Compute size of Spark dataframe - SizeEstimator gives unexpected results
Solution 1
Unfortunately, I was not able to get reliable estimates from SizeEstimator
, but I could find another strategy - if the dataframe is cached, we can extract its size from queryExecution
as follows:
df.cache.foreach(_ => ())
val catalyst_plan = df.queryExecution.logical
val df_size_in_bytes = spark.sessionState.executePlan(
catalyst_plan).optimizedPlan.stats.sizeInBytes
For the example dataframe, this gives exactly 4.8GB (which also corresponds to the file size when writing to an uncompressed Parquet table).
This has the disadvantage that the dataframe needs to be cached, but it is not a problem in my case.
EDIT: Replaced df.cache.foreach(_=>_)
by df.cache.foreach(_ => ())
, thanks to @DavidBenedeki for pointing it out in the comments.
Solution 2
Apart from Size estimator, which you have already tried(good insight)..
below is another option
RDDInfo[] getRDDStorageInfo()
Return information about what RDDs are cached, if they are in mem or on both, how much space they take, etc.
actually spark storage tab uses this.Spark docs
Below is the implementation from spark
/**
* :: DeveloperApi ::
* Return information about what RDDs are cached, if they are in mem or on disk, how much space
* they take, etc.
*/
@DeveloperApi
def getRDDStorageInfo: Array[RDDInfo] = {
getRDDStorageInfo(_ => true)
}
private[spark] def getRDDStorageInfo(filter: RDD[_] => Boolean): Array[RDDInfo] = {
assertNotStopped()
val rddInfos = persistentRdds.values.filter(filter).map(RDDInfo.fromRdd).toArray
rddInfos.foreach { rddInfo =>
val rddId = rddInfo.id
val rddStorageInfo = statusStore.asOption(statusStore.rdd(rddId))
rddInfo.numCachedPartitions = rddStorageInfo.map(_.numCachedPartitions).getOrElse(0)
rddInfo.memSize = rddStorageInfo.map(_.memoryUsed).getOrElse(0L)
rddInfo.diskSize = rddStorageInfo.map(_.diskUsed).getOrElse(0L)
}
rddInfos.filter(_.isCached)
}
yourRDD.toDebugString
from RDD also uses this. code here
General Note :
In my opinion, to get optimal number of records in each partition and check your repartition is correct and they are uniformly distributed, I would suggest to try like below... and adjust your re-partition number. and then measure the size of partition... would be more sensible. to address this kind of problems
yourdf.rdd.mapPartitionsWithIndex{case (index,rows) => Iterator((index,rows.size))}
.toDF("PartitionNumber","NumberOfRecordsPerPartition")
.show
or existing spark functions (based on spark version)
import org.apache.spark.sql.functions._
df.withColumn("partitionId", sparkPartitionId()).groupBy("partitionId").count.show
Solution 3
SizeEstimator
returns the number of bytes an object takes up on the JVM heap. This includes objects referenced by the object, the actual object size will almost always be much smaller.
The discrepancies in sizes you've observed are because when you create new objects on the JVM the references take up memory too, and this is being counted.
Check out the docs here 🤩
https://spark.apache.org/docs/2.2.0/api/scala/index.html#org.apache.spark.util.SizeEstimator$
hiryu
Updated on July 09, 2022Comments
-
hiryu almost 2 years
I am trying to find a reliable way to compute the size (in bytes) of a Spark dataframe programmatically.
The reason is that I would like to have a method to compute an "optimal" number of partitions ("optimal" could mean different things here: it could mean having an optimal partition size, or resulting in an optimal file size when writing to Parquet tables - but both can be assumed to be some linear function of the dataframe size). In other words, I would like to call
coalesce(n)
orrepartition(n)
on the dataframe, wheren
is not a fixed number but rather a function of the dataframe size.Other topics on SO suggest using
SizeEstimator.estimate
fromorg.apache.spark.util
to get the size in bytes of the dataframe, but the results I'm getting are inconsistent.First of all, I'm persisting my dataframe to memory:
df.cache().count
The Spark UI shows a size of 4.8GB in the Storage tab. Then, I run the following command to get the size from
SizeEstimator
:import org.apache.spark.util.SizeEstimator SizeEstimator.estimate(df)
This gives a result of 115'715'808 bytes =~ 116MB. However, applying
SizeEstimator
to different objects leads to very different results. For instance, I try computing the size separately for each row in the dataframe and sum them:df.map(row => SizeEstimator.estimate(row.asInstanceOf[ AnyRef ])).reduce(_+_)
This results in a size of 12'084'698'256 bytes =~ 12GB. Or, I can try to apply
SizeEstimator
to every partition:df.mapPartitions( iterator => Seq(SizeEstimator.estimate( iterator.toList.map(row => row.asInstanceOf[ AnyRef ]))).toIterator ).reduce(_+_)
which results again in a different size of 10'792'965'376 bytes =~ 10.8GB.
I understand there are memory optimizations / memory overhead involved, but after performing these tests I don't see how
SizeEstimator
can be used to get a sufficiently good estimate of the dataframe size (and consequently of the partition size, or resulting Parquet file sizes).What is the appropriate way (if any) to apply
SizeEstimator
in order to get a good estimate of a dataframe size or of its partitions? If there isn't any, what is the suggested approach here? -
hiryu about 6 yearsThanks, I have seen the docs but tbh they confuse me even more. See for instance the 1st example where I apply the
estimate
method directly to the dataframe - the resulting size (116MB) is only a small fraction of the size shown in the UI when caching the object (4.8GB), not more. Moreover, the doc forestimate
states thatThis is useful for determining (...) or the amount of space each object will take when caching objects in deserialized form.
However, I don't see that happening in my tests... -
Saiteja Parsi over 4 yearscan you please help with corresponding method in java? I am having hard time figuring it out. I tried below code. Its gives incorrect values.
LogicalPlan logicalPlan = dataSet.queryExecution().optimizedPlan(); Statistics stats = sparkSession.sessionState().executePlan(logicalPlan).analyzed().stats(dataSet.sqlContext().conf());
-
hiryu over 4 yearsI'm not familiar with the Java API, but in your code, when you assign
logicalPlan
, it seems you are usingoptimizedPlan()
, which is a different plan. In my code, I access thelogical
property and extract it from there. -
Bob almost 4 yearsHow would I do this in pyspark?
-
ss301 over 3 yearsthis is giving me a really big unrealistic figure.
-
hiryu over 3 yearsFor pyspark, you have to access the hidden
_jdf
andjSparkSession
variables, since the Python objects do not expose the needed attributes directly...df.cache().foreach(lambda x: x)
catalyst_plan = df._jdf.queryExecution().logical()
spark._jsparkSession.sessionState().executePlan(catalyst_plan).optimizedPlan().stats().sizeInBytes()
-
thentangler about 3 yearscan this be used in pyspark?
-
hiryu about 3 yearssee the comment just before yours, that code is in pyspark.
-
RafalK almost 3 yearsJust wondering why caching is the 'must' here?.. Could you please elaborate more on this? Thank you in advance.
-
hiryu almost 3 yearsAs you see, we are extracting the byte size of the dataframe from the stats of the optimized plan (more on that: jaceklaskowski.gitbooks.io/mastering-spark-sql/content/…). These stats are usually just an estimation. However, when we cache the dataset, the optimized plan realizes that it can simply access the dataframe in memory, hence the stats become exact. The
.foreach(_=>_)
bit is only an action that forces execution and hence the caching. -
scravy almost 3 years@hiryu should it also work to do a
localCheckpoint()
before? -
David Benedeki over 2 yearsLittle fix for the cache line, as
foreach
doesn't have a return type:df.cache.foreach(_ => Unit)
-
Emma Y over 2 yearsdf.count() is action, why need to use df.cache.foreach()?
-
hiryu over 2 years@DavidBenedeki thanks, not sure why it worked previously? however, it should be
df.cache.foreach(_ => ())
. -
hiryu over 2 years@EmmaY the reason I don't use
count
is that, because of Spark Catalyst optimizations,count
is not guaranteed to materialize all rows: stackoverflow.com/questions/42714291/… Now according to the same link, using it in conjunction with cache should actually materialize the data... but I didn't want to take any chance.foreach
will always materialize the dataframe, full stop. -
hiryu over 2 years@scravy I never used
localCheckpoint
, so you will need to test for yourself. -
Emma Y over 2 years@hiryu. Thanks for sharing, but from this link, it seems that df.rdd.count() or df.cache().count() shall play the trick?
-
hiryu over 2 yearsAs long as the action called after
cache
guarantees the materialization of all rows and columns, it will do the trick. However, I could ask you the reverse question: what's wrong with usingforeach
? It guarantees materialization of all rows, while there are legit concerns withdf.count
- the link tells us that addingcache
is sufficient, but is it confirmed by official docs/code? Moreover,foreach
does not return anything, so it won't pollute stdout (the call tocount
will) and it will be marginally be more efficient (no need to send local counts to the driver for aggregation). -
Carlos Eduardo Bilar Rodrigues over 2 yearshow to use this : df.cache.foreach(_ => ()) val catalyst_plan = df.queryExecution.logical val df_size_in_bytes = spark.sessionState.executePlan( catalyst_plan).optimizedPlan.stats.sizeInBytes In Pyspark?
-
hiryu over 2 years@CarlosEduardoBilarRodrigues see comment from Nov 17 2020.
-
Carlos Eduardo Bilar Rodrigues over 2 years@Hiryu, I tried it, but it crashes when I use df.cache.foreach (_ => ()), it only works with df.cache(), but the result is not satisfactory
-
hiryu over 2 yearsif you look in the comment I pointed you to (Nov 17 2020), it uses
.foreach(lambda x: x)
, not.foreach(_ => ())
. The former is Python code, the latter is Scala. Ifforeach(lambda x: x)
does not work (but it should, just tested), try.foreach(lambda x: None)
. -
Carlos Eduardo Bilar Rodrigues over 2 years@hiryu, this process returns a very large number for a small group of data, can you tell me why? example 9MB translates a number of 100 characters
-
hiryu over 2 yearsyou will need to post a separate question with a minimal reproducible example to get more help, you are not providing enough context. What is described in the answer is a workaround that served me well in the past, but it's by no means an official solution nor a widely tested one.
-
Des0lat0r about 2 yearsI would like to ask how it would be in Java. Would it be correct to write df.cache(); System.out.println(df.logicalPlan().stats().sizeInBytes()) ??