我使用spark-sql-2.4.1v和java8。
必须使用group by对各种条件进行计算,使用 java api,即使用MapFunction和ReduceFunction。
场景。
有如下示例的源数据
+--------+--------------+-----------+-------------+---------+------+
| country|generated_date|industry_id|industry_name| revenue| state|
+--------+--------------+-----------+-------------+---------+------+
|Country1| 2020-03-01| Indus_1| Indus_1_Name| 12789979|State1|
|Country1| 2019-06-01| Indus_1| Indus_1_Name| 56189008|State1|
|Country1| 2019-03-01| Indus_1| Indus_1_Name| 12789979|State1|
|Country1| 2020-03-01| Indus_2| Indus_2_Name| 21789933|State2|
|Country1| 2018-03-01| Indus_2| Indus_2_Name|300789933|State2|
|Country1| 2019-03-01| Indus_3| Indus_3_Name| 27989978|State3|
|Country1| 2017-06-01| Indus_3| Indus_3_Name| 56189008|State3|
|Country1| 2017-03-01| Indus_3| Indus_3_Name| 30014633|State3|
|Country2| 2020-03-01| Indus_4| Indus_4_Name| 41789978|State1|
|Country2| 2018-03-01| Indus_4| Indus_4_Name| 56189008|State1|
|Country3| 2019-03-01| Indus_5| Indus_5_Name| 37899790|State3|
|Country3| 2018-03-01| Indus_5| Indus_5_Name| 56189008|State3|
|Country3| 2017-03-01| Indus_5| Indus_5_Name| 67789978|State3|
|Country1| 2020-03-01| Indus_6| Indus_6_Name| 12789979|State1|
|Country1| 2020-06-01| Indus_6| Indus_6_Name| 37899790|State1|
|Country1| 2018-03-01| Indus_6| Indus_6_Name| 56189008|State1|
|Country3| 2020-03-01| Indus_7| Indus_7_Name| 26689900|State1|
|Country3| 2020-12-01| Indus_7| Indus_7_Name|212359979|State1|
|Country3| 2019-03-01| Indus_7| Indus_7_Name| 12789979|State1|
|Country1| 2018-03-01| Indus_8| Indus_8_Name|212359979|State2|
+--------+--------------+-----------+-------------+---------+------+
需要计算各种计算,如avg(revenue)为每个给定的组给定的日期,能够做到这一点,但在spark-cluster中不缩放。
对于同样的事情,我正在做下面的事情,但这不是在所有的缩放...因此理解我需要使用MapFunction和ReduceFunction的java...不知道如何做到这一点?
//Will get dates to for which I need to calculate , this provided by external source
List<String> datesToCalculate = Arrays.asList("2019-03-01","2020-06-01","2018-09-01");
//Will get groups to calculate , this provided by external source ..will keep changing
//Have around 100s of groups.
List<String> groupsToCalculate = Arrays.asList("Country","Country-State");
//For each data given need to calculate avg(revenue) for each given group
//for those given each date of datesToCalculate for those records whose are later than given date.
//i.e.
//Now I am doing some thing like this..but it is not scaling
datesToCalculate.stream().forEach( cal_date -> {
Dataset<IndustryRevenue> calc_ds = ds.where(col("generated_date").gt(lit(cal_date)));
//this keep changing for each cal_date
Dataset<Row> final_ds = calc_ds
.withColumn("calc_date", to_date(lit(cal_date)).cast(DataTypes.DateType));
//for each group it calcuate separate set
groupsToCalculate.stream().forEach( group -> {
String tempViewName = new String("view_" + cal_date + "_" + group);
final_ds.createOrReplaceTempView(tempViewName);
String query = "select "
+ " avg(revenue) as mean, "
+ "from " + tempViewName
+ " group by " + group;
System.out.println("query : " + query);
Dataset<Row> resultDs = spark.sql(query);
Dataset<Row> finalResultDs = resultDs
.withColumn("calc_date", to_date(lit(cal_date)).cast(DataTypes.DateType))
.withColumn("group", to_date(lit(group)).cast(DataTypes.DateType));
//Writing to each group for each date is taking hell lot of time.
// For each record it is save at a time
// want to move out unioning all finalResultDs and write in batches
finalResultDs
.write().format("parquet")
.mode("append")
.save("/tmp/"+ tempViewName);
spark.catalog().dropTempView(tempViewName);
});
});
由于for-loops,处理几百万条记录需要20多小时,所以如何避免forloops,使其快速运行。
下面是示例代码
https:/github.comBdLearnerrJava-mapReduceblobmasterMapReduceScalingProblem.java。
预期的输出。
+--------------+----------------+--------------+
| group-name | group-value | mean |
+--------------+----------------+--------------+
|country-state |Country1-State1 | 2.53448845E7 |
|country-state |Country3-State3 | 6.7789978E7|
|country-state |Country1-State2 | 1.919319606E8|
|country-state |Country4-State1 | 9.789979E7|
|country-state |Country1-State3 | 2.9339748E7|
|country-state |Country3-State1 | 2.66899E7|
|country-state |Country2-State1 | 4.1789978E7|
|country |Country4 | 9.789979E7|
|country |Country1 | 8.5696311E7|
|country |Country3 | 4.7239939E7|
|country |Country2 | 4.1789978E7|
+--------------+----------------+--------------+
这里是我认为解决你眼前问题的部分方案,但我也留下了一些方面让你填写。还有其他的方法,但这是我根据我的理解快速采取的方法。成功。没有foreach req'd。我可能对你需要的东西有错误的理解。如果是这样的话,请原谅。你可能想在这种方法中考虑.cache。
// Assuming constant names in terms of country names are spelled similarly and consistently
// Not clear if by date or for selected dates. If selected dates then use another list
// This approach will scale due to JOIN and AGG and no foreach, etc.
// Spark will fuse the code together if it can, but there are shuffles
// This is for Country, State. You can apply the approach to just Country and then UNION the 2 DF's with common names and definitions. Try it out
// NB: You make a custom grouping by concatenating the Country & State or you can leave as is, and for 2nd query you can just fill in country and put a blank value into the State.
// I leave that up to you.
import spark.implicits._
import org.apache.spark.sql.functions._
val dfC = Seq(("USA", "Ohio"), ("NZ", "Otago")).toDF("sCountry", "sState") // Your search criteria at Country / State level, cannot so simple .isin - why?
val d = List("23-10-2001", "12-12-2003") // or Array
val dfS = Seq(
("USA", "Ohio", "23-10-2001", 2),
("USA", "Ohio", "23-10-2001", 2),
("USA", "Ohio", "23-10-2011", 2),
("USA", "Texas", "23-10-2001", 2),
("USA", "Virgina", "23-10-2001", 10),
("USA", "Virgina", "23-10-2001", 6),
("USA", "vanDiemensLand", "23-10-2001", 26),
("NL", "vanDiemensLand", "23-10-2001", 16),
("UK", "Middlesex", "23-10-2001", 3)
).toDF("country", "state", "date", "some_val")
dfS.show(false)
// 1. For Country & State
// JOIN acts as a filter as is inner join and alleviates the .isin for multiple cols i.e. Country||State
val df1 = dfS.join(dfC, (dfS("country") === dfC("sCountry")) && (dfS("state") === dfC("sState"))).drop("sCountry").drop("sState")
df1.show(false)
val df2 = df1.filter($"date".isin(d:_*)).groupBy("country", "state").avg("some_val")
df2.show(false)
// 2. For Country only
... to fill in by you
...
// 3. UNION df2 & df3
...
// 4. Save with partitioning.