Arbalest is a Python data pipeline orchestration library for Amazon S3 and Amazon Redshift. It takes care of the heavy lifting of making data queryable at scale in AWS.
It takes care of:
- Ingesting data into Amazon Redshift
- Schema creation and validation
- Creating highly available and scalable data import strategies
- Generating and uploading prerequisite artifacts for import
- Running data import jobs
- Orchestrating idempotent and fault tolerant multi-step ETL pipelines with SQL
Why Arbalest?
- Lightweight library over heavyweight frameworks that can be composed with existing data tools
- Python is a de facto lingua franca for data science
- Configuration as code
- Batteries included, for example, strategies for ingesting time series or sparse data (
arbalest.pipeline), or integration with an existing pipeline topology (arbalest.contrib)
Use Cases
Arbalest is not a MapReduce framework, but rather designed to make Amazon Redshift (and all its strengths) easy to use with typical data workflows and tools. Here are a few examples:
- You are already using a MapReduce framework to process data in S3. Arbalest could make the results of an Elastic MapReduce job queryable with SQL in Redshift. You can then hand off to Arbalest to define additional ETL in plain old SQL.
- You treat S3 as a catch-all data sink, perhaps persisting JSON messages or events from a message system like Kafka or RabbitMQ. Arbalest can expose some or all of this data into a data warehouse using Redshift. The ecosystem of SQL is now available for dashboards, reports, ad-hoc analysis.
- You have complex pipelines that could benefit from a fast, SQL queryable data sink. Arbalest has support out of the box (arbalest.contrib) to integrate with tools like Luigi to be part of a multi-dependency, multi-step pipeline topology.
Getting started is easy with pip:
pip install arbalest
Examples of Arbalest pipeline are in examples/. An overview of concepts and classes are below.
Note
Arbalest depends on psycopg2. However, installing psycopg2 on Windows may not be straightforward.
To install psycopg2 on Windows:
64 bit Python installation:
pip install -e git+https://github.com/nwcell/psycopg2-windows.git@win64-py27#egg=psycopg2
32 bit Python installation:
pip install -e git+https://github.com/nwcell/psycopg2-windows.git@win32-py27#egg=psycopg2
Arbalest orchestrates data loading using pipelines. Each Pipeline
can have one or many steps that are made up of three parts:
metadata: Path in an S3 bucket to store information needed for the copy process.
source: Path in an S3 bucket where data to be copied from is located consisting of JSON object files:
{ "id": "66bc8153-d6d9-4351-bada-803330f22db7", "someNumber": 1 }
schema: Definition of JSON objects to map into Redshift rows.
A schema is defined using a JsonObject mapper which consists of one or many Property declarations.
By default the name of the JSON property is used as the column, but can be set
to a custom column name. Column names have a
maximum length of 127 characters. Column names
longer than 127 characters will be truncated.
Nested properties will create a default column name delimited by an underscore.
Example JSON Object (whitespace for clarity):
{
"id": "66bc8153-d6d9-4351-bada-803330f22db7",
"someNumber": 1,
"child" : {
"someBoolean": true
}
}
Example Schema:
JsonObject('destination_table_name',
Property('id', 'VARCHAR(36)'),
Property('someNumber', 'INTEGER', 'custom_column_name'),
Property('child', Property('someBoolean', 'BOOLEAN')))The S3CopyPipeline supports different strategies for copying data from S3 to Redshift.
Bulk copy
Bulk copy imports all keys in an S3 path into a Redshift table using a staging table. By dropping and reimporting all data, duplication is eliminated. This type of copy is useful for data that does not change very often or will only be ingested once (e.g. immutable time series).
Manifest copy
A manifest copy imports all keys in an S3 path into a Redshift table using a manifest.
In addition, a journal of successfully imported objects is persisted to the metadata path.
Subsequent runs of this copy step will only copy S3 keys that do not exist in the journal.
This type of copy is useful for data in a path that changes often.
Example data copies:
#!/usr/bin/env python
import psycopg2
from arbalest.configuration import env
from arbalest.redshift import S3CopyPipeline
from arbalest.redshift.schema import JsonObject, Property
if __name__ == '__main__':
pipeline = S3CopyPipeline(
aws_access_key_id=env('AWS_ACCESS_KEY_ID'),
aws_secret_access_key=env('AWS_SECRET_ACCESS_KEY'),
bucket=env('BUCKET_NAME'),
db_connection=psycopg2.connect(env('REDSHIFT_CONNECTION')))
pipeline.bulk_copy(metadata='path_to_save_pipeline_metadata',
source='path_of_source_data',
schema=JsonObject('destination_table_name',
Property('id', 'VARCHAR(36)'),
Property('someNumber', 'INTEGER',
'custom_column_name')))
pipeline.manifest_copy(metadata='path_to_save_pipeline_metadata',
source='path_of_incremental_source_data',
schema=JsonObject('incremental_destination_table_name',
Property('id', 'VARCHAR(36)'),
Property('someNumber', 'INTEGER',
'custom_column_name')))
pipeline.run()Pipelines can also have arbitrary SQL steps. Each SQL step can have one or many statements which are executed in a transaction, for example, orchestrating additional ETL (extract, transform, and load). Expanding on the previous example:
#!/usr/bin/env python
import psycopg2
from arbalest.configuration import env
from arbalest.redshift import S3CopyPipeline
from arbalest.redshift.schema import JsonObject, Property
if __name__ == '__main__':
pipeline = S3CopyPipeline(
aws_access_key_id=env('AWS_ACCESS_KEY_ID'),
aws_secret_access_key=env('AWS_SECRET_ACCESS_KEY'),
bucket=env('BUCKET_NAME'),
db_connection=psycopg2.connect(env('REDSHIFT_CONNECTION')))
pipeline.bulk_copy(metadata='path_to_save_pipeline_metadata',
source='path_of_source_data',
schema=JsonObject('destination_table_name',
Property('id', 'VARCHAR(36)'),
Property('someNumber', 'INTEGER',
'custom_column_name')))
pipeline.manifest_copy(metadata='path_to_save_pipeline_metadata',
source='path_of_incremental_source_data',
schema=JsonObject('incremental_destination_table_name',
Property('id', 'VARCHAR(36)'),
Property('someNumber', 'INTEGER',
'custom_column_name')))
pipeline.sql(('SELECT someNumber + %s '
'INTO some_olap_table FROM destination_table_name', 1),
('SELECT * INTO destination_table_name_copy '
'FROM destination_table_name'))
pipeline.run()Included in this project are a variety of orchestration helpers to assist with
the creation of pipelines.
These classes are defined in the arbalest.pipeline and arbalest.contrib modules.
Assuming source data is stored in a sortable series of directories, S3SortedDataSources
facilitates the retrieval of S3 paths in a sequence for import, given a start
and/or end. In addition, it has methods to mark a cursor in an S3 persisted journal.
Examples of data stored as a sorted series
Sequential integers:
s3://bucket/child/1/* s3://bucket/child/2/* s3://bucket/child/3/*
Time series:
s3://bucket/child/2015-01-01/* s3://bucket/child/2015-01-02/* s3://bucket/child/2015-01-03/* s3://bucket/child/2015-01-04/00/*
Example of sorted data source class
S3SortedDataSources(
metadata='',
source='child',
bucket=bucket,
start=env('START'),
end=env('END'))SqlTimeSeriesImport implements a bulk copy and update strategy of data from
a list of time series sources from S3SortedDataSources into an existing
target table.
Example time series import from an S3 time series topology, ingesting a day of objects
Time series path topology:
s3://bucket/child/2015-01-01/* s3://bucket/child/2015-01-02/*
ExamplePipeline(S3CopyPipeline):
def __init__(self,
aws_access_key_id,
aws_secret_access_key,
bucket,
db_connection):
super(ExamplePipeline, self).__init__(
aws_access_key_id,
aws_secret_access_key,
bucket,
db_connection)
# Create table to ingest data into if it does not exist
self.sql('CREATE target_table IF NOT EXISTS target_table(id VARCHAR(36), someNumber INTEGER, timestamp TIMESTAMP);')
time_series = SqlTimeSeriesImport(
destination_table='target_table',
update_date='2015-01-01', # Replace existing events, if any, after this timestamp
sources=S3SortedDataSources(
metadata='',
source='child',
bucket=bucket,
start='2015-01-01',
end='2015-01-02'),
Property('id', 'VARCHAR(36)'),
Property('someNumber', 'INTEGER'),
Property('timestamp', 'TIMESTAMP'))
# Populate target_table using a bulk copy per day
time_series.bulk_copy(
pipeline=self,
metadata='',
max_error=1000, # Maximum errors tolerated by Redshift COPY
order_by_column='timestamp') # Use column named timestamp to sort by and replace existing events, if anyPipelineTask wraps any arbalest.core.Pipeline into a Luigi Task.
This allows for the composition of workflows with dependency graphs, for example,
data pipelines that are dependent on multiple steps or other pipelines. Luigi then takes care of
the heavy lifting of
scheduling and executing
multistep pipelines.
Arbalest is licensed under the MIT License.
Arbalest was built at Dwolla, primarily by Fredrick Galoso. Initial support for Luigi and contributions to orchestration helpers by Hayden Goldstien. We gladly welcome contributions and feedback. If you are using Arbalest we would love to know.