We want to implement SCD2 in Spark using SQL Join. i got reference from Github
https://gist.github.com/rampage644/cc4659edd11d9a288c1b
but it's not very clear.
Can anybody provide any example or reference to implement SCD2 in spark
Regards,
Manish
A little outdated in terms of newer Spark SQL, but here is an example
I trialed a la Ralph Kimball using Spark SQL, that worked and is thus
reliable. You can run it and it works - but file logic and such needs
to be added - this is the body of the ETL SCD2 logic based on 1.6
syntax but run in 2.x - it is not that hard but you will need to trace
through and generate test data and trace through each step:
Some pre-processing required before script initiates, save a copy of existing and copy existing to the DIM_CUSTOMER_EXISTING.
Write new output to DIM_CUSTOMER_NEW and then copy this to target, DIM_CUSTOMER_1 or DIM_CUSTOMER_2.
The feed can also be re-created or LOAD OVERWRITE.
^^^ NEED SOME BETTER SCRIPTING AROUND THIS. ^^^ The Type 2 dimension is simply only Type 2 values, not a mixed Type 1 & Type 2.
DUMPs that are accumulative can be in fact pre-processed to get the delta.
Use case assumes we can have N input for a person with a date validity / extract supplied.
SPARK 1.6 SQL based originally, not updated yet to SPARK 2.x SQL with nested correlated subquery support.
CUST_CODE cannot changes unless a stable Primary Key.
This approach handles no input, delta input, same input, all input, and can catch up and need not be run-date based.
^^^ Works best with deltas, as if pass all data and there is no change then still have make a dummy entry with all the same values else it will have gaps in key range
which means will not be able to link facts to dimensions in all cases. I.e. the discard logic works only in terms of a pure delta feed. All data can be passed but only
the current delta. Problem becomes difficult to solve in that we must then look for changes over different rows and expand date range, a little too complicated imho.
The dummy entries in the dimensions are not a huge issue. The problem is a little more difficult in such a less mutable environment, in KUDU it easier to solve.
Ideally there would be some sort of preprocessor that checks which fields have changed and only then passed on, but that may be a bridge too far.
HENCE IT IS A COMPROMISE ALGORITHM necessarily. ^^^
No Deletions processed.
Multi-step processing for SQL required in some cases. Gaps in key ranges difficult to avoid with set processing.
No out of order processing, that would mean re-processing all. Works on a whole date/day basis, if run more than once per day in batch then would need timestamp instead.
0.1 Any difference analysis on existimg dumps only possible if the dumps are accumulative. If they are transactional deltas only, then this is not required.
Care to be taken here.
0.2 If we want only the last update for a given date, then do that here by method of Partitioning and Ranking and filtering out.
These are all pre-processing steps as are the getting of the dimension data from which table.
0.3 Issue is that of small files, but that is not an issue here at xxx. RAW usage only as written to KUDU in a final step.
Actual coding:
import org.apache.spark.sql.SparkSession
val sparkSession = SparkSession
.builder
.master("local") // Not a good idea
.appName("Type 2 dimension update")
.config("spark.sql.crossJoin.enabled", "true") // Needed to add this
.getOrCreate()
spark.sql("drop table if exists DIM_CUSTOMER_EXISTING")
spark.sql("drop table if exists DIM_CUSTOMER_NEW")
spark.sql("drop table if exists FEED_CUSTOMER")
spark.sql("drop table if exists DIM_CUSTOMER_TEMP")
spark.sql("drop table if exists DIM_CUSTOMER_WORK")
spark.sql("drop table if exists DIM_CUSTOMER_WORK_2")
spark.sql("drop table if exists DIM_CUSTOMER_WORK_3")
spark.sql("drop table if exists DIM_CUSTOMER_WORK_4")
spark.sql("create table DIM_CUSTOMER_EXISTING (DWH_KEY int, CUST_CODE String, CUST_NAME String, ADDRESS_CITY String, SALARY int, VALID_FROM_DT String, VALID_TO_DT String) ROW FORMAT DELIMITED FIELDS TERMINATED BY ',' LINES TERMINATED BY '\n' STORED AS TEXTFILE LOCATION '/FileStore/tables/alhwkf661500326287094' ")
spark.sql("create table DIM_CUSTOMER_NEW (DWH_KEY int, CUST_CODE String, CUST_NAME String, ADDRESS_CITY String, SALARY int, VALID_FROM_DT String, VALID_TO_DT String) ROW FORMAT DELIMITED FIELDS TERMINATED BY ',' LINES TERMINATED BY '\n' STORED AS TEXTFILE LOCATION '/FileStore/tables/DIM_CUSTOMER_NEW_3' ")
spark.sql("CREATE TABLE FEED_CUSTOMER (CUST_CODE String, CUST_NAME String, ADDRESS_CITY String, SALARY int, VALID_DT String) ROW FORMAT DELIMITED FIELDS TERMINATED BY ',' LINES TERMINATED BY '\n' STORED AS TEXTFILE LOCATION '/FileStore/tables/mhiscfsv1500226290781' ")
// 1. Get maximum value in dimension, this differs to other RDD approach, issues in parallel? May be other way to be done! Check, get a DF here and this is the interchangability
val max_val = spark.sql("select max(dwh_key) from DIM_CUSTOMER_EXISTING")
//max_val.show()
val null_count = max_val.filter("max(DWH_KEY) is null").count()
var max_Dim_Key = 0;
if ( null_count == 1 ) {
max_Dim_Key = 0
} else {
max_Dim_Key = max_val.head().getInt(0)
}
//2. Cannot do simple difference processing. The values of certain fields could be flip-flopping over time. A too simple MINUS will not work well. Need to process relative to
// youngest existing record etc. and roll the transactions forward. Hence we will not do any sort of difference analysis between new dimension data and existing dimension
// data in any way.
// DO NOTHING.
//3. Capture new stuff to be inserted.
// Some records for a given business key can be linea recta inserted as there have been no mutations to consider at all as there is nothing in current Staging. Does not mean
// delete.
// Also, the older mutations need not be re-processed, only the youngest! The younger one may need closing off or not, need to decide if it is now
// copied across or subject to updating in this cycle, depends on the requirements.
// Older mutations copied across immediately.
// DELTA not always strictly speaking needed, but common definitions. Some ranking required.
spark.sql("""insert into DIM_CUSTOMER_NEW select *
from DIM_CUSTOMER_EXISTING
where CUST_CODE not in (select distinct CUST_CODE FROM FEED_CUSTOMER) """) // This does not need RANKing, DWH Key retained.
spark.sql("""create table DIM_CUSTOMER_TEMP as select *, dense_rank() over (partition by CUST_CODE order by VALID_FROM_DT desc) as RANK
from DIM_CUSTOMER_EXISTING """)
spark.sql("""insert into DIM_CUSTOMER_NEW select DWH_KEY, CUST_CODE, CUST_NAME, ADDRESS_CITY, SALARY, VALID_FROM_DT, VALID_TO_DT
from DIM_CUSTOMER_TEMP
where CUST_CODE in (select distinct CUST_CODE from FEED_CUSTOMER)
and RANK <> 1 """)
// For updating of youngest record in terms of SLCD, we use use AND RANK <> 1 to filter these out here as we want to close off the period in this record, but other younger
// records can be passed through immediately with their retained DWH Key.
//4. Combine Staging and those existing facts required. The result of this eventually will be stored in DIM_CUSTOMER_NEW which can be used for updating a final target.
// Issue here is that DWH Key not yet set and different columns. DWH key can be set last.
//4.1 Get records to process, the will have the status NEW.
spark.sql("""create table DIM_CUSTOMER_WORK (DWH_KEY int, CUST_CODE String, CUST_NAME String, ADDRESS_CITY String, SALARY int, VALID_FROM_DT String, VALID_TO_DT String, RECSTAT String) """)
spark.sql("""insert into DIM_CUSTOMER_WORK select 0, CUST_CODE, CUST_NAME, ADDRESS_CITY, SALARY, VALID_DT, '2099-12-31', "NEW"
from FEED_CUSTOMER """)
//4.2 Get youngest already existing dimension record to process in conjunction with newer values.
spark.sql("""insert into DIM_CUSTOMER_WORK select DWH_KEY, CUST_CODE, CUST_NAME, ADDRESS_CITY, SALARY, VALID_FROM_DT, VALID_TO_DT, "OLD"
from DIM_CUSTOMER_TEMP
where CUST_CODE in (select distinct CUST_CODE from FEED_CUSTOMER)
and RANK = 1 """)
// 5. ISSUE with first record in a set. It is not a delta or is used for making a delta, need to know what to do or bypass, depends on case.
// Here we are doing deltas, so first rec is a complete delta
// RECSTAT to be filtered out at end
// NEW, 1 = INSERT --> checked, is correct way, can do in others. No delta computation required
// OLD, 1 = DO NOTHING
// else do delta and INSERT
//5.1 RANK and JOIN to get before and after images in CDC format so that we can decide what needs to be closed off.
// Get the new DWH key values + offset, there may exist gaps eventually.
spark.sql(""" create table DIM_CUSTOMER_WORK_2 as select *, rank() over (partition by CUST_CODE order by VALID_FROM_DT asc) as rank FROM DIM_CUSTOMER_WORK """)
//DWH_KEY, CUST_CODE, CUST_NAME, BIRTH_CITY, SALARY,VALID_FROM_DT, VALID_TO_DT, "OLD"
spark.sql(""" create table DIM_CUSTOMER_WORK_3 as
select T1.DWH_KEY as T1_DWH_KEY, T1.CUST_CODE as T1_CUST_CODE, T1.rank as CURR_RANK, T2.rank as NEXT_RANK,
T1.VALID_FROM_DT as CURR_VALID_FROM_DT, T2.VALID_FROM_DT as NEXT_VALID_FROM_DT,
T1.VALID_TO_DT as CURR_VALID_TO_DT, T2.VALID_TO_DT as NEXT_VALID_TO_DT,
T1.CUST_NAME as CURR_CUST_NAME, T2.CUST_NAME as NEXT_CUST_NAME,
T1.SALARY as CURR_SALARY, T2.SALARY as NEXT_SALARY,
T1.ADDRESS_CITY as CURR_ADDRESS_CITY, T2.ADDRESS_CITY as NEXT_ADDRESS_CITY,
T1.RECSTAT as CURR_RECSTAT, T2.RECSTAT as NEXT_RECSTAT
from DIM_CUSTOMER_WORK_2 T1 LEFT OUTER JOIN DIM_CUSTOMER_WORK_2 T2
on T1.CUST_CODE = T2.CUST_CODE AND T2.rank = T1.rank + 1 """)
//5.2 Get the data for computing new Dimension Surrogate DWH Keys, must execute new query or could use DF's and RDS, RDDs, but chosen for SPARK SQL as aeasier to follow
spark.sql(s""" create table DIM_CUSTOMER_WORK_4 as
select *, row_number() OVER( ORDER BY T1_CUST_CODE) as ROW_NUMBER, '$max_Dim_Key' as DIM_OFFSET
from DIM_CUSTOMER_WORK_3 """)
//spark.sql("""SELECT * FROM DIM_CUSTOMER_WORK_4 """).show()
//Execute the above to see results, could not format here.
//5.3 Process accordingly and check if no change at all, if no change can get holes in the sequence numbers, that is not an issue. NB: NOT DOING THIS DUE TO COMPLICATIONS !!!
// See sample data above for decision-making on what to do. NOTE THE FACT THAT WE WOULD NEED A PRE_PROCCESOR TO CHECK IF FIELD OF INTEREST ACTUALLY CHANGED
// to get the best result.
// We could elaborate and record via an extra step if there were only two records per business key and if all the current and only next record fields were all the same,
// we could disregard the first and the second record. Will attempt that later as an extra optimization. As soon as there are more than two here, then this scheme packs up
// Some effort still needed.
//5.3.1 Records that just need to be closed off. The previous version gets an appropriate DATE - 1. Dates must not overlap.
// No check on whether data changed or not due to issues above.
spark.sql("""insert into DIM_CUSTOMER_NEW select T1_DWH_KEY, T1_CUST_CODE, CURR_CUST_NAME, CURR_ADDRESS_CITY, CURR_SALARY,
CURR_VALID_FROM_DT, cast(date_sub(cast(NEXT_VALID_FROM_DT as DATE), 1) as STRING)
from DIM_CUSTOMER_WORK_4
where CURR_RECSTAT = 'OLD' """)
//5.3.2 Records that are the last in the sequence must have high end 2099-12-31 set, which has already been done.
// No check on whether data changed or not due to issues above.
spark.sql("""insert into DIM_CUSTOMER_NEW select ROW_NUMBER + DIM_OFFSET, T1_CUST_CODE, CURR_CUST_NAME, CURR_ADDRESS_CITY, CURR_SALARY,
CURR_VALID_FROM_DT, CURR_VALID_TO_DT
from DIM_CUSTOMER_WORK_4
where NEXT_RANK is null """)
//5.3.3
spark.sql("""insert into DIM_CUSTOMER_NEW select ROW_NUMBER + DIM_OFFSET, T1_CUST_CODE, CURR_CUST_NAME, CURR_ADDRESS_CITY, CURR_SALARY,
CURR_VALID_FROM_DT, cast(date_sub(cast(NEXT_VALID_FROM_DT as DATE), 1) as STRING)
from DIM_CUSTOMER_WORK_4
where CURR_RECSTAT = 'NEW'
and NEXT_RANK is not null""")
spark.sql("""SELECT * FROM DIM_CUSTOMER_NEW """).show()
// So, the question is if we could have done without JOINing and just sorted due to gap processing. This was derived off the delta processing but it turned out a little
// different.
// Well we did need the JOIN for next date at least, so if we add some optimization it still holds.
// My logic applied here per different steps, may well be less steps, left as is.
//6. The copy / insert to get a new big target table version and re-compile views. Outside of this actual processing. Logic performed elsewhere.
// NOTE now that 2.x supports nested correlated sub-queries are supported, so would need to re-visit this at a later point, but can leave as is.
// KUDU means no more restating.
Sample data so you know what to generate for the examples:
+-------+---------+----------------+------------+------+-------------+-----------+
|DWH_KEY|CUST_CODE| CUST_NAME|ADDRESS_CITY|SALARY|VALID_FROM_DT|VALID_TO_DT|
+-------+---------+----------------+------------+------+-------------+-----------+
| 230| E222222| Pete Saunders| Leeds| 75000| 2013-03-09| 2099-12-31|
| 400| A048901| John Alexander| Calgary| 22000| 2015-03-24| 2017-10-22|
| 402| A048901| John Alexander| Wellington| 47000| 2017-10-23| 2099-12-31|
| 403| B787555| Mark de Wit|Johannesburg| 49500| 2017-10-02| 2099-12-31|
| 406| C999666| Daya Dumar| Mumbai| 50000| 2016-12-16| 2099-12-31|
| 404| C999666| Daya Dumar| Mumbai| 49000| 2016-11-11| 2016-12-14|
| 405| C999666| Daya Dumar| Mumbai| 50000| 2016-12-15| 2016-12-15|
| 300| A048901| John Alexander| Calgary| 15000| 2014-03-24| 2015-03-23|
+-------+---------+----------------+------------+------+-------------+-----------+
Here's the detailed implementation of slowly changing dimension type 2 in Spark (Data frame and SQL) using exclusive join approach.
Assuming that the source is sending a complete data file i.e. old, updated and new records.
Steps:
Load the recent file data to STG table
Select all the expired records from HIST table
1. select * from HIST_TAB where exp_dt != '2099-12-31'
Select all the records which are not changed from STG and HIST using inner join and filter on HIST.column = STG.column as below
2. select hist.* from HIST_TAB hist inner join STG_TAB stg on hist.key = stg.key where hist.column = stg.column
Select all the new and updated records which are changed from STG_TAB using exclusive left join with HIST_TAB and set expiry and effective date as below
3. select stg.*, eff_dt (yyyy-MM-dd), exp_dt (2099-12-31) from STG_TAB stg left join (select * from HIST_TAB where exp_dt = '2099-12-31') hist
on hist.key = stg.key where hist.key is null or hist.column != stg.column
Select all updated old records from the HIST table using exclusive left join with STG table and set their expiry date as shown below:
4. select hist.*, exp_dt(yyyy-MM-dd) from (select * from HIST_TAB where exp_dt = '2099-12-31') hist left join STG_TAB stg
on hist.key= stg.key where hist.key is null or hist.column!= stg.column
unionall queries from 1-4 and insert overwrite result to HIST table
More detailed implementation of SCD type 2 in Scala and Pyspark can be found here-
https://github.com/sahilbhange/spark-slowly-changing-dimension
Hope this helps!
scala spark: https://georgheiler.com/2020/11/19/sparkling-scd2/
NOTICE: this is not a full SCD2 - it assumes one table of events and it determines/ deduplicates valid_from/valid_to from them i.e. no merge/upsert is implemented
val df = Seq(("k1","foo", "2020-01-01"), ("k1","foo", "2020-02-01"), ("k1","baz", "2020-02-01"),
("k2","foo", "2019-01-01"), ("k2","foo", "2019-02-01"), ("k2","baz", "2019-02-01")).toDF("key", "value_1", "date").withColumn("date", to_date(col("date")))
df.show
+---+-------+----------+
|key|value_1| date|
+---+-------+----------+
| k1| foo|2020-01-01|
| k1| foo|2020-02-01|
| k1| baz|2020-02-01|
| k2| foo|2019-01-01|
| k2| foo|2019-02-01|
| k2| baz|2019-02-01|
+---+-------+----------+
df.printSchema
root
|-- key: string (nullable = true)
|-- value_1: string (nullable = true)
|-- date: date (nullable = true)
df.transform(deduplicateScd2(Seq("key"), Seq("date"), "date", Seq())).show
+---+-------+----------+----------+
|key|value_1|valid_from| valid_to|
+---+-------+----------+----------+
| k1| foo|2020-01-01|2020-02-01|
| k1| baz|2020-02-01|2020-11-18|
| k2| foo|2019-01-01|2019-02-01|
| k2| baz|2019-02-01|2020-11-18|
+---+-------+----------+----------+
import org.apache.spark.sql.DataFrame
import org.apache.spark.sql.expressions.Window
import org.apache.spark.sql.functions.col
import org.apache.spark.sql.functions.lag
import org.apache.spark.sql.functions.lead
import org.apache.spark.sql.functions.when
import org.apache.spark.sql.functions.current_date
def deduplicateScd2(
key: Seq[String],
sortChangingIgnored: Seq[String],
timeColumn: String,
columnsToIgnore: Seq[String]
)(df: DataFrame): DataFrame = {
val windowPrimaryKey = Window
.partitionBy(key.map(col): _*)
.orderBy(sortChangingIgnored.map(col): _*)
val columnsToCompare =
df.drop(key ++ sortChangingIgnored: _*).drop(columnsToIgnore: _*).columns
val nextDataChange = lead(timeColumn, 1).over(windowPrimaryKey)
val deduplicated = df
.withColumn(
"data_changes_start",
columnsToCompare
.map(e => {
val previous = lag(col(e), 1).over(windowPrimaryKey)
val self = col(e)
// 3 cases: 1.: start (previous is NULL), 2: in between, try to collapse 3: end (= next is null)
// first, filter to only start & end events (= updates/invalidations of records)
//self =!= previous or self =!= next or previous.isNull or next.isNull
self =!= previous or previous.isNull
})
.reduce(_ or _)
)
.withColumn(
"data_changes_end",
columnsToCompare
.map(e => {
val next = lead(col(e), 1).over(windowPrimaryKey)
val self = col(e)
// 3 cases: 1.: start (previous is NULL), 2: in between, try to collapse 3: end (= next is null)
// first, filter to only start & end events (= updates/invalidations of records)
self =!= next or next.isNull
})
.reduce(_ or _)
)
.filter(col("data_changes_start") or col("data_changes_end"))
.drop("data_changes")
deduplicated //.withColumn("valid_to", nextDataChange)
.withColumn(
"valid_to",
when(col("data_changes_end") === true, col(timeColumn))
.otherwise(nextDataChange)
)
.filter(col("data_changes_start") === true)
.withColumn(
"valid_to",
when(nextDataChange.isNull, current_date()).otherwise(col("valid_to"))
)
.withColumnRenamed(timeColumn, "valid_from")
.drop("data_changes_end", "data_changes_start")
}
}
Here an updated answer with MERGE.
Note it will not work with Spark Structured Streaming, but can be used with Spark Kafka Batch Integration.
// 0. Standard, start of program.
// Handles multiple business keys in a single run. DELTA tables.
// Schema evolution also handled.
import org.apache.spark.sql.SparkSession
import org.apache.spark.sql.functions._
import org.apache.spark.sql.types._
val sparkSession = SparkSession.builder
.master("local") // Not realistic
.appName("REF Zone History stuff and processing")
.enableHiveSupport() // Standard in Databricks.
.getOrCreate()
// 1. Read newer data to process in some way. Create tempView.
// In general we should have few rows to process, i.e. not at scale.
val dfA = spark.read.option("multiLine",false).json("/FileStore/tables/new_customers_json_multiple_alt3.txt") // New feed.
dfA.createOrReplaceTempView("newFeed")
// 2. First create the target for data at rest if it does not exist. Add an ASC col_key. Should only occur once.
val save_path = "/some_loc_fix/ref/atRest/data" // Make dynamic.
val table_name = "CUSTOMERS_AT_REST"
spark.sql("CREATE TABLE IF NOT EXISTS " + table_name + " LOCATION '" + save_path + "'" + " AS SELECT * from newFeed WHERE 1 = 0 " ) // Can also use limit 0 instead of WHERE 1 = 0.
// Add an ASC col_key column if it does not exist.
// I have in input valid_from_dt, but it could be different so we would need to add in reality as well. Mark to decide.
try {
spark.sql("ALTER TABLE " + table_name + " ADD COLUMNS (col_key BIGINT FIRST, valid_to_dt STRING) ")
} catch {
case unknown: Exception => {
None
}
}
// 3. Get maximum value for target. This is a necessity.
val max_val = spark.sql("select max(col_key) from " + table_name)
//max_val.show()
val null_count = max_val.filter("max(col_key) is null").count()
var max_Col_Key: BigInt = 0;
if ( null_count == 1 ) {
max_Col_Key = 0
} else {
max_Col_Key = max_val.head().getLong(0) // Long and BIGINT interoperable.
}
// 4.1 Create a temporary table for getting the youngest records from the existing data. table_name as variable, newFeed tempView as string. Then apply processing.
val dfB = spark.sql(" select O.* from (select A.cust_code, max(A.col_key) as max_col_key from " + table_name + " A where A.cust_code in (select B.cust_code from newFeed B ) group by A.cust_code ) Z, " + table_name + " O where O.col_key = Z.max_col_key ") // Most recent records.
// No tempView required.
// 4.2 Get the set of data to actually process. New feed + youngest records in feed.
val dfC =dfA.unionByName(dfB, true)
dfC.createOrReplaceTempView("cusToProcess")
// 4.3 RANK
val df1 = spark.sql("""select *, dense_rank() over (partition by CUST_CODE order by VALID_FROM_DT desc) as RANK from CusToProcess """)
df1.createOrReplaceTempView("CusToProcess2")
// 4.4 JOIN adjacent records & process closing off dates etc.
val df2 = spark.sql("""select A.*, B.rank as B_rank, cast(date_sub(cast(B.valid_from_dt as DATE), 1) as STRING) as untilMinus1
from CusToProcess2 A LEFT OUTER JOIN CusToProcess2 B
on A.cust_code = B.cust_code and A.RANK = B.RANK + 1 """)
val df3 = df2.drop("valid_to_dt").withColumn("valid_to_dt", $"untilMinus1").drop("untilMinus1").drop("B_rank")
val df4 = df3.withColumn("valid_to_dt", when($"valid_to_dt".isNull, lit("2099-12-31")).otherwise($"valid_to_dt")).drop("RANK")
df4.createOrReplaceTempView("CusToProcess3")
val df5 = spark.sql(s""" select *, row_number() OVER( ORDER BY cust_code ASC, valid_from_dt ASC) as ROW_NUMBER, '$max_Col_Key' as col_OFFSET
from CusToProcess3 """)
// Add new ASC col_key, gaps can result, not an issue must always be ascending.
val df6 = df5.withColumn("col_key", when($"col_key".isNull, ($"ROW_NUMBER" + $"col_OFFSET")).otherwise($"col_key"))
val df7 = df6.withColumn("col_key", col("col_key").cast(LongType)).drop("ROW_NUMBER").drop("col_OFFSET")
// 5. ACTUAL MERGE, is very simple.
// More than one Merge key possible? Need then to have a col_key if only one such possible.
df7.createOrReplaceTempView("CUST_DELTA")
spark.sql("SET spark.databricks.delta.schema.autoMerge.enabled = true")
spark.sql(""" MERGE INTO CUSTOMERS_AT_REST
USING CUST_DELTA
ON CUSTOMERS_AT_REST.col_key = CUST_DELTA.col_key
WHEN MATCHED THEN
UPDATE SET *
WHEN NOT MATCHED THEN
INSERT *
""")
I'm having an issue with scheduling job in Data Factory.
I'm trying to approach a scheduled job per hour which will execute the same script each hour with different condition.
Consider I have a bunch of Avro Files spread in Azure Data Lake Store with following pattern.
/Data/SomeEntity/{date:yyyy}/{date:MM}/{date:dd}/SomeEntity_{date:yyyy}{date:MM}{date:dd}__{date:H}
Each hour new files are added to Data Lake Store.
In order to process the files only once I decided to handle them by help of U-SQL virtual file set column and some SyncTable which i created in Data Lake Store.
My query looks like following.
DECLARE #file_set_path string = /Data/SomeEntity/{date:yyyy}/{date:MM}/{date:dd}/SomeEntity_{date:yyyy}_{date:MM}_{date:dd}__{date:H};
#result = EXTRACT [Id] long,
....
date DateTime
FROM #file_set_path
USING someextractor;
#rdate =
SELECT MAX(ProcessedDate) AS ProcessedDate
FROM dbo.SyncTable
WHERE EntityName== "SomeEntity";
#finalResult = SELECT [Id],... FROM #result
CROSS JOIN #rdate AS r
WHERE date >= r.ProcessedDate;
since I can't use rowset variable in where clause I'm cross joining the singe row with set , however even in this case U-SQL won't find the correct files and always return all files set.
Is there any workaround or other approach ?
I think this approach should work unless there is something not quite right somewhere, ie can you confirm the datatypes of the dbo.SyncTable table? Dump out #rdate and make sure the value you get there is what you expect.
I put together a simple demo which worked as expected. My copy of SyncTable had one record with the value of 01/01/2018:
#working =
SELECT *
FROM (
VALUES
( (int)1, DateTime.Parse("2017/12/31") ),
( (int)2, DateTime.Parse("2018/01/01") ),
( (int)3, DateTime.Parse("2018/02/01") )
) AS x ( id, someDate );
#rdate =
SELECT MAX(ProcessedDate) AS maxDate
FROM dbo.SyncTable;
//#output =
// SELECT *
// FROM #rdate;
#output =
SELECT *, (w.someDate - r.maxDate).ToString() AS diff
FROM #working AS w
CROSS JOIN
#rdate AS r
WHERE w.someDate >= r.maxDate;
OUTPUT #output TO "/output/output.csv"
USING Outputters.Csv();
I did try this with a filepath (full script here). The thing to remember is the custom date format H represents the hour as a number from 0 to 23. If your SyncTable date does not have a time component to it when you insert it, it will default to midnight (0), meaning the whole day will be collected. Your file structure should look something like this according to your pattern:
"D:\Data Lake\USQLDataRoot\Data\SomeEntity\2017\12\31\SomeEntity_2017_12_31__8\test.csv"
I note your filepath has underscores in the second section and a double underscore before the hour section (which will be between 0 and 23, single digit up to the hour 10). I notice your fileset path does not have a file type or quotes - I've used test.csv in my tests. My results:
Basically I think the approach will work, but there is something not quite right, maybe in your file structure, the value in your SyncTable, the datatype etc. You need to go over the details, dump out intermediate values to check until you find the problem.
Doesn't the gist of wBob's full script resolve your issue? Here is a very slightly edited version of wBob's full script to address some of the issues you raised:
Ability to filter on SyncTable,
last part of pattern is file name and not folder. Sample file and structure: \Data\SomeEntity\2018\01\01\SomeEntity_2018_01_01__1
DECLARE #file_set_path string = #"/Data/SomeEntity/{date:yyyy}/{date:MM}/{date:dd}/SomeEntity_{date:yyyy}_{date:MM}_{date:dd}__{date:H}";
#input =
EXTRACT [Id] long,
date DateTime
FROM #file_set_path
USING Extractors.Text();
// in lieu of creating actual table
#syncTable =
SELECT * FROM
( VALUES
( "SomeEntity", new DateTime(2018,01,01,01,00,00) ),
( "AnotherEntity", new DateTime(2018,01,01,01,00,00) ),
( "SomeEntity", new DateTime(2018,01,01,00,00,00) ),
( "AnotherEntity", new DateTime(2018,01,01,00,00,00) ),
( "SomeEntity", new DateTime(2017,12,31,23,00,00) ),
( "AnotherEntity", new DateTime(2017,12,31,23,00,00) )
) AS x ( EntityName, ProcessedDate );
#rdate =
SELECT MAX(ProcessedDate) AS maxDate
FROM #syncTable
WHERE EntityName== "SomeEntity";
#output =
SELECT *,
date.ToString() AS dateString
FROM #input AS i
CROSS JOIN
#rdate AS r
WHERE i.date >= r.maxDate;
OUTPUT #output
TO "/output/output.txt"
ORDER BY Id
USING Outputters.Text(quoting:false);
Also please note that file sets cannot perform partition elimination on dynamic joins, since the values are not known to the optimizer during the preparation phase.
I would suggest to pass the Sync point as a parameter from ADF to the processing script. Then the value is known to the optimizer and file set partition elimination will kick in. In the worst case, you would have to read the value from your sync table in a previous script and use it as a parameter in the next.
So as far as I know Apache Spark doesn't has a functionality that imitates the update SQL command. Like, I can change a single value in a column given a certain condition. The only way around that is to use the following command I was instructed to use (here in Stackoverflow): withColumn(columnName, where('condition', value));
However, the condition should be of column type, meaning I have to use the built in column filtering functions apache has (equalTo, isin, lt, gt, etc). Is there a way I can instead use an SQL statement instead of those built in functions?
The problem is I'm given a text file with SQL statements, like WHERE ID > 5 or WHERE AGE != 50, etc. Then I have to label values based on those conditions, and I thought of following the withColumn() approach but I can't plug-in an SQL statement in that function. Any idea of how I can go around this?
I found a way to go around this:
You want to split your dataset into two sets: the values you want to update and the values you don't want to update
Dataset<Row> valuesToUpdate = dataset.filter('conditionToFilterValues');
Dataset<Row> valuesNotToUpdate = dataset.except(valuesToUpdate);
valueToUpdate = valueToUpdate.withColumn('updatedColumn', lit('updateValue'));
Dataset<Row> updatedDataset = valuesNotToUpdate.union(valueToUpdate);
This, however, doesn't keep the same order of records as the original dataset, so if order is of importance to you, this won't suffice your needs.
In PySpark you have to use .subtract instead of .except
If you are using DataFrame, you can register that dataframe as temp table,
using df.registerTempTable("events")
Then you can query like,
sqlContext.sql("SELECT * FROM events "+)
when clause translates into case clause which you can relate to SQL case clause.
Example
scala> val condition_1 = when(col("col_1").isNull,"NA").otherwise("AVAILABLE")
condition_1: org.apache.spark.sql.Column = CASE WHEN (col_1 IS NULL) THEN NA ELSE AVAILABLE END
or you can chain when clause as well
scala> val condition_2 = when(col("col_1") === col("col_2"),"EQUAL").when(col("col_1") > col("col_2"),"GREATER").
| otherwise("LESS")
condition_2: org.apache.spark.sql.Column = CASE WHEN (col_1 = col_2) THEN EQUAL WHEN (col_1 > col_2) THEN GREATER ELSE LESS END
scala> val new_df = df.withColumn("condition_1",condition_1).withColumn("condition_2",condition_2)
Still if you want to use table, then you can register your dataframe / dataset as temperory table and perform sql queries
df.createOrReplaceTempView("tempTable")//spark 2.1 +
df.registerTempTable("tempTable")//spark 1.6
Now, you can perform sql queries
spark.sql("your queries goes here with case clause and where condition!!!")//spark 2.1
sqlContest.sql("your queries goes here with case clause and where condition!!!")//spark 1.6
If you are using java dataset
you can update dataset by below.
here is the code
Dataset ratesFinal1 = ratesFinal.filter(" on_behalf_of_comp_id != 'COMM_DERIVS' ");
ratesFinal1 = ratesFinal1.filter(" status != 'Hit/Lift' ");
Dataset ratesFinalSwap = ratesFinal1.filter (" on_behalf_of_comp_id in ('SAPPHIRE','BOND') and cash_derivative != 'cash'");
ratesFinalSwap = ratesFinalSwap.withColumn("ins_type_str",functions.lit("SWAP"));
adding new column with value from existing column
ratesFinalSTW = ratesFinalSTW.withColumn("action", ratesFinalSTW.col("status"));
I have an operation that I want to perform within PySpark 2.0 that would be easy to perform as a df.rdd.map, but since I would prefer to stay inside the Dataframe execution engine for performance reasons, I want to find a way to do this using Dataframe operations only.
The operation, in RDD-style, is something like this:
def precision_formatter(row):
formatter = "%.{}f".format(row.precision)
return row + [formatter % row.amount_raw / 10 ** row.precision]
df = df.rdd.map(precision_formatter)
Basically, I have a column that tells me, for each row, what the precision for my string formatting operation should be, and I want to selectively format the 'amount_raw' column as a string depending on that precision.
I don't know of a way to use the contents of one or more columns as input to another Column operation. The closest I can come is suggesting the use of Column.when with an externally-defined set of boolean operations that correspond to the set of possible boolean conditions/cases within the column or columns.
In this specific case, for instance, if you can obtain (or better yet, already have) all possible values of row.precision, then you can iterate over that set and apply a Column.when operation for each value in the set. I believe this set can be obtained with df.select('precision').distinct().collect().
Because the pyspark.sql.functions.when and Column.when operations themselves return a Column object, you can iterate over the items in the set (however it was obtained) and keep 'appending' when operations to each other programmatically until you have exhausted the set:
import pyspark.sql.functions as PSF
def format_amounts_with_precision(df, all_precisions_set):
amt_col = PSF.when(df['precision'] == 0, df['amount_raw'].cast(StringType()))
for precision in all_precisions_set:
if precision != 0: # this is a messy way of having a base case above
fmt_str = '%.{}f'.format(precision)
amt_col = amt_col.when(df['precision'] == precision,
PSF.format_string(fmt_str, df['amount_raw'] / 10 ** precision)
return df.withColumn('amount', amt_col)
You can do it with a python UDF. They can take as many input values (values from columns of a Row) and spit out a single output value. It would look something like this:
from pyspark.sql import types as T, functions as F
from pyspark.sql.function import udf, col
# Create example data frame
schema = T.StructType([
T.StructField('precision', T.IntegerType(), False),
T.StructField('value', T.FloatType(), False)
])
data = [
(1, 0.123456),
(2, 0.123456),
(3, 0.123456)
]
rdd = sc.parallelize(data)
df = sqlContext.createDataFrame(rdd, schema)
# Define UDF and apply it
def format_func(precision, value):
format_str = "{:." + str(precision) + "f}"
return format_str.format(value)
format_udf = F.udf(format_func, T.StringType())
new_df = df.withColumn('formatted', format_udf('precision', 'value'))
new_df.show()
Also, if instead of the column precision value you wanted to use a global one, you could use the lit(..) function when you call it like this:
new_df = df.withColumn('formatted', format_udf(F.lit(2), 'value'))