As always, brilliant and best on the tube. Small suggestion: we don't need a UDF (everybody hates UDFs), we can simply use the pyspark.sql.functions rand and floor: salted_emp = emp.withColumn("salted_dept_id", concat("department_id", lit("_"), floor(rand()*16)))
very well explained...thanks...in this example we have smaller dept table hence cross joining not impacting the number of records what if we have huge size of dept table also (both our tables are huge) second does salting works on string columns as well (if joining key is string)
In DW concepts DIM tables are basically smaller that Fact tables. But that's a trade-off you have handle if you want to go for manual salting. And yes STRING columns can be used for salting. If you like my content, Please make sure to share with your network over LinkedIn 👍
@@easewithdata if the DIM table is that small, why not broadcast it? Or it wouldn't then solve the skewness? Isn't aggregate a better use case for salting?
thank you for explanation one question is there at 4:41 that 39.6 MB is data first loaded in memory then it deserialized then it become 79mb or it is more than that means( deserialized data - 79 MB ) is able to process at one time then after processing it that 79 MB will be processed
Spillage shows the amount of data it is not able to fit in memory for processing, both in deserialized and serialized form. This data will get processed as soon Spark is able to fit the same after processing some data.
@@easewithdata thanks...like in sort merge join suppose we are joining two tables let suppose after shuffling the shuffled partition of one table have very big size so that it can not fully fit in memory.. I have seen in such a scenario spark is able to join the data with the help of sort merge join how it is possible as we know both table partitions have to be fit in memory for join
Why you made 32 shuffle partition if you have 8core, If one partition is going to process on single core, from where it will get other remaining 24 cores?
If we repartition shuffle partitions after joining, then its again an extra step of exchange. And if we try to repartition before shuffle then the data will again get skewed in shuffle partitions. So, its a choice of optimisation completely based on scenario.
Hello, Unfortunately GitHub is not allowing me to push GB files anymore. I am trying to locate another file sharing service to upload the bigger files.
if we can set the shuffle partition to 32 and don't use the salting technique, just do the joining on original department_id columns then what will happen
Shuffle setting doesn't guarantee even data distribution among executors. In order to make sure the data is distributed properly, we are using slating technique.
As always, brilliant and best on the tube. Small suggestion: we don't need a UDF (everybody hates UDFs), we can simply use the pyspark.sql.functions rand and floor:
salted_emp = emp.withColumn("salted_dept_id", concat("department_id", lit("_"), floor(rand()*16)))
1. do we need to keep -> random numbers and no of shuffle partitions same ? how did you choose 16 ?
very well explained...thanks...in this example we have smaller dept table hence cross joining not impacting the number of records what if we have huge size of dept table also (both our tables are huge) second does salting works on string columns as well (if joining key is string)
In DW concepts DIM tables are basically smaller that Fact tables. But that's a trade-off you have handle if you want to go for manual salting. And yes STRING columns can be used for salting.
If you like my content, Please make sure to share with your network over LinkedIn 👍
@@easewithdata if the DIM table is that small, why not broadcast it? Or it wouldn't then solve the skewness?
Isn't aggregate a better use case for salting?
Great video. Thank you bhaiya.
This is what exactly I was looking for , such a great way to explain.Can you please explain why did you choose 16 as a number of partitions ?
The number partitions need to factor of cores. subhamkharwal.medium.com/pyspark-the-factor-of-cores-e884b2d5af6c
V nice explanation.
thank you for explanation one question is there at 4:41 that 39.6 MB is data first loaded in memory then it deserialized then it become 79mb or it is more than that means( deserialized data - 79 MB ) is able to process at one time then after processing it that 79 MB will be processed
Spillage shows the amount of data it is not able to fit in memory for processing, both in deserialized and serialized form. This data will get processed as soon Spark is able to fit the same after processing some data.
@@easewithdata thanks...like in sort merge join suppose we are joining two tables let suppose after shuffling the shuffled partition of one table have very big size so that it can not fully fit in memory.. I have seen in such a scenario spark is able to join the data with the help of sort merge join how it is possible as we know both table partitions have to be fit in memory for join
Hi Sir,
emp_records_skewed.csv file is missing in the GIT. Could you please upload the file.
Unfortunately the file is too big to upload in Git. You can create the skewed file by merging same type of data multiple times to the same file.
Why you made 32 shuffle partition if you have 8core,
If one partition is going to process on single core, from where it will get other remaining 24 cores?
The 8 cores will process all the 32 partitions in 4 iterations each. (8X4 = 32)
Awesome sir.. why don't you create courses on udemy?? You have a great ability to explain technical aspects ❤❤
Thank you ❤️ I want all good content to be available for free.
If you like my content, just make sure to share with your network over LinkedIn 👍
@@easewithdata definitely sir😊🙏
great work
Thank you for your feedback 💓 Please make sure to share it with your network over LinkedIn 👍
I couldn't find the emp_skewed_data in the repo, can you please share it here in the link or on the git repo.
The dataset is too huge to be uploaded in Github. I am trying an external source to upload the same.
Mass solutions bruhh🎉
Amazing video.
Why not use repartition using id_departament? Isn't it simpler?
If we repartition shuffle partitions after joining, then its again an extra step of exchange. And if we try to repartition before shuffle then the data will again get skewed in shuffle partitions.
So, its a choice of optimisation completely based on scenario.
Watch out for AQE video for the most simpler option to fix Skewness.
@@easewithdata thanks bro, I'm going to watch.
Hi Shubham,
The skewed employee dataset is not available, could you please push it to the git.
Hello,
Unfortunately GitHub is not allowing me to push GB files anymore. I am trying to locate another file sharing service to upload the bigger files.
@@easewithdata Hi Please create file at your end using below code
# Read Employee data
_schema = "first_name string, last_name string, job_title string, dob string, email string, phone string, salary double, department_id int"
emp = spark.read.format("csv").schema(_schema).option("header", True).load("data/employee_records.csv")
from pyspark.sql.functions import lit,count
emp.groupBy("department_id").agg(count("first_name").alias("count")).show()
dept_3 = spark.range(40).select(lit(3).alias("department_id_temp"))
dept_7 = spark.range(40).select(lit(7).alias("department_id_temp"))
emp_inc_3 = emp.join(dept_3, emp["department_id"] == dept_3["department_id_temp"], "left_outer")
emp_inc_3 = emp_inc_3.drop("department_id_temp")
emp_inc_3.groupBy("department_id").agg(count("first_name").alias("count")).show()
emp_inc_7 = emp_inc_3.join(dept_7, emp_inc_3["department_id"] == dept_7["department_id_temp"], "left_outer")
emp_inc_7 = emp_inc_7.drop("department_id_temp")
emp_inc_7.groupBy("department_id").agg(count("first_name").alias("count")).show()
emp_inc_7.repartition(1).write.format("csv").mode("overwrite").option("header", True).save("data/output/emp_record_skewed.csv")
Hi, how can we getvthe data to practice as same@@easewithdata
Great Video👍
if we can set the shuffle partition to 32 and don't use the salting technique, just do the joining on original department_id columns then what will happen
Shuffle setting doesn't guarantee even data distribution among executors. In order to make sure the data is distributed properly, we are using slating technique.
can you pls provide data
All datasets are uploaded at: github.com/subhamkharwal/pyspark-zero-to-hero/tree/master/datasets