Question Purpose
Sorting a parquet files provides a number of benefits:
more efficient filtering using file metadata
more efficient compression rate
There may be other benefits for this. There is a lot of discussion about this on the Internet. For this reason, the discussion of this question is not about the cause of sorting. Rather, the purpose of this question is to talk about how to sort, which is mentioned in all Internet links with the least explanation (about 30%) and the challenges of data sorting are not mentioned at all. The purpose of this question is to get help from all friends who are experts and experienced in this field and to determine the best method (based on cost and benefit) for sorting.
Brief explanation about Apache parquet library
Before starting discussing Spark, I will explain about the tool used to produce parquet files. The parquet-mr library (I use Java for example, but it can probably be extended to other languages) writes to a disk and memory at the same time when we create a parquet file. This library also has a feature called getDataSize() that returns the exact final size of the file after it is completely closed on the disk, so we can use it to achieve the following two conditions when we write parquet files:
Do not make parquet files with small size (which is not good for query engines)
All parquet files can be produced with a certain minimum size or fixed size (for example, 1 GB each file)
Since this library writes to disk and memory at the same time, it does not allow data to be sorted unless all the data is first sorted in memory and then given to the library. (But this is not possible with large volumes of data.) We also implicitly assume that data is being generated as a stream that we intend to store. (In the case of a fixed data, the problem stated in this question will be meaningless because it can be said that the whole data is arranged once and for all and the problem is over. But we assume that there is a flow of data, in which case it is important to have an optimal way to sort the data)
One advantage mentioned above for the Apache parquet library is that we can fix the exact size of the output parquet file. This is an advantage in my opinion. Because, for example, if I know that the size of Hadoop blocks is equal to 128 MB and the size of parquet row-group is 128 MB, I can fix the parquet file size to 1 GB. Then I know that all parquet files will have 8 blocks and HDFS storage will be used best and all parquet files will be the same. (Because in HDFS, when the block size is 128 MB, the smaller file will take up the same amount of space) This may not be an advantage for everyone, and we'd be happy for experienced people to critique it if needed.
Parquet File Sorting Challenges
One point before we start is that we are looking for permanent data sorting because we are going to use it in the next thousands of queries. Almost so far, the above descriptions have identified some of challenges for sorting, but I will describe all of the challenges below:
Parquet tools do not allow you to write sorted data. So one way is to keep all the data in memory and after sorting, give it to the parquet library to be written in the parquet file. This method has two drawbacks: 1) It is not possible to keep all data in memory. 2) Because all the data is in memory, the size of the parquet file is not known and may be less than or more than 1 GB or any amount after writing, and the advantage of being fixed parquet size is lost.
Suppose we want to do this sorting in a parallel process instead of doing it in real time and stream. In this way, if we want to use parquet library, we will still have the problem that we have to bring the whole data to the memory for sorting, which is not possible. So let's say we use a tool like Spark for sorting. A specific cost we give in this section is that cluster resources are used for sorting, and in practice each data is written twice. (Once the parquet writing time and once the sorting) The next point is that even if we skip these two cases, after sorting the data, depending on the other columns in the parquet file, the amount of parquet compression for that particular column and for the whole data may change and increase or decrease. For this reason, after the parquet file is written, small files may be created or the fixed size (for example, 1 GB) may change. Unfortunately, Spark does not provide a way to control the file size (it may not be possible in practice), and therefore if we want to restore the fixed file size, we may need to use methods such as the mentioned link, which will not be free (causes to write the file several times apart from the cluster resources that are consumed and the exact file size will not be fixed):How do you control the size of the output file
Maybe there is no other way and the only ways are the mentioned one at the above. In which case, I would be happy for this note to be expressed by experts so that others know that there is no other way right now.
Challenges In Summary
For this reason, we generally observed 2 types of problems in these solutions:
How to do sorting at a reasonable cost and time (in stream flow)
How to keep the size of parquet files fixed
For this reason, although it is said everywhere that sorting is very good (and the results of surveys, both on the Internet and by myself, show that it is really useful), there is no mention at all of its methods and challenges. I ask experienced and expert friends in this field to help me in this direction (hoping that it will help others as well) and if ways or points are missed in this explanation, please state it.
Sorry if there is a typo in some parts due to my weakness in English language. Thanks.
Related
What does this property spark.sql.files.openCostInBytes do ?
This is official document definition:
The estimated cost to open a file, measured by the number of bytes could be scanned in the same time. This is used when putting multiple files into a partition. It is better to over-estimated, then the partitions with small files will be faster than partitions with bigger files (which is scheduled first). This configuration is effective only when using file-based sources such as Parquet, JSON and ORC.
But still didn't get it. Can anyone explain with small example that why and where its useful?
When saving to a delta table we avoid 'df.coalesce(1)' but when saving to csv or parquet we(my team) add 'df.coalesce(1)'. Is it a common practise? Why? Is it mandatory?
In most cases when I have seen df.coalesce(1) it was done to generate only one file, for example, import CSV file into Excel, or for Parquet file into the Pandas-based program. But if you're doing .coalesce(1), then the write happens via single task, and it's becoming the performance bottleneck because you need to get data from other executors, and write it.
If you're consuming data from Spark or other distributed system, having multiple files will be beneficial for performance because you can write & read them in parallel. By default, Spark writes N files into the directory where N is the number of partitions. As #pltc noticed, this may generate the big number of files that's often not desirable because you'll get performance overhead from accessing them. So we need to have a balance between the number of files and their size - for Parquet and Delta (that is based on Parquet), having the bigger files bring several performance advantages - you read less files, you can get better compression for data inside the file, etc.
For Delta specifically, having .coalesce(1) having the same problem as for other file formats - you're writing via one task. Relying on default Spark behaviour and writing multiple files is beneficial from performance point of view - each node is writing its data in parallel, but you can get too many small files (so you may use .coalesce(N) to write bigger files). For Databricks Delta, as it was correctly pointed by #Kafels, there are some optimizations that will allow to remove that .coalesce(N) and do automatic tuning achieve the best throughput (so called "Optimized Writes"), and create bigger files ("Auto compaction") - but they should be used carefully.
Overall, the topic of optimal file size for Delta is an interesting topic - if you have big files (1Gb is used by default by OPTIMIZE command), you can get better read throughput, but if you're rewriting them with MERGE/UPDATE/DELETE, then big files are bad from performance standpoint, and it's better to have smaller (16-64-128Mb) files, so you can rewrite less data.
TL;DR: it's not mandatory, it depends on the size of your dataframe.
Long answer:
If your dataframe is 10Mb, and you have 1000 partitions for example, each file would be about 10Kb. And having so many small files would reduce Spark performance dramatically, not to mention when you have too many files, you'll eventually reach OS limitation of the number of files. Any how, when your dataset is small enough, you should merge them into a couple of files by coalesce.
However, if your dataframe is 100G, technically you still can use coalesce(1) and save to a single file, but later on you will have to deal with less parallelism when reading from it.
I've been reading few questions regarding this topic and also several forums, and in all of them they seem to be mentioning that each of resulting .parquet files coming out from Spark should be either 64MB or 1GB size, but still can't make my mind around which case scenarios belong to each of those file sizes and the reasons behind apart from HDFS splitting them in 64MB blocks.
My current testing scenario is the following.
dataset
.coalesce(n) # being 'n' 4 or 48 - reasons explained below.
.write
.mode(SaveMode.Append)
.partitionBy(CONSTANTS)
.option("basepath", outputPath)
.parquet(outputPath)
I'm currently handling a total of 2.5GB to 3GB of daily data, that will be split and saved into daily buckets per year. The reasons behind 'n' being 4 or 48 is just for testing purposes, as I know the size of my testing set in advance, I try to get a number as close to 64MB or 1GB as I can. I haven't implemented code to buffer the needed data until I get the exact size I need prior saving.
So my question here is...
Should I take the size that much into account if I'm not planning to use HDFS and merely store and retrieve data from S3?
And also, which should be the optimal size for daily datasets of around 10GB maximum if I'm planning to use HDFS to store my resulting .parquet files?
Any other optimization tip would be really appreciated!
You can control the split size of parquet files, provided you save them with a splittable compression like snappy. For the s3a connector, just set fs.s3a.block.size to a different number of bytes.
Smaller split size
More workers can work on a file simultaneously. Speedup if you have idle workers.
More startup overhead scheduling work, starting processing, committing tasks
Creates more files from the output, unless you repartition.
Small files vs large files
Small files:
you get that small split whether or not you want it.
even if you use unsplittable compression.
takes longer to list files. Listing directory trees on s3 is very slow
impossible to ask for larger block sizes than the file length
easier to save if your s3 client doesn't do incremental writes in blocks. (Hadoop 2.8+ does if you set spark.hadoop.fs.s3a.fast.upload true.
Personally, and this is opinion, and some benchmark driven -but not with your queries
Writing
save to larger files.
with snappy.
shallower+wider directory trees over deep and narrow
Reading
play with different block sizes; treat 32-64 MB as a minimum
Hadoop 3.1, use the zero-rename committers. Otherwise, switch to v2
if your FS connector supports this make sure random IO is turned on (hadoop-2.8 + spark.hadoop.fs.s3a.experimental.fadvise random
save to larger files via .repartion().
Keep an eye on how much data you are collecting, as it is very easy to run up large bills from storing lots of old data.
see also Improving Spark Performance with S3/ADLS/WASB
new Spark user here. i wasn't able to find any information about filesize comparison between JSON and parquet output of the same dataFrame via Spark.
testing with a very small data set for now, doing a df.toJSON().collect() and then writing to disk creates a 15kb file. but doing a df.write.parquet creates 105 files at around 1.1kb each. why is the total file size so much larger with parquet in this case than with JSON?
thanks in advance
what you're doing with df.toJSON.collect is you get a single JSON from all your data (15kb in your case) and you save that to disk - this is not something scalable for situations you'd want to use Spark in any way.
For saving parquet you are using spark built-in function and it seems that for some reason you have 105 partitions (probably the result of the manipulation you did) so you get 105 files. Each of these files has the overhead of the file structure and probably stores 0,1 or 2 records. if you want to save a single file you should coalesce(1) before you save (again this just for the toy example you have) so you'd get 1 file. Note that it still might be larger due to the file format overhead (i.e. the overhead might still be larger than the compression benefit)
Conan, it is very hard to answer your question precisely without knowing the nature of the data (you don't even tell amount of row in your DataFrame). But let me speculate.
First. Text files containing JSON usually take more space on disk then parquet. At least when one store millions-billions rows. The reason for that is parquet is highly optimized column based storage format which uses a binary encoding to store your data
Second. I would guess that you have a very small dataframe with 105 partitions (and probably 105 rows). When you store something that small the disk footprint should not bother you but if it does you need to be aware that each parquet file has a relatively sizeable header describing the data you store.
[Disclaimer: While this question is somewhat specific, I think it circles a very generic issue with Hadoop/Spark.]
I need to process a large dataset (~14TB) in Spark. Not doing aggregations, mostly filtering. Given ~30k files (250 part files, per month for 10 years, each part being ~ 200MB), I would like to load them into a RDD/DataFrame and filter out items based on some arbitrary filters.
To make the listing of the files efficient (I'm on google dataproc/cloud storage, so the driver doing a wildcard glob was very serial and very slow), I precalculate an RDD of the file names, then load them into an RDD (I'm using avro, but file type shouldn't be relevant), e.g.
#returns an array of files to load
files = sc.textFile('/list/of/files/').collect()
#load the files into a dataframe
documents = sqlContext.read.format('com.databricks.spark.avro').load(files)
When I do this, even on a 50-worker cluster, it seems that only one executor is doing the work of reading the files. I've experimented with broadcasting the files list and read a dozen different approaches but I can't seem to crack the issue.
So, is there an efficient way to create a very large dataframe from multiple files? How do I best take advantage of all the potential computing power when creating this RDD?
This approach works very well on smaller sets but, at this size, I see a large number of symptoms like long-running processes with no feedback. Is there some treasure trove of knowledge -- besides #zero323 :-) -- on optimizing spark at this scale?
Listing 30k files shouldn't be an issue for GCS - even if single GCS list request that lists up to 500 files at a time will take 1 second each, all 30k files will be listed in a minute or so. There could be some corner cases with some glob patterns that make it slow, but there were recent optimizations in GCS connector globbing implementation that could help.
That's why it should be good enough for you to just rely on default Spark API with globbing:
val df = sqlContext.read.avro("gs://<BUCKET>/path/to/files/")