zhengruifeng edited a comment on pull request #34367:
URL: https://github.com/apache/spark/pull/34367#issuecomment-949516811
a simple skewed window example:
```
import org.apache.spark.sql.expressions.Window
val df1 = spark.range(0, 100000000, 1, 9).select(when('id < 90000000,
123).otherwise('id).as("key1"), 'id as "value1").withColumn("hash1",
abs(hash(col("key1"))).mod(1000))
df1.withColumn("rank",
row_number().over(Window.partitionBy("hash1").orderBy("value1"))).where(col("rank")
<= 1).write.mode("overwrite").parquet("/tmp/tmp1")
spark.conf.set("spark.sql.rankLimit.enabled", "true")
df1.withColumn("rank",
row_number().over(Window.partitionBy("hash1").orderBy("value1"))).where(col("rank")
<= 1).write.mode("overwrite").parquet("/tmp/tmp2")
```
existing plan took 60 sec, while the new plan with `RankLimit` took only 8
sec.
and the shuffle write was reduced from 544.9 MiB to 26.7 KiB
```
== Physical Plan ==
Execute InsertIntoHadoopFsRelationCommand (20)
+- AdaptiveSparkPlan (19)
+- == Final Plan ==
* Filter (12)
+- Window (11)
+- RankLimit (10)
+- * Sort (9)
+- AQEShuffleRead (8)
+- ShuffleQueryStage (7)
+- Exchange (6)
+- RankLimit (5)
+- * Sort (4)
+- * Project (3)
+- * Project (2)
+- * Range (1)
+- == Initial Plan ==
Filter (18)
+- Window (17)
+- RankLimit (16)
+- Sort (15)
+- Exchange (14)
+- RankLimit (13)
+- Sort (4)
+- Project (3)
+- Project (2)
+- Range (1)
(1) Range [codegen id : 1]
Output [1]: [id#0L]
Arguments: Range (0, 100000000, step=1, splits=Some(9))
(2) Project [codegen id : 1]
Output [2]: [CASE WHEN (id#0L < 90000000) THEN 123 ELSE id#0L END AS
key1#2L, id#0L AS value1#3L]
Input [1]: [id#0L]
(3) Project [codegen id : 1]
Output [3]: [key1#2L, value1#3L, (abs(hash(key1#2L, 42), false) % 1000) AS
hash1#6]
Input [2]: [key1#2L, value1#3L]
(4) Sort [codegen id : 1]
Input [3]: [key1#2L, value1#3L, hash1#6]
Arguments: [hash1#6 ASC NULLS FIRST], false, 0
(5) RankLimit
Input [3]: [key1#2L, value1#3L, hash1#6]
Arguments: [hash1#6], [value1#3L ASC NULLS FIRST], row_number(), 1, Partial
(6) Exchange
Input [3]: [key1#2L, value1#3L, hash1#6]
Arguments: hashpartitioning(hash1#6, 200), ENSURE_REQUIREMENTS, [id=#119]
(7) ShuffleQueryStage
Output [3]: [key1#2L, value1#3L, hash1#6]
Arguments: 0
(8) AQEShuffleRead
Input [3]: [key1#2L, value1#3L, hash1#6]
Arguments: coalesced
(9) Sort [codegen id : 2]
Input [3]: [key1#2L, value1#3L, hash1#6]
Arguments: [hash1#6 ASC NULLS FIRST], false, 0
(10) RankLimit
Input [3]: [key1#2L, value1#3L, hash1#6]
Arguments: [hash1#6], [value1#3L ASC NULLS FIRST], row_number(), 1, Final
(11) Window
Input [3]: [key1#2L, value1#3L, hash1#6]
Arguments: [row_number() windowspecdefinition(hash1#6, value1#3L ASC NULLS
FIRST, specifiedwindowframe(RowFrame, unboundedpreceding$(), currentrow$())) AS
rank#21], [hash1#6], [value1#3L ASC NULLS FIRST]
(12) Filter [codegen id : 3]
Input [4]: [key1#2L, value1#3L, hash1#6, rank#21]
Condition : (rank#21 <= 1)
(13) RankLimit
Input [3]: [key1#2L, value1#3L, hash1#6]
Arguments: [hash1#6], [value1#3L ASC NULLS FIRST], row_number(), 1, Partial
(14) Exchange
Input [3]: [key1#2L, value1#3L, hash1#6]
Arguments: hashpartitioning(hash1#6, 200), ENSURE_REQUIREMENTS, [id=#100]
(15) Sort
Input [3]: [key1#2L, value1#3L, hash1#6]
Arguments: [hash1#6 ASC NULLS FIRST], false, 0
(16) RankLimit
Input [3]: [key1#2L, value1#3L, hash1#6]
Arguments: [hash1#6], [value1#3L ASC NULLS FIRST], row_number(), 1, Final
(17) Window
Input [3]: [key1#2L, value1#3L, hash1#6]
Arguments: [row_number() windowspecdefinition(hash1#6, value1#3L ASC NULLS
FIRST, specifiedwindowframe(RowFrame, unboundedpreceding$(), currentrow$())) AS
rank#21], [hash1#6], [value1#3L ASC NULLS FIRST]
(18) Filter
Input [4]: [key1#2L, value1#3L, hash1#6, rank#21]
Condition : (rank#21 <= 1)
(19) AdaptiveSparkPlan
Output [4]: [key1#2L, value1#3L, hash1#6, rank#21]
Arguments: isFinalPlan=true
(20) Execute InsertIntoHadoopFsRelationCommand
Input [4]: [key1#2L, value1#3L, hash1#6, rank#21]
Arguments: file:/tmp/tmp2, false, Parquet, [path=/tmp/tmp2], Overwrite,
[key1, value1, hash1, rank]
```
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
