Pyspark - Read data inside a map function? - apache-spark

I have a dataframe as below:
id | file_path
--------------------------
abc | s3://data/file1.json
def | s3://data/file2.json
For every row in this dataframe, I want to read the contents of the file located in file_path in a distributed manner.
Here's what I tried:
rdd_paths = df.rdd.map(lambda x: x.file_path)
rdd_contents = rdd_paths.map(lambda y: spark.read.parquet(y))
rdd_contents.take(2)
This gave me the following error:
Exception: It appears that you are attempting to reference SparkContext from a broadcast variable, action, or transformation. SparkContext can only be used on the driver, not in code that it run on workers. For more information, see SPARK-5063.
I checked SPARK-5063 but did not get clear pointers to solve this. How can I read from the file paths in a distributed manner?

Spark context can only be accessed from driver node, since map() is executed in the worker nodes, it doesnt have access to spark in your code that does spark.read.parquet(y).
You need to modify you logic, for example(not a very good example) you can the column values for s3 paths, pass this to textFile which provide key as file name and value as file content.
paths = df.select('file_path').collect()
rdd4 = sc.textFile(paths)
Now, you can apply map or iterate to store values (file content) as columns in the dataframe. For example with join.
rdd4.foreach(f=>{
println(f)
})

Related

Spark Reading and Writing to same S3 Path Giving Unable to infer Schema Error

I need to perform update insert (Upsert) on old data with new data.
Pseudo Code:
old_data = spark.read.parquet('s3://bucket/old_data/')
new_data= spark.read.parquet('s3://bucket/new_data/')
common_records = old_data.join(new_data,on=opk,how="inner")
non_match_records = old_data.join(new_data,on=opk,how="left_anti")
new_records = new_data.join(old_data,on=opk,how="left_anti")
dfs = [common_records , non_match_records , new_records ]
final_data = reduce(DataFrame.unionAll, dfs)
final_data .cache()
final_data.write.parquet('s3://bucket/old_data/')
Error :
Even I cached data, Still It's Looking for old_data path, Is there anyway to directly write to old data s3 path.
I have tried it to write to some temp path and read from it and write to main path Like below It worked but It's taking time process when I have data in Billons.
final_data.write.parquet('s3://bucket/temp/')
df = spark.read.parquet('s3://bucket/temp/')
df.write.parquet('s3://bucket/old_data/')
I want to reduce this Temp writing & reading part.
Thanks in advance :)
You need to perform an action to trigger dataframe caching. Thus you should modify the last lines of your code snippet as follow:
...
final_data = final_data.cache()
final_data.count()
final_data.write.parquet('s3://bucket/old_data/')
By performing a count action on your dataframe, you trigger caching process and then be able to write on same directory from where your read.
However, I don't know if this will improve performance of your application as cache fallback to disk write when dataframe is too huge for memory. If your use case is updating parquet files, I advise you to look at DeltaLake that was created to solve this.

Convert CSV files from multiple directory into parquet in PySpark

I have CSV files from multiple paths that are not parent directories in s3 bucket. All the tables have the same partition keys.
the directory of the s3:
table_name_1/partition_key_1 = <pk_1>/partition_key_2 = <pk_2>/file.csv
table_name_2/partition_key_1 = <pk_1>/partition_key_2 = <pk_2>/file.csv
...
I need to convert these csv files into parquet files and store them in another s3 bucket that has the same directory structure.
the directory of another s3:
table_name_1/partition_key_1 = <pk_1>/partition_key_2 = <pk_2>/file.parquet
table_name_2/partition_key_1 = <pk_1>/partition_key_2 = <pk_2>/file.parquet
...
I have a solution is iterating through the s3 bucket and find the CSV file and convert it to parquet and save to the another S3 path. I find this way is not efficient, because i have a loop and did the conversion one file by one file.
I want to utilize the spark library to improve the efficiency.
Then, I tried:
spark.read.csv('s3n://bucket_name/table_name_1/').write.partitionBy('partition_key_1', 'partition_key_2').parquet('s3n://another_bucket/table_name_1')
This way works good for each table, but to optimize it more, I want to take the table_name as a parameter, something like:
TABLE_NAMES = [table_name_1, table_name_2, ...]
spark.read.csv('s3n://bucket_name/{*TABLE_NAMES}/').write.partitionBy('partition_key_1', 'partition_key_2').parquet('s3n://another_bucket/{*TABLE_NAMES}')
Thanks
The mentioned question provides solutions for reading multiple files at once. The method spark.read.csv(...) accepts one or multiple paths as shown here. For reading the files you can apply the same logic. Although, when it comes to writing, Spark will merge all the given dataset/paths into one Dataframe. Therefore it is not possible to generate from one single dataframe multiple dataframes without applying a custom logic first. So to conclude, there is not such a method for extracting the initial dataframe directly into multiple directories i.e df.write.csv(*TABLE_NAMES).
The good news is that Spark provides a dedicated function namely input_file_name() which returns the file path of the current record. You can use it in combination with TABLE_NAMES to filter on the table name.
Here it is one possible untested PySpark solution:
from pyspark.sql.functions import input_file_name
TABLE_NAMES = [table_name_1, table_name_2, ...]
source_path = "s3n://bucket_name/"
input_paths = [f"{source_path}/{t}" for t in TABLE_NAMES]
all_df = spark.read.csv(*input_paths) \
.withColumn("file_name", input_file_name()) \
.cache()
dest_path = "s3n://another_bucket/"
def write_table(table_name: string) -> None:
all_df.where(all_df["file_name"].contains(table_name))
.write
.partitionBy('partition_key_1','partition_key_2')
.parquet(f"{dest_path}/{table_name}")
for t in TABLE_NAMES:
write_table(t)
Explanation:
We generate and store the input paths into input_paths. This will create paths such as: s3n://bucket_name/table1, s3n://bucket_name/table2 ... s3n://bucket_name/tableN.
Then we load all the paths into one dataframe in which we add a new column called file_name, this will hold the path of each row. Notice that we also use cache here, this is important since we have multiple len(TABLE_NAMES) actions in the following code. Using cache will prevent us from loading the datasource again and again.
Next we create the write_table which is responsible for saving the data for the given table. The next step is to filter based on the table name using all_df["file_name"].contains(table_name), this will return only the records that contain the value of the table_name in the file_name column. Finally we save the filtered data as you already did.
In the last step we call write_table for every item of TABLE_NAMES.
Related links
How to import multiple csv files in a single load?
Get HDFS file path in PySpark for files in sequence file format

Spark unzip and write CSV as parquet in executor

My issue is that my CSV files come in ZIP .csv.zip format so I cannot just use Sparks .csv directly as I could with .csv.gzip | .csv.gz. Which means I need to decompress the file, read the contents (the files are quite big ~5gb) and write them as parquet files.
My approach is as such:
String paths = "s3a://...,s3a://...,...";
JavaRDD<Tuple2<String, PortableDataStream>> zipRDD = context.binaryFiles(paths, sparkContext.context.defaultParallelism()).toJavaRDD();
JavaRDD<Tuple2<String, List<Row>>> filenameRowsRDD = zipRDD.flatMap(new ConvertLinesToRows());
The first JavaRDD returns a pair of Filename, InputStream. This is then passed to class ConvertLinesToRows which calls ZipInputStream reads the contents of CSV files and for each line creates a new spark Row and finally returning tuple pair of Filename, List<Row> where the list contains all lines from CSV converted to Row.
I now want to save each read CSV as parquet file.
filenameRowsRDD.foreach(tuple -> {
SparkContext newContext = MySparkConfig.createNewContext();
SparkSession newSpark = SparkSession.builder()
.sparkContext(newContext)
.getOrCreate();
Dataset<Row> dataset = newSpark.createDataFrame(tuple._2, datasetSchema.value());
dataset.write().parquet("s3a://...");
});
I recreate the SparkSession in my executor so to be able to use SparkSession.write.
The idea is that this will all run in an executor (I'm hoping). However with this approach, reading file is not an issue problem comes when executor wants to write this output file. Throwing exception for: A master URL must be set in your configuration.
This seems like I'm doing something anti-spark way. It also does not work. I have also tried broadcasting my SparkSession however that throws a NPE inside SparkSession before trying to write.
What would be the correct way to approach my problem here ?
What would be the spark way of doing this.
All of the above code in in my main() method. Am I correct in assuming that first zipRDD is run on master node and than the second filenameRowsRDD runs on executor nodes as well as the .foreach.

Dataframe not able to write on S3

I am creating a dataframe from existing hive table.Table is partitioned on date and site column.Now, when i am trying to overwrite the data in this same table after some computation with previous day data.It is successfully getting loaded.
But when i am trying to write final dataframe at S3 bucket. I am getting error saying file not found.Now the file it is mentioning is previous day file which is now overwritten.
If i write dataframe first and then overwrite table then its running fine.
For writing at S3 location , what it has to do with table partition file?
Below is the error and code.
java.io.FileNotFoundException: No such file or directory: s3://bucket_1/DM/web_fact_tbl/local_dt=2018-05-10/site_name=ABC/part-00000-882a6e29-eb6a-477c-8b88-6fe853956674.c000
fact_tbl = spark.table('db.web_fact_tbl')
fact_lkp = fact_tbl.filter(fact_tbl['local_dt']=='2018-05-10')
fact_join = fact_lkp.alias('a').join(fact_tbl.alias('b'),(col('a.id') == col('b.id')),"inner").select('a.*')
fact_final = fact_join.union(fact_tbl)
fact_final.coalesce(2).createOrReplaceTempView('cwf')
spark.sql('INSERT OVERWRITE TABLE dm.web_fact_tbl PARTITION (local_dt, site_name) \
SELECT * FROM cwf')
fact_final.write.csv('s3://bucket_1/yahoo')
Before last line fact_final is just a "lazy" dataframe object that contains definitions only. It does not contain any data. But it has pointer to exact data files, where data is stored actually.
When you try to perform actual operations (does not matter it's writing to S3, or executing query like fact_final.count()) you'll get the error as above. It looks like partition local_dt=2018-05-10 does not exists anymore (files/folder that sits behind it does not exists).
You can try to re-initialize dataframe once again, before final write (it's another lazy operation - all work is done in your case while you writing it on S3).

How to overwrite the output directory in spark

I have a spark streaming application which produces a dataset for every minute.
I need to save/overwrite the results of the processed data.
When I tried to overwrite the dataset org.apache.hadoop.mapred.FileAlreadyExistsException stops the execution.
I set the Spark property set("spark.files.overwrite","true") , but there is no luck.
How to overwrite or Predelete the files from spark?
UPDATE: Suggest using Dataframes, plus something like ... .write.mode(SaveMode.Overwrite) ....
Handy pimp:
implicit class PimpedStringRDD(rdd: RDD[String]) {
def write(p: String)(implicit ss: SparkSession): Unit = {
import ss.implicits._
rdd.toDF().as[String].write.mode(SaveMode.Overwrite).text(p)
}
}
For older versions try
yourSparkConf.set("spark.hadoop.validateOutputSpecs", "false")
val sc = SparkContext(yourSparkConf)
In 1.1.0 you can set conf settings using the spark-submit script with the --conf flag.
WARNING (older versions): According to #piggybox there is a bug in Spark where it will only overwrite files it needs to to write it's part- files, any other files will be left unremoved.
since df.save(path, source, mode) is deprecated, (http://spark.apache.org/docs/1.5.0/api/scala/index.html#org.apache.spark.sql.DataFrame)
use df.write.format(source).mode("overwrite").save(path)
where df.write is DataFrameWriter
'source' can be ("com.databricks.spark.avro" | "parquet" | "json")
From the pyspark.sql.DataFrame.save documentation (currently at 1.3.1), you can specify mode='overwrite' when saving a DataFrame:
myDataFrame.save(path='myPath', source='parquet', mode='overwrite')
I've verified that this will even remove left over partition files. So if you had say 10 partitions/files originally, but then overwrote the folder with a DataFrame that only had 6 partitions, the resulting folder will have the 6 partitions/files.
See the Spark SQL documentation for more information about the mode options.
The documentation for the parameter spark.files.overwrite says this: "Whether to overwrite files added through SparkContext.addFile() when the target file exists and its contents do not match those of the source." So it has no effect on saveAsTextFiles method.
You could do this before saving the file:
val hadoopConf = new org.apache.hadoop.conf.Configuration()
val hdfs = org.apache.hadoop.fs.FileSystem.get(new java.net.URI("hdfs://localhost:9000"), hadoopConf)
try { hdfs.delete(new org.apache.hadoop.fs.Path(filepath), true) } catch { case _ : Throwable => { } }
Aas explained here:
http://apache-spark-user-list.1001560.n3.nabble.com/How-can-I-make-Spark-1-0-saveAsTextFile-to-overwrite-existing-file-td6696.html
df.write.mode('overwrite').parquet("/output/folder/path") works if you want to overwrite a parquet file using python. This is in spark 1.6.2. API may be different in later versions
val jobName = "WordCount";
//overwrite the output directory in spark set("spark.hadoop.validateOutputSpecs", "false")
val conf = new
SparkConf().setAppName(jobName).set("spark.hadoop.validateOutputSpecs", "false");
val sc = new SparkContext(conf)
This overloaded version of the save function works for me:
yourDF.save(outputPath, org.apache.spark.sql.SaveMode.valueOf("Overwrite"))
The example above would overwrite an existing folder. The savemode can take these parameters as well (https://spark.apache.org/docs/1.4.0/api/java/org/apache/spark/sql/SaveMode.html):
Append: Append mode means that when saving a DataFrame to a data source, if data/table already exists, contents of the DataFrame are expected to be appended to existing data.
ErrorIfExists: ErrorIfExists mode means that when saving a DataFrame to a data source, if data already exists, an exception is expected to be thrown.
Ignore: Ignore mode means that when saving a DataFrame to a data source, if data already exists, the save operation is expected to not save the contents of the DataFrame and to not change the existing data.
Spark – Overwrite the output directory:
Spark by default doesn’t overwrite the output directory on S3, HDFS, and any other file systems, when you try to write the DataFrame contents to an existing directory, Spark returns runtime error hence. To overcome this Spark provides an enumeration org.apache.spark.sql.SaveMode.Overwrite to overwrite the existing folder.
We need to use this Overwrite as an argument to mode() function of the DataFrameWrite class, for example.
df. write.mode(SaveMode.Overwrite).csv("/tmp/out/foldername")
or you can use the overwrite string.
df.write.mode("overwrite").csv("/tmp/out/foldername")
Besides Overwrite, SaveMode also offers other modes like SaveMode.Append, SaveMode.ErrorIfExists and SaveMode.Ignore
For older versions of Spark, you can use the following to overwrite the output directory with the RDD contents.
sparkConf.set("spark.hadoop.validateOutputSpecs", "false")
val sparkContext = SparkContext(sparkConf)
If you are willing to use your own custom output format, you would be able to get the desired behaviour with RDD as well.
Have a look at the following classes:
FileOutputFormat,
FileOutputCommitter
In file output format you have a method named checkOutputSpecs, which is checking whether the output directory exists.
In FileOutputCommitter you have the commitJob which is usually transferring data from the temporary directory to its final place.
I wasn't able to verify it yet (would do it, as soon as I have few free minutes) but theoretically: If I extend FileOutputFormat and override checkOutputSpecs to a method that doesn't throw exception on directory already exists, and adjust the commitJob method of my custom output committer to perform which ever logic that I want (e.g. Override some of the files, append others) than I may be able to achieve the desired behaviour with RDDs as well.
The output format is passed to: saveAsNewAPIHadoopFile (which is the method saveAsTextFile called as well to actually save the files). And the Output committer is configured at the application level.

Resources