I am reading some data from cloud storage in parquet format into Spark DataFrame using PySpark (~10 executors and 4-5 cores per executor). The spark.sql.files.maxPartitionBytes is set before loading so that I could control the size of each partition and in turn # of partitions I could get. After loading in the data, I apply Spark functions/UDFs on the data. So there should be no shuffle happening (join and groupby). I am expecting that each partition has relatively equal partition size after loading in but the reality is the data loaded is very highly skewed.
When looking in YARN, the min, 25%, median, 75% partition size are all 21 B (basically empty partition) while the max partition size is say 100 MB where all rows are loaded.
What I am doing now to solve this skew is to do a df.repartition() to evenly distribute it right after loading and it works well. But this introduces one more shuffle of the data which is not ideal.
So the question is why the partitions are so highly skewed by default after loading? Is there a way that I could load them in with relatively even partition size and skip that df.repartition() step?
Thanks!
Related
I am reading a table from cassandra table in spark. I have big partition in cassandra and when partition size of cassandra exceeds 64 MB , in that case cassandra partition is going to be equal to spark partition. Due to large partition I am getting memory issues in spark.
My question is if I do repartition at the beginning after reading data from cassandra, would number of spark partitions change ? and would it not lead to spark memory issues ?
My assumption is at very first place spark would read data from cassandra and hence at this stage cassandra large partition won't split due to repartition . Repartition will work on underlying data loaded from cassandra.
I am just wondering for answer if repartition could change data distribution when reading data from spark , rather than doing partitioning again ?
If you repartition your data using some arbitrary key then yes, it will be redistributed among the Spark partitions.
Technically, Cassandra partitions do not get split into Spark partitions when you retrieve the data but once you're done reading, you can repartition on a different key to break up the rows of a large Cassandra partition.
For the record, it doesn't avoid the memory issues of reading large Cassandra partitions in the first place because the default input split size of 64MB is just a notional target that Spark uses to calculate how many Spark partitions are required based on the estimated Cassandra table size and C* partition sizes. But since the calculation is based on estimates, the Spark partitions don't actually end up being 64MB in size.
If you are interested, I've explained in detail how Spark partitions are calculated in this post -- https://community.datastax.com/questions/11500/.
To illustrate with an example, let's say that based on the estimated table size and estimated number of C* partitions, each Spark partition is mapped to 200 token ranges in Cassandra.
For the first Spark partition, the token range might only contain 2 Cassandra partitions of size 3MB and 15MB so the actual size of the data in Sthe park partition is just 18MB.
But in the next Spark partition, the token range contains 28 Cassandra partitions that are mostly 1 to 4MB but there is one partition that is 56MB. The total size of this Spark partition ends up being a lot more than 64MB.
In these 2 cases, one Spark partition was just 18MB in size while the other is bigger than the 64MB target size. I've explained this issue in a bit more detail in this post -- https://community.datastax.com/questions/11565/. Cheers!
I am looking through spark partitioning and I see different answers for the question.
Is spark partition size is equal to HDFS block size or depends on the number of cores available on all executors?, and Does the performance improves by repartitioning the data in skewed data case? (I assume the data related to the same join key is again shuffled back to a single executor during the join). Please help me understand this. Thanks!
It really depends on your data where from you are reading. If you are reading from HDFS, then one block will be one partition. But if you are reading a parquet file, then one parquet file is one partition as it is not splittable, so depending on the block in case of HDFS and files count in case of parquet, it creates partitions.
Regarding the skewed data, the more data one partition has, the more time it takes to finish the execution. The other tasks will finished quickly as they have less data so the resources are not being utilized properly. Therefore, it is always better to repartition the skewed data properly, so all executors can evenly do the execution.
You can look here for all the available RDDs, and how they are creating partitions:
https://github.com/apache/spark/tree/master/core/src/main/scala/org/apache/spark/rdd
I have an Spark application that keeps running out of memory, the cluster has two nodes with around 30G of RAM, and the input data size is about few hundreds of GBs.
The application is a Spark SQL job, it reads data from HDFS and create a table and cache it, then do some Spark SQL queries and writes the result back to HDFS.
Initially I split the data into 64 partitions and I got OOM, then I was able to fix the memory issue by using 1024 partitions. But why using more partitions helped me solve the OOM issue?
The solution to big data is partition(divide and conquer). Since not all data could be fit into the memory, and it also could not be processed in a single machine.
Each partition could fit into memory and processed(map) in relative short time. After the data is processed for each partition. It need be merged (reduce). This is tradition map reduce
Splitting data to more partitions means that each partition getting smaller.
[Edit]
Spark using revolution concept called Resilient Distributed DataSet(RDD).
There are two types of operations, transformation and acton
Transformations are mapping from one RDD to another. It is lazy evaluated. Those RDD could be treated as intermediate result we don't wanna get.
Actions is used when you really want get the data. Those RDD/data could be treated as what we want it, like take top failing.
Spark will analysed all the operation and create a DAG(Directed Acyclic Graph) before execution.
Spark start compute from source RDD when actions are fired. Then forget it.
(source: cloudera.com)
I made a small screencast for a presentation on Youtube Spark Makes Big Data Sparking.
Spark's operators spill data to disk if it does not fit in memory,
allowing it to run well on any sized data". The issue with large
partitions generating OOM
Partitions determine the degree of parallelism. Apache Spark doc says that, the partitions size should be atleast equal to the number of cores in the cluster.
Less partitions results in
Less concurrency,
Increase memory pressure for transformation which involves shuffle
More susceptible for data skew.
Many partitions might also have negative impact
Too much time spent in scheduling multiple tasks
Storing your data on HDFS, it will be partitioned already in 64 MB or 128 MB blocks as per your HDFS configuration When reading HDFS files with spark, the number of DataFrame partitions df.rdd.getNumPartitions depends on following properties
spark.default.parallelism (Cores available for the application)
spark.sql.files.maxPartitionBytes (default 128MB)
spark.sql.files.openCostInBytes (default 4MB)
Links :
https://spark.apache.org/docs/latest/tuning.html
https://databricks.com/session/a-deeper-understanding-of-spark-internals
https://spark.apache.org/faq.html
During Spark Summit Aaron Davidson gave some tips about partitions tuning. He also defined a reasonable number of partitions resumed to below 3 points:
Commonly between 100 and 10000 partitions (note: two below points are more reliable because the "commonly" depends here on the sizes of dataset and the cluster)
lower bound = at least 2*the number of cores in the cluster
upper bound = task must finish within 100 ms
Rockie's answer is right, but he does't get the point of your question.
When you cache an RDD, all of his partitions are persisted (in term of storage level) - respecting spark.memory.fraction and spark.memory.storageFraction properties.
Besides that, in an certain moment Spark can automatically drop's out some partitions of memory (or you can do this manually for entire RDD with RDD.unpersist()), according with documentation.
Thus, as you have more partitions, Spark is storing fewer partitions in LRU so that they are not causing OOM (this may have negative impact too, like the need to re-cache partitions).
Another importante point is that when you write result back to HDFS using X partitions, then you have X tasks for all your data - take all the data size and divide by X, this is the memory for each task, that are executed on each (virtual) core. So, that's not difficult to see that X = 64 lead to OOM, but X = 1024 not.
I am planning to use Spark to process data where each individual element/row in the RDD or DataFrame may occasionally be large (up to several GB).
The data will probably be stored in Avro files in HDFS.
Obviously, each executor must have enough RAM to hold one of these "fat rows" in memory, and some to spare.
But are there other limitations on row size for Spark/HDFS or for the common serialisation formats (Avro, Parquet, Sequence File...)? For example, can individual entries/rows in these formats be much larger than the HDFS block size?
I am aware of published limitations for HBase and Cassandra, but not Spark...
There are currently some fundamental limitations related to block size, both for partitions in use and for shuffle blocks - both are limited to 2GB, which is the maximum size of a ByteBuffer (because it takes an int index, so is limited to Integer.MAX_VALUE bytes).
The maximum size of an individual row will normally need to be much smaller than the maximum block size, because each partition will normally contain many rows, and the largest rows might not be evenly distributed among partitions - if by chance a partition contains an unusually large number of big rows, this may push it over the 2GB limit, crashing the job.
See:
Why does Spark RDD partition has 2GB limit for HDFS?
Related Jira tickets for these Spark issues:
https://issues.apache.org/jira/browse/SPARK-1476
https://issues.apache.org/jira/browse/SPARK-5928
https://issues.apache.org/jira/browse/SPARK-6235
I have not been able to find much information on this topic but lets say we use a dataframe to read in a parquet file that is 10 Blocks spark will naturally create 10 partitions. But when the dataframe reads in the file to process it, won't it be processing a large data to partition ratio because if it was processing the file uncompressed the block size would have been much larger meaning the partitions would be larger as well.
So let me clarify, parquet compressed (these numbers are not fully accurate).
1GB Par = 5 Blocks = 5 Partitions which might be decompressed to 5GB making it 25 blocks/25 partitions. But unless you repartition the 1GB par file you will be stuck with just 5 partitions when optimally it would be 25 partitions? Or is my logic wrong.
Would make sense to repartition to increase speed? Or am I thinking about this wrong. Can anyone shed some light on this?
Assumptions:
1 Block = 1 Partition For Spark
1 Core operated on 1 Partition
Spark DataFrame doesn't load parquet files in memory. It uses Hadoop/HDFS API to read it during each operation. So the optimal number of partitions depends on HDFS block size (different from a Parquet block size!).
Spark 1.5 DataFrame partitions parquet file as follows:
1 partition per HDFS block
If HDFS block size is less than configured in Spark parquet block size a partition will be created for multiple HDFS blocks such as total size of partition is no less than parquet block size
I saw the other answer but I thought I can clarify more on this. If you are reading Parquet from posix filesystem then you can increase number of partitioning readings by just having more workers in Spark.
But in order to control the balance of data that comes into workers one may use the hierarchical data structure of the Parquet files, and later in the workers you may point to different partitions or parts of the Parquet file. This will give you control over how much of data should go to each worker according to the domain of your dataset (if by balancing data in workers you mean equal batch of data per worker is not efficient).