[apache-spark] Spark - repartition() vs coalesce()

According to Learning Spark

Keep in mind that repartitioning your data is a fairly expensive operation. Spark also has an optimized version of repartition() called coalesce() that allows avoiding data movement, but only if you are decreasing the number of RDD partitions.

One difference I get is that with repartition() the number of partitions can be increased/decreased, but with coalesce() the number of partitions can only be decreased.

If the partitions are spread across multiple machines and coalesce() is run, how can it avoid data movement?

Justin's answer is awesome and this response goes into more depth.

The repartition algorithm does a full shuffle and creates new partitions with data that's distributed evenly. Let's create a DataFrame with the numbers from 1 to 12.

val x = (1 to 12).toList
val numbersDf = x.toDF("number")

numbersDf contains 4 partitions on my machine.

numbersDf.rdd.partitions.size // => 4

Here is how the data is divided on the partitions:

Partition 00000: 1, 2, 3
Partition 00001: 4, 5, 6
Partition 00002: 7, 8, 9
Partition 00003: 10, 11, 12

Let's do a full-shuffle with the repartition method and get this data on two nodes.

val numbersDfR = numbersDf.repartition(2)

Here is how the numbersDfR data is partitioned on my machine:

Partition A: 1, 3, 4, 6, 7, 9, 10, 12
Partition B: 2, 5, 8, 11

The repartition method makes new partitions and evenly distributes the data in the new partitions (the data distribution is more even for larger data sets).

Difference between coalesce and repartition

coalesce uses existing partitions to minimize the amount of data that's shuffled. repartition creates new partitions and does a full shuffle. coalesce results in partitions with different amounts of data (sometimes partitions that have much different sizes) and repartition results in roughly equal sized partitions.

Is coalesce or repartition faster?

coalesce may run faster than repartition, but unequal sized partitions are generally slower to work with than equal sized partitions. You'll usually need to repartition datasets after filtering a large data set. I've found repartition to be faster overall because Spark is built to work with equal sized partitions.

N.B. I've curiously observed that repartition can increase the size of data on disk. Make sure to run tests when you're using repartition / coalesce on large datasets.

Read this blog post if you'd like even more details.

When you'll use coalesce & repartition in practice

• See this question on how to use coalesce & repartition to write out a DataFrame to a single file
• It's critical to repartition after running filtering queries. The number of partitions does not change after filtering, so if you don't repartition, you'll have way too many memory partitions (the more the filter reduces the dataset size, the bigger the problem). Watch out for the empty partition problem.
• partitionBy is used to write out data in partitions on disk. You'll need to use repartition / coalesce to partition your data in memory properly before using partitionBy.

In a simple way COALESCE :- is only for decreases the no of partitions , No shuffling of data it just compress the partitions

REPARTITION:- is for both increase and decrease the no of partitions , But shuffling takes place

Example:-

val rdd = sc.textFile("path",7)
rdd.repartition(10)
rdd.repartition(2)

Both works fine

But we go generally for this two things when we need to see output in one cluster,we go with this.

To all the great answers I would like to add that repartition is one the best option to take advantage of data parallelization. While coalesce gives a cheap option to reduce the partitions and it is very useful when writing data to HDFS or some other sink to take advantage of big writes.

I have found this useful when writing data in parquet format to get full advantage.

Also another difference is taking into consideration a situation where there is a skew join and you have to coalesce on top of it. A repartition will solve the skew join in most cases, then you can do the coalesce.

Another situation is, suppose you have saved a medium/large volume of data in a data frame and you have to produce to Kafka in batches. A repartition helps to collectasList before producing to Kafka in certain cases. But, when the volume is really high, the repartition will likely cause serious performance impact. In that case, producing to Kafka directly from dataframe would help.

side notes: Coalesce does not avoid data movement as in full data movement between workers. It does reduce the number of shuffles happening though. I think that's what the book means.

Repartition: Shuffle the data into a NEW number of partitions.

Eg. Initial data frame is partitioned in 200 partitions.

df.repartition(500): Data will be shuffled from 200 partitions to new 500 partitions.

Coalesce: Shuffle the data into existing number of partitions.

df.coalesce(5): Data will be shuffled from remaining 195 partitions to 5 existing partitions.

But also you should make sure that, the data which is coming coalesce nodes should have highly configured, if you are dealing with huge data. Because all the data will be loaded to those nodes, may lead memory exception. Though reparation is costly, i prefer to use it. Since it shuffles and distribute the data equally.

Be wise to select between coalesce and repartition.

I would like to add to Justin and Power's answer that -

repartition will ignore existing partitions and create new ones. So you can use it to fix data skew. You can mention partition keys to define the distribution. Data skew is one of the biggest problems in the 'big data' problem space.

coalesce will work with existing partitions and shuffle a subset of them. It can't fix the data skew as much as repartition does. Therefore even if it is less expensive it might not be the thing you need.

The repartition algorithm does a full shuffle of the data and creates equal sized partitions of data. coalesce combines existing partitions to avoid a full shuffle.

Coalesce works well for taking an RDD with a lot of partitions and combining partitions on a single worker node to produce a final RDD with less partitions.

Repartition will reshuffle the data in your RDD to produce the final number of partitions you request. The partitioning of DataFrames seems like a low level implementation detail that should be managed by the framework, but it’s not. When filtering large DataFrames into smaller ones, you should almost always repartition the data. You’ll probably be filtering large DataFrames into smaller ones frequently, so get used to repartitioning.

Read this blog post if you'd like even more details.

What follows from the code and code docs is that coalesce(n) is the same as coalesce(n, shuffle = false) and repartition(n) is the same as coalesce(n, shuffle = true)

Thus, both coalesce and repartition can be used to increase number of partitions

With shuffle = true, you can actually coalesce to a larger number of partitions. This is useful if you have a small number of partitions, say 100, potentially with a few partitions being abnormally large.

Another important note to accentuate is that if you drastically decrease number of partitions you should consider using shuffled version of coalesce (same as repartition in that case). This will allow your computations be performed in parallel on parent partitions (multiple task).

However, if you're doing a drastic coalesce, e.g. to numPartitions = 1, this may result in your computation taking place on fewer nodes than you like (e.g. one node in the case of numPartitions = 1). To avoid this, you can pass shuffle = true. This will add a shuffle step, but means the current upstream partitions will be executed in parallel (per whatever the current partitioning is).

Please also refer to the related answer here

All the answers are adding some great knowledge into this very often asked question.

So going by tradition of this question's timeline, here are my 2 cents.

I found the repartition to be faster than coalesce, in very specific case.

In my application when the number of files that we estimate is lower than the certain threshold, repartition works faster.

Here is what I mean

if(numFiles > 20)
    df.coalesce(numFiles).write.mode(SaveMode.Overwrite).parquet(dest)
else
    df.repartition(numFiles).write.mode(SaveMode.Overwrite).parquet(dest)

In above snippet, if my files were less than 20, coalesce was taking forever to finish while repartition was much faster and so the above code.

Of course, this number (20) will depend on the number of workers and amount of data.

Hope that helps.

For someone who had issues generating a single csv file from PySpark (AWS EMR) as an output and saving it on s3, using repartition helped. The reason being, coalesce cannot do a full shuffle, but repartition can. Essentially, you can increase or decrease the number of partitions using repartition, but can only decrease the number of partitions (but not 1) using coalesce. Here is the code for anyone who is trying to write a csv from AWS EMR to s3:

df.repartition(1).write.format('csv')\
.option("path", "s3a://my.bucket.name/location")\
.save(header = 'true')

There is a use-case for repartition >> coalesce even where the partition number decreases mentioned in @Rob's answer, that is writing data to a single file.

@Rob's answer hints in the good direction, but I think that some further explanation is needed to understand what's going on under the hood.

If you need to filter your data before writing, then repartition is much more suitable than coalesce, since coalesce will be pushed-down right before the loading operation.

For instance: load().map(…).filter(…).coalesce(1).save()

translates to: load().coalesce(1).map(…).filter(…).save()

This means that all your data will collapse into a single partition, where it will be filtered, losing all parallelism. This happens even for very simple filters like column='value'.

This does not happen with repartition: load().map(…).filter(…).repartition(1).save()

In such case, filtering happens in parallel on the original partitions.

Just to give an order of magnitude, in my case when filtering 109M rows (~105G) with ~1000 partitions after loading from a Hive table, the runtime dropped from the ~6h for coalesce(1) to ~2m for repartition(1).

The specific example is taken from this article from AirBnB, which is pretty good and covers even more aspects of repartitioning techniques in Spark.

repartition - it's recommended to use it while increasing the number of partitions, because it involve shuffling of all the data.

coalesce - it's is recommended to use it while reducing the number of partitions. For example if you have 3 partitions and you want to reduce it to 2, coalesce will move the 3rd partition data to partition 1 and 2. Partition 1 and 2 will remains in the same container. On the other hand, repartition will shuffle data in all the partitions, therefore the network usage between the executors will be high and it will impacts the performance.

coalesce performs better than repartition while reducing the number of partitions.

One additional point to note here is that, as the basic principle of Spark RDD is immutability. The repartition or coalesce will create new RDD. The base RDD will continue to have existence with its original number of partitions. In case the use case demands to persist RDD in cache, then the same has to be done for the newly created RDD.

scala> pairMrkt.repartition(10)
res16: org.apache.spark.rdd.RDD[(String, Array[String])] =MapPartitionsRDD[11] at repartition at <console>:26

scala> res16.partitions.length
res17: Int = 10

scala>  pairMrkt.partitions.length
res20: Int = 2

Basically Repartition allows you to increase or decrease the number of partitions. Repartition re-distributes the data from all the partitions and this leads to full shuffle which is very expensive operation.

Coalesce is the optimized version of Repartition where you can only reduce the number of partitions. As we are only able to reduce the number of partitions what it does is merge some of the partitions to be a single partition. By merging partitions, the movement of the data across the partition is lower compared to Repartition. So in Coalesce is minimum data movement but saying that coalesce does not do data movement is completely wrong statement.

Other thing is in repartition by providing the number of partitions, it tries to redistribute the data uniformly on all the partitions while in case of Coalesce we could still have skew data in some cases.