Managing data workflows with Luigi
7 likes•6,267 views
The document discusses managing data workflows, particularly through the use of Luigi, a tool developed by Spotify for data pipelines. It details the extraction, transformation, and loading (ETL) processes involved in customer analytics and emphasizes the importance of handling complex data workflows. Additionally, it highlights features of Luigi such as task dependencies, targets, and the ability to resume workflows after failures.
![Target
class MongoTarget(Luigi.Target):
def __init__(self, database, collection, predicate):
self.client = MongoClient()
self.database = database
self.collection = collection
self.predicate = predicate
def exists(self):
db = self.client[self.database]
one = db[self.collection].find_one(self.predicate)
return one is not None](https://cdn.statically.io/img/image.slidesharecdn.com/luigiatmetricsmonday-160607061302/85/Managing-data-workflows-with-Luigi-19-320.jpg)
![Multiple dependencies
class TransformAllSales(Luigi.Task):
def requires(self):
for i in range(1000):
return [ImportInitialSaleFile(index=i)]
def run(self): #…
def output(self): #…](https://cdn.statically.io/img/image.slidesharecdn.com/luigiatmetricsmonday-160607061302/85/Managing-data-workflows-with-Luigi-30-320.jpg)
Managing data workflows with Luigi
- 1. 1 by Teemu Kurppa www.teemukurppa.net Metrics Monday at Custobar, Helsinki, 30.5.2016 Managing data workflows with Luigi
- 2. 2 Customer analytics and marketing tool for retailers I’m an advisor at your host:
- 3. teemu@ouraring.com www.ouraring.com the world's first wellness ring Head of Software: Cloud & Mobile I work at
- 5. gunzip -c /var/log/syslog.3.gz | grep -e UFW
- 6. Complex data workflow Let’s analyse if the weather affects sleep quality: • Get sleep data of all study participants • Get location data of all study participants • Fetch weather data for each day and location • Fetch historical weather data for each location • Calculate difference from an average weather for each data point • Do a statistical analysis over users and days, comparing weather data and sleep quality data A lot of can go wrong on each step. Rerunning takes time
- 7. Case Custobar:ETL Extract - Transform - Load
- 8. Case Custobar:ETL Fetch custom sales.csv from SFTP Transform custom sales.csv to standard sales.json Validate and throw away invalid fields Load valid sales data to database
- 9. Case Custobar:ETL Fetch custom sales.csv from SFTP Transform custom sales.csv to standard sales.json Validate and throw away invalid fields Load valid sales data to database Transform Load Extract
- 10. Case Custobar:ETL Fetch custom sales.csv from SFTP Transform custom sales.csv to standard sales.json Validate and throw away invalid fields Load valid sales data to database Do this, for millions of rows of initial data, and continue doing it every day, for products customers sales
- 12. Data workflow tools Pinball by Pinterest Luigi by Spotify Airflow by AirBnB
- 14. Luigi Concepts Get Changed Customers sql: Customers table Tasks Targets Export Changed Customers to FTP file://data/ customers.csv sftp://data/ customers.csv Dependencies
- 15. Luigi Concepts Get Changed Customers sql: Customers table Tasks Targets Export Changed Customers to FTP file://data/ customers.csv sftp://data/ customers.csv Dependencies output()input() input() output() requires()
- 16. Luigi Concepts Get Changed Customers sql: Customers table Tasks Targets Export Changed Customers to FTP file://data/ customers.csv sftp://data/ customers.csv Dependencies company: Parameter date: DateParameter company: Parameter date: DateParameter Parameters
- 18. Target Target is simply something that exists or doesn’t exist For example • a file in a local file system • a file in a remote file system • a file in an Amazon S3 bucket • a database row in a SQL database
- 19. Target class MongoTarget(Luigi.Target): def __init__(self, database, collection, predicate): self.client = MongoClient() self.database = database self.collection = collection self.predicate = predicate def exists(self): db = self.client[self.database] one = db[self.collection].find_one(self.predicate) return one is not None
- 20. Target Lots of ready-made targets in Luigi: • local file • HDFS file • S3 key/value target • SSH remote target • SFTP remote target • SQL table row target • Amazon Redshift table row target • ElasticSearch target
- 22. Task: basic structure class TransformDailySalesCSVtoJSON(Luigi.Task): def requires(self): #… def run(self): # … def output(self): #…
- 23. Task: parameters class TransformDailySalesCSVtoJSON(Luigi.Task): date = luigi.DateParameter() def requires(self): #… def run(self): # … def output(self): #…
- 24. Task: requires class TransformDailySalesCSVtoJSON(Luigi.Task): date = luigi.DateParameter() def requires(self): return ImportDailyCSVFromSFTP(self.date) def run(self): # … def output(self): #…
- 25. Task: output class TransformDailySalesCSVtoJSON(Luigi.Task): date = luigi.DateParameter() def requires(self): # … def run(self): # … def output(self): path = “/d/sales_%s.json” % (self.date.stftime(‘%Y%m%d’)) return luigi.LocalTarget(path)
- 26. Task: run class TransformDailySalesCSVtoJSON(Luigi.Task): date = luigi.DateParameter() def requires(self): #… def run(self): # Note: luigi’s input() and output() takes care of atomicity with self.input().open(‘r’) as infile: data = transform_csv_to_dict(infile) with self.output().open(‘w’) as outfile: json.dump(data, outfile) def output(self): #…
- 27. Task class TransformDailySalesCSVtoJSON(Luigi.Task): date = luigi.DateParameter() def requires(self): return ImportDailyCSVFromSFTP(self.date) def run(self): with self.input().open(‘r’) as infile: data = transform_csv_to_dict(infile) with self.output().open(‘w’) as outfile: json.dump(data, outfile) def output(self): path = “/d/sales_%s.json” % (self.date.stftime(‘%Y%m%d’)) return luigi.LocalTarget(path)
- 28. Tasks Lots of ready-made tasks in Luigi: • dump data to SQL table • copy to Redshift Table • run Hadoop job • query SalesForce • copy to Redshift Table • Load ElasticSearch index • …
- 30. Multiple dependencies class TransformAllSales(Luigi.Task): def requires(self): for i in range(1000): return [ImportInitialSaleFile(index=i)] def run(self): #… def output(self): #…
- 31. Dynamic dependencies class LoadDailyAPIData(Luigi.Task): date = luigi.DateParameter() def run(self): for filepath in os.listdir(‘/d/api_data/*.json’): TransformDailyAPIData(filepath)
- 32. Wrapper task class LoadAllDailyData(Luigi.WrapperTask): date = luigi.DateParameter() def run(self): yield LoadDailyProducts(self.date) yield LoadDailyCustomers(self.date) yield LoadDailySales(self.date)
- 33. Why to use data workflow tools? 33
- 34. 34 1. Resume the data workflow after a failure 2. Parametrize and rerun tasks every day 3. Organise code with shared patterns
- 35. 35 Thanks! Questions? Custobar is hiring! Approach Juha, Tatu or me to learn more Follow @teemu on Twitter to stay in touch.