onETL

What is onETL?

Python ETL/ELT library powered by Apache Spark & other open-source tools.

Goals

• Provide unified classes to extract data from (E) & load data to (L) various stores.
• Provides Spark DataFrame API for performing transformations (T) in terms of ETL.
• Provide direct assess to database, allowing to execute SQL queries, as well as DDL, DML, and call functions/procedures. This can be used for building up ELT pipelines.
• Support different read strategies for incremental and batch data fetching.
• Provide hooks & plugins mechanism for altering behavior of internal classes.

Non-goals

• onETL is not a Spark replacement. It just provides additional functionality that Spark does not have, and improves UX for end users.
• onETL is not a framework, as it does not have requirements to project structure, naming, the way of running ETL/ELT processes, configuration, etc. All of that should be implemented in some other tool.
• onETL is deliberately developed without any integration with scheduling software like Apache Airflow. All integrations should be implemented as separated tools.
• Only batch operations, no streaming. For streaming prefer Apache Flink.

Requirements

• Python 3.7 - 3.12
• PySpark 2.3.x - 3.5.x (depends on used connector)
• Java 8+ (required by Spark, see below)
• Kerberos libs & GCC (required by Hive, HDFS and SparkHDFS connectors)

Supported storages

Database

Type	Storage	Powered by
Database	Clickhouse	Apache Spark JDBC Data Source
	MSSQL
	MySQL
	Postgres
	Oracle
	Teradata
	Hive	Apache Spark Hive integration
	Kafka	Apache Spark Kafka integration
	Greenplum	VMware Greenplum Spark connector
	MongoDB	MongoDB Spark connector
File	HDFS	HDFS Python client
	S3	minio-py client
	SFTP	Paramiko library
	FTP	FTPUtil library
	FTPS
	WebDAV	WebdavClient3 library
	Samba	pysmb library
Files as DataFrame	SparkLocalFS	Apache Spark File Data Source
	SparkHDFS
	SparkS3	Hadoop AWS library

Documentation

See https://onetl.readthedocs.io/

How to install

Minimal installation

Base onetl package contains:

• DBReader, DBWriter and related classes
• FileDownloader, FileUploader, FileMover and related classes, like file filters & limits
• FileDFReader, FileDFWriter and related classes, like file formats
• Read Strategies & HWM classes
• Plugins support

It can be installed via:

pip install onetl

Warning

This method does NOT include any connections.

This method is recommended for use in third-party libraries which require for onetl to be installed, but do not use its connection classes.

With DB and FileDF connections

All DB connection classes (Clickhouse, Greenplum, Hive and others) and all FileDF connection classes (SparkHDFS, SparkLocalFS, SparkS3) require Spark to be installed.

Firstly, you should install JDK. The exact installation instruction depends on your OS, here are some examples:

yum install java-1.8.0-openjdk-devel  # CentOS 7 + Spark 2
dnf install java-11-openjdk-devel  # CentOS 8 + Spark 3
apt-get install openjdk-11-jdk  # Debian-based + Spark 3

Compatibility matrix

Spark	Python	Java	Scala
2.3.x	3.7 only	8 only	2.11
2.4.x	3.7 only	8 only	2.11
3.2.x	3.7 - 3.10	8u201 - 11	2.12
3.3.x	3.7 - 3.10	8u201 - 17	2.12
3.4.x	3.7 - 3.12	8u362 - 20	2.12
3.5.x	3.8 - 3.12	8u371 - 20	2.12

Then you should install PySpark via passing spark to extras:

pip install onetl[spark]  # install latest PySpark

or install PySpark explicitly:

pip install onetl pyspark==3.5.3  # install a specific PySpark version

or inject PySpark to sys.path in some other way BEFORE creating a class instance. Otherwise connection object cannot be created.

With File connections

All File (but not FileDF) connection classes (FTP, SFTP, HDFS and so on) requires specific Python clients to be installed.

Each client can be installed explicitly by passing connector name (in lowercase) to extras:

pip install onetl[ftp]  # specific connector
pip install onetl[ftp,ftps,sftp,hdfs,s3,webdav,samba]  # multiple connectors

To install all file connectors at once you can pass files to extras:

pip install onetl[files]

Otherwise class import will fail.

With Kerberos support

Most of Hadoop instances set up with Kerberos support, so some connections require additional setup to work properly.

• HDFS Uses requests-kerberos and GSSApi for authentication. It also uses kinit executable to generate Kerberos ticket.
• Hive and SparkHDFS require Kerberos ticket to exist before creating Spark session.

So you need to install OS packages with:

• krb5 libs
• Headers for krb5
• gcc or other compiler for C sources

The exact installation instruction depends on your OS, here are some examples:

dnf install krb5-devel gcc  # CentOS, OracleLinux
apt install libkrb5-dev gcc  # Debian-based

Also you should pass kerberos to extras to install required Python packages:

pip install onetl[kerberos]

Full bundle

To install all connectors and dependencies, you can pass all into extras:

pip install onetl[all]

# this is just the same as
pip install onetl[spark,files,kerberos]

Warning

This method consumes a lot of disk space, and requires for Java & Kerberos libraries to be installed into your OS.

Quick start

MSSQL → Hive

Read data from MSSQL, transform & write to Hive.

# install onETL and PySpark
pip install onetl[spark]

# Import pyspark to initialize the SparkSession
from pyspark.sql import SparkSession

# import function to setup onETL logging
from onetl.log import setup_logging

# Import required connections
from onetl.connection import MSSQL, Hive

# Import onETL classes to read & write data
from onetl.db import DBReader, DBWriter

# change logging level to INFO, and set up default logging format and handler
setup_logging()

# Initialize new SparkSession with MSSQL driver loaded
maven_packages = MSSQL.get_packages()
spark = (
    SparkSession.builder.appName("spark_app_onetl_demo")
    .config("spark.jars.packages", ",".join(maven_packages))
    .enableHiveSupport()  # for Hive
    .getOrCreate()
)

# Initialize MSSQL connection and check if database is accessible
mssql = MSSQL(
    host="mssqldb.demo.com",
    user="onetl",
    password="onetl",
    database="Telecom",
    spark=spark,
    # These options are passed to MSSQL JDBC Driver:
    extra={"applicationIntent": "ReadOnly"},
).check()

# >>> INFO:|MSSQL| Connection is available

# Initialize DBReader
reader = DBReader(
    connection=mssql,
    source="dbo.demo_table",
    columns=["on", "etl"],
    # Set some MSSQL read options:
    options=MSSQL.ReadOptions(fetchsize=10000),
)

# checks that there is data in the table, otherwise raises exception
reader.raise_if_no_data()

# Read data to DataFrame
df = reader.run()
df.printSchema()
# root
#  |-- id: integer (nullable = true)
#  |-- phone_number: string (nullable = true)
#  |-- region: string (nullable = true)
#  |-- birth_date: date (nullable = true)
#  |-- registered_at: timestamp (nullable = true)
#  |-- account_balance: double (nullable = true)

# Apply any PySpark transformations
from pyspark.sql.functions import lit

df_to_write = df.withColumn("engine", lit("onetl"))
df_to_write.printSchema()
# root
#  |-- id: integer (nullable = true)
#  |-- phone_number: string (nullable = true)
#  |-- region: string (nullable = true)
#  |-- birth_date: date (nullable = true)
#  |-- registered_at: timestamp (nullable = true)
#  |-- account_balance: double (nullable = true)
#  |-- engine: string (nullable = false)

# Initialize Hive connection
hive = Hive(cluster="rnd-dwh", spark=spark)

# Initialize DBWriter
db_writer = DBWriter(
    connection=hive,
    target="dl_sb.demo_table",
    # Set some Hive write options:
    options=Hive.WriteOptions(if_exists="replace_entire_table"),
)

# Write data from DataFrame to Hive
db_writer.run(df_to_write)

# Success!

SFTP → HDFS

Download files from SFTP & upload them to HDFS.

# install onETL with SFTP and HDFS clients, and Kerberos support
pip install onetl[hdfs,sftp,kerberos]

# import function to setup onETL logging
from onetl.log import setup_logging

# Import required connections
from onetl.connection import SFTP, HDFS

# Import onETL classes to download & upload files
from onetl.file import FileDownloader, FileUploader

# import filter & limit classes
from onetl.file.filter import Glob, ExcludeDir
from onetl.file.limit import MaxFilesCount

# change logging level to INFO, and set up default logging format and handler
setup_logging()

# Initialize SFTP connection and check it
sftp = SFTP(
    host="sftp.test.com",
    user="someuser",
    password="somepassword",
).check()

# >>> INFO:|SFTP| Connection is available

# Initialize downloader
file_downloader = FileDownloader(
    connection=sftp,
    source_path="/remote/tests/Report",  # path on SFTP
    local_path="/local/onetl/Report",  # local fs path
    filters=[
        # download only files matching the glob
        Glob("*.csv"),
        # exclude files from this directory
        ExcludeDir("/remote/tests/Report/exclude_dir/"),
    ],
    limits=[
        # download max 1000 files per run
        MaxFilesCount(1000),
    ],
    options=FileDownloader.Options(
        # delete files from SFTP after successful download
        delete_source=True,
        # mark file as failed if it already exist in local_path
        if_exists="error",
    ),
)

# Download files to local filesystem
download_result = downloader.run()

# Method run returns a DownloadResult object,
# which contains collection of downloaded files, divided to 4 categories
download_result

#  DownloadResult(
#      successful=[
#          LocalPath('/local/onetl/Report/file_1.json'),
#          LocalPath('/local/onetl/Report/file_2.json'),
#      ],
#      failed=[FailedRemoteFile('/remote/onetl/Report/file_3.json')],
#      ignored=[RemoteFile('/remote/onetl/Report/file_4.json')],
#      missing=[],
#  )

# Raise exception if there are failed files, or there were no files in the remote filesystem
download_result.raise_if_failed() or download_result.raise_if_empty()

# Do any kind of magic with files: rename files, remove header for csv files, ...
renamed_files = my_rename_function(download_result.success)

# function removed "_" from file names
# [
#    LocalPath('/home/onetl/Report/file1.json'),
#    LocalPath('/home/onetl/Report/file2.json'),
# ]

# Initialize HDFS connection
hdfs = HDFS(
    host="my.name.node",
    user="someuser",
    password="somepassword",  # or keytab
)

# Initialize uploader
file_uploader = FileUploader(
    connection=hdfs,
    target_path="/user/onetl/Report/",  # hdfs path
)

# Upload files from local fs to HDFS
upload_result = file_uploader.run(renamed_files)

# Method run returns a UploadResult object,
# which contains collection of uploaded files, divided to 4 categories
upload_result

#  UploadResult(
#      successful=[RemoteFile('/user/onetl/Report/file1.json')],
#      failed=[FailedLocalFile('/local/onetl/Report/file2.json')],
#      ignored=[],
#      missing=[],
#  )

# Raise exception if there are failed files, or there were no files in the local filesystem, or some input file is missing
upload_result.raise_if_failed() or upload_result.raise_if_empty() or upload_result.raise_if_missing()

# Success!

S3 → Postgres

Read files directly from S3 path, convert them to dataframe, transform it and then write to a database.

# install onETL and PySpark
pip install onetl[spark]

# Import pyspark to initialize the SparkSession
from pyspark.sql import SparkSession

# import function to setup onETL logging
from onetl.log import setup_logging

# Import required connections
from onetl.connection import Postgres, SparkS3

# Import onETL classes to read files
from onetl.file import FileDFReader
from onetl.file.format import CSV

# Import onETL classes to write data
from onetl.db import DBWriter

# change logging level to INFO, and set up default logging format and handler
setup_logging()

# Initialize new SparkSession with Hadoop AWS libraries and Postgres driver loaded
maven_packages = SparkS3.get_packages(spark_version="3.5.3") + Postgres.get_packages()
spark = (
    SparkSession.builder.appName("spark_app_onetl_demo")
    .config("spark.jars.packages", ",".join(maven_packages))
    .getOrCreate()
)

# Initialize S3 connection and check it
spark_s3 = SparkS3(
    host="s3.test.com",
    protocol="https",
    bucket="my-bucket",
    access_key="somekey",
    secret_key="somesecret",
    # Access bucket as s3.test.com/my-bucket
    extra={"path.style.access": True},
    spark=spark,
).check()

# >>> INFO:|SparkS3| Connection is available

# Describe file format and parsing options
csv = CSV(
    delimiter=";",
    header=True,
    encoding="utf-8",
)

# Describe DataFrame schema of files
from pyspark.sql.types import (
    DateType,
    DoubleType,
    IntegerType,
    StringType,
    StructField,
    StructType,
    TimestampType,
)

df_schema = StructType(
    [
        StructField("id", IntegerType()),
        StructField("phone_number", StringType()),
        StructField("region", StringType()),
        StructField("birth_date", DateType()),
        StructField("registered_at", TimestampType()),
        StructField("account_balance", DoubleType()),
    ],
)

# Initialize file df reader
reader = FileDFReader(
    connection=spark_s3,
    source_path="/remote/tests/Report",  # path on S3 there *.csv files are located
    format=csv,  # file format with specific parsing options
    df_schema=df_schema,  # columns & types
)

# Read files directly from S3 as Spark DataFrame
df = reader.run()

# Check that DataFrame schema is same as expected
df.printSchema()
# root
#  |-- id: integer (nullable = true)
#  |-- phone_number: string (nullable = true)
#  |-- region: string (nullable = true)
#  |-- birth_date: date (nullable = true)
#  |-- registered_at: timestamp (nullable = true)
#  |-- account_balance: double (nullable = true)

# Apply any PySpark transformations
from pyspark.sql.functions import lit

df_to_write = df.withColumn("engine", lit("onetl"))
df_to_write.printSchema()
# root
#  |-- id: integer (nullable = true)
#  |-- phone_number: string (nullable = true)
#  |-- region: string (nullable = true)
#  |-- birth_date: date (nullable = true)
#  |-- registered_at: timestamp (nullable = true)
#  |-- account_balance: double (nullable = true)
#  |-- engine: string (nullable = false)

# Initialize Postgres connection
postgres = Postgres(
    host="192.169.11.23",
    user="onetl",
    password="somepassword",
    database="mydb",
    spark=spark,
)

# Initialize DBWriter
db_writer = DBWriter(
    connection=postgres,
    # write to specific table
    target="public.my_table",
    # with some writing options
    options=Postgres.WriteOptions(if_exists="append"),
)

# Write DataFrame to Postgres table
db_writer.run(df_to_write)

# Success!