How to handle data skew in the spark data frame for outer join

Question

I have two data frames and I am performing outer join on 5 columns . Below is example of my data set .

uniqueFundamentalSet|^|PeriodId|^|SourceId|^|StatementTypeCode|^|StatementCurrencyId|^|FinancialStatementLineItem.lineItemId|^|FinancialAsReportedLineItemName|^|FinancialAsReportedLineItemName.languageId|^|FinancialStatementLineItemValue|^|AdjustedForCorporateActionValue|^|ReportedCurrencyId|^|IsAsReportedCurrencySetManually|^|Unit|^|IsTotal|^|StatementSectionCode|^|DimentionalLineItemId|^|IsDerived|^|EstimateMethodCode|^|EstimateMethodNote|^|EstimateMethodNote.languageId|^|FinancialLineItemSource|^|IsCombinedItem|^|IsExcludedFromStandardization|^|DocByteOffset|^|DocByteLength|^|BookMark|^|ItemDisplayedNegativeFlag|^|ItemScalingFactor|^|ItemDisplayedValue|^|ReportedValue|^|EditedDescription|^|EditedDescription.languageId|^|ReportedDescription|^|ReportedDescription.languageId|^|AsReportedInstanceSequence|^|PhysicalMeasureId|^|FinancialStatementLineItemSequence|^|SystemDerivedTypeCode|^|AsReportedExchangeRate|^|AsReportedExchangeRateSourceCurrencyId|^|ThirdPartySourceCode|^|FinancialStatementLineItemValueUpperRange|^|FinancialStatementLineItemLocalLanguageLabel|^|FinancialStatementLineItemLocalLanguageLabel.languageId|^|IsFinal|^|FinancialStatementLineItem.lineItemInstanceKey|^|StatementSectionIsCredit|^|CapitalChangeAdjustmentDate|^|ParentLineItemId|^|EstimateMethodId|^|StatementSectionId|^|SystemDerivedTypeCodeId|^|UnitEnumerationId|^|FiscalYear|^|IsAnnual|^|PeriodPermId|^|PeriodPermId.objectTypeId|^|PeriodPermId.objectType|^|AuditID|^|AsReportedItemId|^|ExpressionInstanceId|^|ExpressionText|^|FFAction|!|
192730239205|^|235|^|1|^|FTN|^|500186|^|221|^|Average Age of Employees|^|505074|^|30.00000|^||^||^|False|^|1.00000|^|False|^|EMP|^||^|False|^|ARV|^||^|505074|^||^|False|^|False|^||^||^||^||^|0|^||^||^||^|505074|^||^|505074|^||^||^|122880|^|NA|^||^||^|TK |^||^||^|505126|^|True|^|1235002211206722736|^|True|^||^||^|3019656|^|3013652|^|3019679|^|1010066|^|1976|^|True|^||^|1000220295|^||^||^||^||^||^|I|!|
192730239205|^|235|^|1|^|FTN|^|500186|^|498|^|Shareholders' Equity Per Share|^|505074|^|91.37000|^|678.74654|^|500186|^|False|^|1.00000|^|False|^|TAN|^||^|False|^|ARV|^||^|505074|^||^|False|^|False|^||^||^||^||^|0|^||^||^||^|505074|^||^|505074|^||^||^|474880|^|NA|^||^||^|TK |^||^||^|505126|^|True|^|1235004981302988315|^|True|^||^||^|3019656|^|3013751|^|3019679|^|1010066|^|1976|^|True|^||^|1000220295|^||^||^||^||^||^|I|!|
192730239205|^|235|^|1|^|FTN|^|500186|^|500|^|Number of Shares Outstanding at Period End-Common Shares|^|505074|^|90000000.00000|^|12115420.96161|^||^|False|^|1000.00000|^|False|^|TAN|^||^|False|^|ARV|^||^|505074|^||^|False|^|False|^||^||^||^||^|3|^||^||^||^|505074|^||^|505074|^||^||^|499712|^|NA|^||^||^|TK |^||^||^|505126|^|True|^|1235005001178855709|^|True|^||^||^|3019656|^|3013751|^|3019679|^|1010067|^|1976|^|True|^||^|1000220295|^||^||^||^||^||^|I|!|
192730239205|^|235|^|1|^|FTN|^|500186|^|562|^|Number of Employees|^|505074|^|2924.00000|^||^||^|False|^|1.00000|^|False|^|EMP|^||^|False|^|ARV|^||^|505074|^||^|False|^|False|^||^||^||^||^|0|^||^||^||^|505074|^||^|505074|^||^||^|464864|^|NA|^||^||^|TK |^||^||^|505126|^|True|^|1235005621461877526|^|True|^||^||^|3019656|^|3013652|^|3019679|^|1010066|^|1976|^|True|^||^|1000220295|^||^||^||^||^||^|I|!|
192730239205|^|235|^|1|^|FTN|^|500186|^|655|^|Total number of shareholders|^|505074|^|11792.00000|^||^||^|False|^|1.00000|^|False|^|OTH|^||^|False|^|ARV|^||^|505074|^||^|False|^|False|^||^||^||^||^|0|^||^||^||^|505074|^||^|505074|^||^||^|466927|^|NA|^||^||^|TK |^||^||^|505126|^|True|^|1235006551335570418|^|True|^||^||^|3019656|^|3013716|^|3019679|^|1010066|^|1976|^|True|^||^|1000220295|^||^||^||^||^||^|I|!|
192730239205|^|235|^|1|^|FTN|^|500186|^|657|^|Total dividends paid (common stock)|^|505074|^|540000000.00000|^||^|500186|^|False|^|1000000.00000|^|False|^|OTH|^||^|False|^|ARV|^||^|505074|^||^|False|^|False|^||^||^||^||^|6|^||^||^||^|505074|^||^|505074|^||^||^|233463|^|NA|^||^||^|TK |^||^||^|505126|^|True|^|12350065712483219|^|True|^||^||^|3019656|^|3013716|^|3019679|^|1010068|^|1976|^|True|^||^|1000220295|^||^||^||^||^||^|I|!|
192730239205|^|235|^|1|^|FTN|^|500186|^|1452|^|Order received|^|505074|^|26936000000.00000|^||^|500186|^|False|^|1000000.00000|^|False|^|OTH|^||^|False|^|ARV|^||^|505074|^||^|False|^|False|^||^||^||^||^|6|^||^||^||^|505074|^||^|505074|^||^||^|350195|^|NA|^||^||^|TK |^||^||^|505126|^|True|^|1235014521608462544|^|True|^||^||^|3019656|^|3013716|^|3019679|^|1010068|^|1976|^|True|^||^|1000220295|^||^||^||^||^||^|I|!|
192730239205|^|235|^|1|^|FTN|^|500186|^|1453|^|Order backlogs|^|505074|^|1447000000.00000|^||^|500186|^|False|^|1000000.00000|^|False|^|OTH|^||^|False|^|ARV|^||^|505074|^||^|False|^|False|^||^||^||^||^|6|^||^||^||^|505074|^||^|505074|^||^||^|350195|^|NA|^||^||^|TK |^||^||^|505126|^|True|^|1235014531922884465|^|True|^||^||^|3019656|^|3013716|^|3019679|^|1010068|^|1976|^|True|^||^|1000220295|^||^||^||^||^||^|I|!|
192730239205|^|235|^|1|^|FTN|^|500186|^|1457|^|Export amount|^|505074|^|3924000000.00000|^||^|500186|^|False|^|1000000.00000|^|False|^|OTH|^||^|False|^|ARV|^||^|505074|^||^|False|^|False|^||^||^||^||^|6|^||^||^||^|505074|^||^|505074|^||^||^|291829|^|NA|^||^||^|TK |^||^||^|505126|^|True|^|1235014571728332413|^|True|^||^||^|3019656|^|3013716|^|3019679|^|1010068|^|1976|^|True|^||^|1000220295|^||^||^||^||^||^|I|!|
192730239205|^|235|^|1|^|FTN|^|500186|^|1459|^|Capital expenditures (Note)|^|505074|^|659000000.00000|^||^|500186|^|False|^|1000000.00000|^|False|^|OTH|^||^|False|^|ARV|^||^|505074|^||^|False|^|False|^||^||^||^||^|6|^||^||^||^|505074|^||^|505074|^||^||^|350195|^|NA|^||^||^|TK |^||^||^|505126|^|True|^|1235014591148256870|^|True|^||^||^|3019656|^|3013716|^|3019679|^|1010068|^|1976|^|True|^||^|1000220295|^||^||^||^||^||^|I|!|
192730239285|^|236|^|1|^|FTN|^|500186|^|255|^|Number of Employees|^|505074|^|10152.00000|^||^||^|False|^|1.00000|^|False|^|EMP|^||^|False|^|ARV|^||^|505074|^||^|False|^|False|^||^||^||^||^|0|^||^||^||^|505074|^||^|505074|^||^||^|12288|^|NA|^||^||^|TK |^||^||^|505126|^|True|^|1236002551128894330|^|True|^||^||^|3019656|^|3013652|^|3019679|^|1010066|^|1976|^|True|^||^|1000220295|^||^||^||^||^||^|I|!|
192730239285|^|236|^|1|^|FTN|^|500186|^|256|^|Average Age of Employees|^|505074|^|34.00000|^||^||^|False|^|1.00000|^|False|^|EMP|^||^|False|^|ARV|^||^|505074|^||^|False|^|False|^||^||^||^||^|0|^||^||^||^|505074|^||^|505074|^||^||^|122880|^|NA|^||^||^|TK |^||^||^|505126|^|True|^|1236002561111316467|^|True|^||^||^|3019656|^|3013652|^|3019679|^|1010066|^|1976|^|True|^||^|1000220295|^||^||^||^||^||^|I|!|
192730239285|^|236|^|1|^|FTN|^|500186|^|542|^|Shareholders' Equity Per Share|^|505074|^|160.20000|^|691.93184|^|500186|^|False|^|1.00000|^|False|^|TAN|^||^|False|^|ARV|^||^|505074|^||^|False|^|False|^||^||^||^||^|0|^||^||^||^|505074|^||^|505074|^||^||^|471038|^|NA|^||^||^|TK |^||^||^|505126|^|True|^|1236005421170597389|^|True|^||^||^|3019656|^|3013751|^|3019679|^|1010066|^|1976|^|True|^||^|1000220295|^||^||^||^||^||^|I|!|
192730239285|^|236|^|1|^|FTN|^|500186|^|545|^|Number of Shares Outstanding at Period End-Common Shares|^|505074|^|679468000.00000|^|157314300.64243|^||^|False|^|1000.00000|^|False|^|TAN|^||^|False|^|ARV|^||^|505074|^||^|False|^|False|^||^||^||^||^|3|^||^||^||^|505074|^||^|505074|^||^||^|472064|^|NA|^||^||^|TK |^||^||^|505126|^|True|^|1236005451445165969|^|True|^||^||^|3019656|^|3013751|^|3019679|^|1010067|^|1976|^|True|^||^|1000220295|^||^||^||^||^||^|I|!|
192730239285|^|236|^|1|^|FTN|^|500186|^|718|^|Total dividends paid (common stock)|^|505074|^|4750000000.00000|^||^|500186|^|False|^|1000000.00000|^|False|^|OTH|^||^|False|^|ARV|^||^|505074|^||^|False|^|False|^||^||^||^||^|6|^||^||^||^|505074|^||^|505074|^||^||^|458752|^|NA|^||^||^|TK |^||^||^|505126|^|True|^|1236007181118043352|^|True|^||^||^|3019656|^|3013716|^|3019679|^|1010068|^|1976|^|True|^||^|1000220295|^||^||^||^||^||^|I|!|
192730239285|^|236|^|1|^|FTN|^|500186|^|1364|^|Export amount|^|505074|^|15379000000.00000|^||^|500186|^|False|^|1000000.00000|^|False|^|OTH|^||^|False|^|ARV|^||^|505074|^||^|False|^|False|^||^||^||^||^|6|^||^||^||^|505074|^||^|505074|^||^||^|459752|^|NA|^||^||^|TK |^||^||^|505126|^|True|^|1236013641649895533|^|True|^||^||^|3019656|^|3013716|^|3019679|^|1010068|^|1976|^|True|^||^|1000220295|^||^||^||^||^||^|I|!|
192730239285|^|236|^|1|^|FTN|^|500186|^|1407|^|Total number of shareholders|^|505074|^|57288.00000|^||^||^|False|^|1.00000|^|False|^|OTH|^||^|False|^|ARV|^||^|505074|^||^|False|^|False|^||^||^||^||^|0|^||^||^||^|505074|^||^|505074|^||^||^|460752|^|NA|^||^||^|TK |^||^||^|505126|^|True|^|1236014071623011361|^|True|^||^||^|3019656|^|3013716|^|3019679|^|1010066|^|1976|^|True|^||^|1000220295|^||^||^||^||^||^|I|!|

The structure of the second data set is also same

I am performing on first 5 columns . As you can see the combination of all first 5 columns does not provide me enough partition and that leads to data skew .

The spark job stuck on some of the Executor .

The size of the first dataset is 270 GB and second is 5 GB but expected to increase .

Total no of partition 1128

This is how I perform my join

val dfMainOutput = (dataMain.join(latestForEachKey, Seq("uniqueFundamentalSet", "PeriodId", "SourceId", "StatementTypeCode", "StatementCurrencyId", "FinancialStatementLineItem_lineItemId"), "outer") select (exprsExtended: _*)).filter(!$"FFAction|!|".contains("D|!|"))

I tried implementing Broadcast Join but no impact .

So in this case can I use salting or hashing on join key so that the joining key will become random and skew will not occur I guess .

Here is my query and app details

Here is the cluster details when we are loading the data .

And here is cluster details when most of the container is idle.

Adding the details of the task where some are 10 and on some executor only 3 to 4 .

Hello @Atharv, first of all a quick note, if possible to apply filter `!$"FFAction|!|".contains("D|!|"` before join. Also, could post some more information over your cluster size and DAG execution plan? — abiratsis, Mar 26 '18 at 12:56
@AlexandrosBiratsis this check is required because only after joining we are removing the rows from all rows.. Added the details also — Atharv Thakur, Mar 26 '18 at 13:20
How many executors you have @Atharv and how much memory for each one of them? If you can fit the 2nd dataset to cluster memory try to use persist(cache) after retrieving it. That could significantly improve your performance although depend if the dataset can fit into cluster memory. — abiratsis, Mar 26 '18 at 14:01
also you can try to repartition to increase the number of partitions in order to face skewing and improve balance of the keys distribution — abiratsis, Mar 26 '18 at 14:05
@AlexandrosBiratsis this Is my configuration `--num-executors 60 --conf spark.yarn.executor.memoryOverhead=9216 --executor-memory 72G --conf spark.yarn.driver.memoryOverhead=3072 --driver-memory 26G --executor-cores 10 --driver-cores 3 --conf spark.default.parallelism=1200` ...I have second data set of size from 10 MB to 10GB also .. — Atharv Thakur, Mar 26 '18 at 14:09
I will try to perform cache second data set and will run again .. — Atharv Thakur, Mar 26 '18 at 14:10
one more thing @Atharv :) please post a photo of the skewed data — abiratsis, Mar 26 '18 at 14:11
@AlexandrosBiratsis thanks for your suggestion ...But how can I exactly find the skewed data pic I am guessing data is skewed because the pattern is same and some executors takes very long time ..Sorry I am new to this so asking simple questions — Atharv Thakur, Mar 26 '18 at 14:14
You can see it on your task details page how many tasks are executed on each executor. They should be equally distributed if not then you have one more symptom on the same page would be some tasks last much longer than other — abiratsis, Mar 26 '18 at 14:31
Hello @AlexandrosBiratsis I have added the task details also ...This is the task details after adding the cache in the smaller data set .. — Atharv Thakur, Mar 26 '18 at 15:59
Let us [continue this discussion in chat](https://chat.stackoverflow.com/rooms/167577/discussion-between-alexandros-biratsis-and-atharv-thakur). — abiratsis, Mar 26 '18 at 17:24

score 5 · Answer 1 · answered Mar 27 '18 at 07:39

Please consider the following points:

1) Since you have 60 executors and 10 cores per executor your partitions should be at least 60 x 10 = 600 partitions

2) In your case you have 270GB / 1128 ~ 241MB this should approximately be the partition size which looks quite big to me (considering data exchange during shuffling). Try first to re-partition to something more realistic for instance 8K or even 16K.

3) Since I can not see clearly how many executors participate on job execution you need to check it again and figure out the exact number of participating executors and if data is equally distributed. If data deviation between executors is low then your data is well distributed otherwise you face skewing.

4) If after re-partition skewing insists try to redistribute the join keys as described here

@Atharv I think would be better 1) to post the complete code 2) try to figure out how many executors are being involved in total 3) try to repartition with more partitions e.g 8192 partitions using repartition method of Dataset class check Spark API [here](https://spark.apache.org/docs/2.2.0/api/java/index.html?org/apache/spark/sql/Dataset.html) 4) use repartition by specifying columns as well eg df.repartition(8192, col("c1"), col("c2")) the columns here should those you use in your join statement — abiratsis, Mar 27 '18 at 11:31

How to handle data skew in the spark data frame for outer join

1 Answers1

Linked