Let's say I'm reading 100 files from an S3 folder. Each file is of size 10 MB. When I execute df = spark.read.parquet(s3 path), how do the files (or rather partitions) get distributed across tasks? E.g. in this case df is going to have 100 partitions, and if spark has 10 tasks running for reading contents of this folder into the data frame, how the partitions are getting assigned to the 10 tasks? Is it in a round-robin fashion, or each task gets equal proportions of all partitions in a range based distribution, or something else? Any pointer to relevant resources would also be very helpful. Thank you.
Tasks are directly proportional to the number of partitions.
Spark tries to partition the rows directly from original partitions without bringing anything to the driver.
The partition logic is to start with a randomly picked target partition and then assign partitions to the rows in a round-robin method. Note that "start" partition is picked for each source partition and there could be collisions.
The final distribution depends on many factors: a number of source/target partitions and the number of rows in your dataframe.
Related
Problem outline: Say I have 300+ GB of data being processed with spark on an EMR cluster in AWS. This data has three attributes used to partition on the filesystem for use in Hive: date, hour, and (let's say) anotherAttr. I want to write this data to a fs in such a way that minimizes the number of files written.
What I'm doing right now is getting the distinct combinations of date, hour, anotherAttr, and a count of how many rows make up combination. I collect them into a List on the driver, and iterate over the list, building a new DataFrame for each combination, repartitioning that DataFrame using the number of rows to guestimate file size, and writing the files to disk with DataFrameWriter, .orc finishing it off.
We aren't using Parquet for organizational reasons.
This method works reasonably well, and solves the problem that downstream teams using Hive instead of Spark don't see performance issues resulting from a high number of files. For example, if I take the whole 300 GB DataFrame, do a repartition with 1000 partitions (in spark) and the relevant columns, and dumped it to disk, it all dumps in parallel, and finishes in ~9 min with the whole thing. But that gets up to 1000 files for the larger partitions, and that destroys Hive performance. Or it destroys some kind of performance, honestly not 100% sure what. I've just been asked to keep the file count as low as possible. With the method I'm using, I can keep the files to whatever size I want (relatively close anyway), but there is no parallelism and it takes ~45 min to run, mostly waiting on file writes.
It seems to me that since there's a 1-to-1 relationship between some source row and some destination row, and that since I can organize the data into non-overlapping "folders" (partitions for Hive), I should be able to organize my code/DataFrames in such a way that I can ask spark to write all the destination files in parallel. Does anyone have suggestions for how to attack this?
Things I've tested that did not work:
Using a scala parallel collection to kick off the writes. Whatever spark was doing with the DataFrames, it didn't separate out the tasks very well and some machines were getting massive garbage collection problems.
DataFrame.map - I tried to map across a DataFrame of the unique combinations, and kickoff writes from inside there, but there's no access to the DataFrame of the data that I actually need from within that map - the DataFrame reference is null on the executor.
DataFrame.mapPartitions - a non-starter, couldn't come up with any ideas for doing what I want from inside mapPartitions
The word 'partition' is also not especially helpful here because it refers both to the concept of spark splitting up the data by some criteria, and to the way that the data will be organized on disk for Hive. I think I was pretty clear in the usages above. So if I'm imagining a perfect solution to this problem, it's that I can create one DataFrame that has 1000 partitions based on the three attributes for fast querying, then from that create another collection of DataFrames, each one having exactly one unique combination of those attributes, repartitioned (in spark, but for Hive) with the number of partitions appropriate to the size of the data it contains. Most of the DataFrames will have 1 partition, a few will have up to 10. The files should be ~3 GB, and our EMR cluster has more RAM than that for each executor, so we shouldn't see a performance hit from these "large" partitions.
Once that list of DataFrames is created and each one is repartitioned, I could ask spark to write them all to disk in parallel.
Is something like this possible in spark?
One thing I'm conceptually unclear on: say I have
val x = spark.sql("select * from source")
and
val y = x.where(s"date=$date and hour=$hour and anotherAttr=$anotherAttr")
and
val z = x.where(s"date=$date and hour=$hour and anotherAttr=$anotherAttr2")
To what extent is y is a different DataFrame than z? If I repartition y, what effect does the shuffle have on z, and on x for that matter?
We had the same problem (almost) and we ended up by working directly with RDD (instead of DataFrames) and implementing our own partitioning mechanism (by extending org.apache.spark.Partitioner)
Details: we are reading JSON messages from Kafka. The JSON should be grouped by customerid/date/more fields and written in Hadoop using Parquet format, without creating too many small files.
The steps are (simplified version):
a)Read the messages from Kafka and transform them to a structure of RDD[(GroupBy, Message)]. GroupBy is a case class containing all the fields that are used for grouping.
b)Use a reduceByKeyLocally transformation and obtain a map of metrics (no of messages/messages size/etc) for each group - eg Map[GroupBy, GroupByMetrics]
c)Create a GroupPartitioner that's using the previously collected metrics (and some input parameters like the desired Parquet size etc) to compute how many partitions should be created for each GroupBy object. Basically we are extending org.apache.spark.Partitioner and overriding numPartitions and getPartition(key: Any)
d)we partition the RDD from a) using the previously defined partitioner: newPartitionedRdd = rdd.partitionBy(ourCustomGroupByPartitioner)
e)Invoke spark.sparkContext.runJob with two parameters: the first one is the RDD partitioned at d), the second one is a custom function (func: (TaskContext, Iterator[T]) that will write the messages taken from Iterator[T] into Hadoop/Parquet
Let's say that we have 100 mil messages, grouped like that
Group1 - 2 mil
Group2 - 80 mil
Group3 - 18 mil
and we decided that we have to use 1.5 mil messages per partition to obtain Parquet files greater than 500MB. We'll end up with 2 partitions for Group1, 54 for Group2, 12 for Group3.
This statement:
I collect them into a List on the driver, and iterate over the list,
building a new DataFrame for each combination, repartitioning that
DataFrame using the number of rows to guestimate file size, and
writing the files to disk with DataFrameWriter, .orc finishing it off.
is completely off-beam where Spark is concerned. Collecting to driver is never a good approach, volumes and OOM issues and latency in your approach is high.
Use so the below so as to simplify and get parallelism of Spark benefits saving time and money for your boss:
df.repartition(cols...)...write.partitionBy(cols...)...
shuffle occurs via repartition, no shuffling ever with partitionBy.
That simple, with Spark's default parallelism utilized.
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
Suppose we have a PySpark dataframe with data spread evenly across 2048 partitions, and we want to coalesce to 32 partitions to write the data back to HDFS. Using coalesce is nice for this because it does not require an expensive shuffle.
But one of the downsides of coalesce is that it typically results in an uneven distribution of data across the new partitions. I assume that this is because the original partition IDs are hashed to the new partition ID space, and the number of collisions is random.
However, in principle it should be possible to coalesce evenly, so that the first 64 partitions from the original dataframe are sent to the first partition of the new dataframe, the next 64 are send to the second partition, and so end, resulting in an even distribution of partitions. The resulting dataframe would often be more suitable for further computations.
Is this possible, while preventing a shuffle?
I can force the relationship I would like between initial and final partitions using a trick like in this question, but Spark doesn't know that everything from each original partition is going to a particular new partition. Thus it can't optimize away the shuffle, and it runs much slower than coalesce.
In your case you can safely coalesce the 2048 partitions into 32 and assume that Spark is going to evenly assign the upstream partitions to the coalesced ones (64 for each in your case).
Here is an extract from the Scaladoc of RDD#coalesce:
This results in a narrow dependency, e.g. if you go from 1000 partitions to 100 partitions, there will not be a shuffle, instead each of the 100 new partitions will claim 10 of the current partitions.
Consider that also how your partitions are physically spread across the cluster influence the way in which coalescing happens. The following is an extract from CoalescedRDD's ScalaDoc:
If there is no locality information (no preferredLocations) in the parent, then the coalescing is very simple: chunk parents that are close in the Array in chunks.
If there is locality information, it proceeds to pack them with the following four goals:
(1) Balance the groups so they roughly have the same number of parent partitions
(2) Achieve locality per partition, i.e. find one machine which most parent partitions prefer
(3) Be efficient, i.e. O(n) algorithm for n parent partitions (problem is likely NP-hard)
(4) Balance preferred machines, i.e. avoid as much as possible picking the same preferred machine
I use Spark 2.
Actually I am not the one executing the queries so I cannot include query plans. I have been asked this question by the data science team.
We are having hive table partitioned into 2000 partitions and stored in parquet format. When this respective table is used in spark, there are exactly 2000 tasks that are executed among the executors. But we have a block size of 256 MB and we are expecting the (total size/256) number of partitions which will be much lesser than 2000 for sure. Is there any internal logic that spark uses physical structure of data to create partitions. Any reference/help would be greatly appreciated.
UPDATE: It is the other way around. Actually our table is very huge like 3 TB having 2000 partitions. 3TB/256MB would actually come to 11720 but we are having exactly same number of partitions as the table is partitioned physically. I just want to understand how the tasks are generated on data volume.
In general Hive partitions are not mapped 1:1 to Spark partitions. 1 Hive partition can be split into multiple Spark partitions, and one Spark partition can hold multiple hive-partitions.
The number of Spark partitions when you load a hive-table depends on the parameters:
spark.files.maxPartitionBytes (default 128MB)
spark.files.openCostInBytes (default 4MB)
You can check the partitions e.g. using
spark.table(yourtable).rdd.partitions
This will give you an Array of FilePartitions which contain the physical path of your files.
Why you got exactly 2000 Spark partitions from your 2000 hive partitions seems a coincidence to me, in my experience this is very unlikely to happen. Note that the situation in spark 1.6 was different, there the number of spark partitions resembled the number of files on the filesystem (1 spark partition for 1 file, unless the file was very large)
I just want to understand how the tasks are generated on data volume.
Tasks are a runtime artifact and their number is exactly the number of partitions.
The number of tasks does not correlate to data volume in any way. It's a Spark developer's responsibility to have enough partitions to hold the data.
Say I have lots data in a couple of s3 files, about 5 GB each, which I read in using sc.textFile
I need to join the data from the two files, therefore, I opt to use the HashPartitioner technique, and I set a partition count of 20. The submitted job to 8 worker nodes fails without any meaningful messages. Now I am thinking maybe I need to pick a proper number of partitions.
Obviously, the idea for spark to partition up all the data based on a chosen key. In order to load them up into 20 partitions, I imagine spark will have to read thru every line of data, compute its hash, and load into the memory of the matching partition, which resides in one of the 8 worker nodes. If there is enough collective memory in the worker nodes, I assume this goes smoothly. At the end of the read, all the data is in the proper partition, in the right node's memory. Am I right so far?
However, if the total memory can not fit all the data, I imagine Spark will work on certain partitions first. And after processing these first partitions, it flushes the original partitions and reads from the source files again, loading remaining data into new partitions. This would mean reading the same file as many time as necessary to process all partitions using available memory. Is this also correct?
Should I should calculate the number of partitions so that at least one full partition would fit into a single node's memory. Are there other guidelines to follow?