 中没有 Year、Month 或 Day 列；我有这个时间戳，我可以从中派生这些列，但我不希望我的结果项序列化这些列之一.

I'm using PySpark to do classic ETL job (load dataset, process it, save it) and want to save my Dataframe as files/directory partitioned by a "virtual" column; what I mean by "virtual" is that I have a column Timestamp which is a string containing an ISO 8601 encoded date, and I'd want to partition by Year / Month / Day; but I don't actually have either a Year, Month or Day column in the DataFrame; I have this Timestamp from which I can derive these columns though, but I don't want my resultat items to have one of these columns serialized.

将 DataFrame 保存到磁盘所产生的文件结构应如下所示:

File structure resulting from saving the DataFrame to disk should look like:

/ 
    year=2016/
        month=01/
            day=01/
                part-****.gz

有没有办法用 Spark/Pyspark 做我想做的事?

Is there a way to do what I want with Spark / Pyspark ?

推荐答案

用于分区的列不包含在序列化数据本身中.例如，如果您像这样创建 DataFrame:

Columns which are used for partitioning are not included in the serialized data itself. For example if you create DataFrame like this:

df = sc.parallelize([
    (1, "foo", 2.0, "2016-02-16"),
    (2, "bar", 3.0, "2016-02-16")
]).toDF(["id", "x", "y", "date"])

并写成如下:

import tempfile
from pyspark.sql.functions import col, dayofmonth, month, year
outdir = tempfile.mktemp()

dt = col("date").cast("date")
fname = [(year, "year"), (month, "month"), (dayofmonth, "day")]
exprs = [col("*")] + [f(dt).alias(name) for f, name in fname]

(df
    .select(*exprs)
    .write
    .partitionBy(*(name for _, name in fname))
    .format("json")
    .save(outdir))

单个文件不会包含分区列:

individual files won't contain partition columns:

import os

(sqlContext.read
    .json(os.path.join(outdir, "year=2016/month=2/day=16/"))
    .printSchema())

## root
##  |-- date: string (nullable = true)
##  |-- id: long (nullable = true)
##  |-- x: string (nullable = true)
##  |-- y: double (nullable = true)

分区数据仅存储在目录结构中，不会在序列化文件中重复.只有在您阅读完整或部分目录树时才会附加它:

Partitioning data is stored only in a directory structure and not duplicated in serialized files. It will be attached only when your read complete or partial directory tree:

sqlContext.read.json(outdir).printSchema()

## root
##  |-- date: string (nullable = true)
##  |-- id: long (nullable = true)
##  |-- x: string (nullable = true)
##  |-- y: double (nullable = true)
##  |-- year: integer (nullable = true)
##  |-- month: integer (nullable = true)
##  |-- day: integer (nullable = true)

sqlContext.read.json(os.path.join(outdir, "year=2016/month=2/")).printSchema()

## root
##  |-- date: string (nullable = true)
##  |-- id: long (nullable = true)
##  |-- x: string (nullable = true)
##  |-- y: double (nullable = true)
##  |-- day: integer (nullable = true)

这篇关于Spark:保存由“虚拟"分区的 DataFrame柱子的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！