In a previous post, we discussed why forecasting models and engines are in high demand and ways in which AI and machine learning (ML) technologies have been satisfying the needs of organizations that want to take advantage of the new Data Economy.

Now with the Amazon Forecast service (announced November 2018), companies can easily leverage ML to combine time series data with additional variables to build forecasts and estimate operational metrics, business metrics, resource requirements, and more.

Building from the example we described in our previous post, we will now go through the simple Python script you can deploy to run the integration between Snowflake and Amazon Forecast. As a quick recap, here’s how the integration works:

Integration Steps:

  1. Extract time series: The user isolates a set of time series training data from the user’s Snowflake database and saves it to Amazon S3.
  2. Forecast: The user runs the data through Amazon Forecast using a Python script, receives a baseline forecast, and then loads the data back into Snowflake.
  3. Connect to a share: The user connects to another Snowflake user’s data through Snowflake Secure Data Sharing. The first user then isolates one or more time series data sets from the other user and saves those to Amazon S3.
  4. Re-forecast with enriched data:To receive an improved time series forecast, the user runs both the original and shared time series data sets through Amazon Forecast using a Python script. The user also provides the shared time as additional inputs to Amazon Forecast as a related time series. Amazon Forecast then uses the inputs to improve the accuracy of the forecast. The user then loads the resulting forecast into Snowflake.

The Integration Script

First, we connect to the Amazon account, and then we connect to the Snowflake account, defining the data owner and data consumer. The owner is the account that will be sharing data, and the consumer is the account that will read the shared data (named the Processor account in our code example).

# coding=utf-8
import snowflake.connector
import boto3
import subprocess
from time import sleep


#DEFINE PROCESSOR PARAMETERS
PROCESSOR_ACCOUNT =  '<ACCOUNT_NAME>'
PROCESSOR_USER =  '<USER_NAME>'
PROCESSOR_PASSWORD = '<PASSWORD>'
PROCESSOR_WAREHOUSE = 'FORECAST_WH'
PROCESSOR_DATABASE =  'FORECAST_RESULTS'
PROCESSOR_SCHEMA = 'FORECAST_DATA'
PROCESSOR_TABLE = 'FORECAST_RESULTS'
PROCESSOR_STAGE = 'FORECAST_STAGE'
PROCESSOR_S3 = '<S3_PATH>'
PROCESSOR_S3_PUBLIC_KEY = '<OWNER_PUBLIC_KEY>'
PROCESSOR_S3_PRIVATE_KEY = '<OWNER_PRIVATE_KEY>'


#DEFINE OWNER PARAMETERS
OWNER_ACCOUNT = '<ACCOUNT_NAME>'
OWNER_USER = '<USER_NAME>'
OWNER_PASSWORD = '<PASSWORD>'
OWNER_WAREHOUSE = 'FORECAST_WH'
OWNER_DATABASE = 'FORECAST_DATA'
OWNER_SCHEMA = 'FORECAST_DATA'
OWNER_TABLE = 'FORECAST_DATA'
OWNER_STAGE = 'FORECAST_STAGE'
OWNER_S3 = '<S3_PATH>'
OWNER_S3_PUBLIC_KEY = '<OWNER_PUBLIC_KEY>'
OWNER_S3_PRIVATE_KEY = '<OWNER_PRIVATE_KEY>'

#DEFINE FORECAST VARIABLES 
S3_BUCKET = PROCESSOR_S3 
FILENAME = 'data_0_0_0.csv.gz' 

DATASETNAME = 'snowflake_ds_1' 
DATASETGROUPNAME = 'snowflake_dsg_1' 
PREDICTORNAME = 'snowflake_f_1' 
RAW_FILEPATH = PROCESSOR_S3 + '/raw/' + FILENAME

def snowflake_connect(ACCOUNT, USER, PASSWORD, WAREHOUSE): 
       con = snowflake.connector.connect( 
       account = ACCOUNT, 
       user = USER, 
       password = PASSWORD, 
       ) 

cur = con.cursor() 
cur.execute('USE ROLE ACCOUNTADMIN') 
cur.execute('USE WAREHOUSE ' + WAREHOUSE) 

return con.cursor()

Now we are going to create and set up the new database and tables, and then load the data set into Snowflake.

def setup_owner():
       #SETUP INITIAL DATASET
       cursor = snowflake_connect(OWNER_ACCOUNT, OWNER_USER, 
OWNER_PASSWORD, OWNER_WAREHOUSE)
       cursor.execute('CREATE OR REPLACE DATABASE ' + OWNER_DATABASE)
       cursor.execute('CREATE OR REPLACE SCHEMA ' + OWNER_SCHEMA)
       cursor.execute('CREATE OR REPLACE STAGE ' + OWNER_STAGE + ' 
url='' + OWNER_S3 + '' credentials=(aws_key_id='' + 
OWNER_S3_PUBLIC_KEY + '' aws_secret_key='' + OWNER_S3_PRIVATE_KEY + '')')
       cursor.execute('CREATE OR REPLACE TABLE ' + OWNER_TABLE + ' 
(ts timestamp, demand float, id string)')
       cursor.execute('COPY INTO ' + OWNER_TABLE + ' FROM @' + 
OWNER_STAGE + '/item-demand-time.csv')

The next step is to create a share (that is, a shareable object) in Snowflake and mount the database.

It is important to note that Snowflake is not copying data to a different area for sharing; all sharing is accomplished through Snowflake’s unique services layer and metadata store, which means that instead of creating duplicate copies, we are using a pointer to the same original copy of the data.

Snowflake’s Secure Data Sharing feature guarantees you always have a single source of truth and avoids the traditional and cumbersome ETL processes required when sharing data across traditional RDBMSs.

As we create the shared database object, we are also granting specific access privileges (GRANT USAGE and GRANT SELECT statements). These privileges can be set for sharing specific objects such as tables, secure views, and secure user-defined functions. We then grant access to the Processor account to this new share.

#CREATE SHARE
       cursor.execute('CREATE OR REPLACE SHARE FORECAST_SHARE')
       cursor.execute('GRANT USAGE ON DATABASE ' + OWNER_DATABASE + ' TO 
SHARE FORECAST_SHARE')
       cursor.execute('GRANT USAGE ON SCHEMA ' + OWNER_SCHEMA + ' TO SHARE 
FORECAST_SHARE')
       cursor.execute('GRANT SELECT ON TABLE ' + OWNER_TABLE + ' TO SHARE 
FORECAST_SHARE')
       cursor.execute('ALTER SHARE FORECAST_SHARE ADD ACCOUNTS = ' + PROCESSOR_ACCOUNT)

def setup_processor():
       #CREATE INITIAL SHARE
       cursor = snowflake_connect(PROCESSOR_ACCOUNT, PROCESSOR_USER, 
PROCESSOR_PASSWORD, PROCESSOR_WAREHOUSE)
       cursor.execute('CREATE OR REPLACE DATABASE ' + OWNER_DATABASE + ' FROM SHARE ' 
+ OWNER_ACCOUNT + '.FORECAST_SHARE')

We now need to create the table to host the results from Amazon Forecast. As we create the resulting table, we are also granting access to the necessary objects (database, schema, table) and granting the owner account access to the database.

#CREATE RESULT SHARE
       cursor.execute('CREATE OR REPLACE DATABASE ' + 
PROCESSOR_DATABASE)
       cursor.execute('CREATE OR REPLACE SCHEMA ' + 
PROCESSOR_SCHEMA)
       cursor.execute('CREATE OR REPLACE STAGE ' + PROCESSOR_STAGE + 
' url='' + PROCESSOR_S3 + '' credentials=(aws_key_id='' + 
PROCESSOR_S3_PUBLIC_KEY + '' aws_secret_key='' + 
PROCESSOR_S3_PRIVATE_KEY + '')')
       cursor.execute('CREATE OR REPLACE TABLE ' + PROCESSOR_TABLE + ' 
(date datetime, first_observation_date datetime, item_id string, 
last_observation_date datetime, mean float, p10 
float, p50 float, p90 float)')
       cursor.execute('CREATE OR REPLACE SHARE FORECAST_RESULT_SHARE')
       cursor.execute('GRANT USAGE ON DATABASE ' + PROCESSOR_DATABASE + ' TO SHARE 
FORECAST_RESULT_SHARE')
       cursor.execute('GRANT USAGE ON SCHEMA ' + PROCESSOR_SCHEMA + ' TO SHARE 
FORECAST_RESULT_SHARE')
       cursor.execute('GRANT SELECT ON TABLE ' + PROCESSOR_TABLE + ' TO SHARE 
FORECAST_RESULT_SHARE')
      cursor.execute('ALTER SHARE FORECAST_RESULT_SHARE ADD ACCOUNTS = ' + OWNER_ACCOUNT)


#Instantiate Forecast Session
session = boto3.Session(region_name='us-west-2')
forecast = session.client(service_name='forecast')
forecastquery = session.client(service_name='forecastquery')

s3 = session.client('s3')
accountId = boto3.client('sts').get_caller_identity().get('Account')
ROLE_ARN = 'arn:aws:iam::%s:role/amazonforecast'%accountId

Next, we configure the data set and schema to be read through Amazon S3, unloading the data from Snowflake and loading it into Amazon Forecast.

#Unload Data From Snowflake to S3

def unload_data_snowflake():
       cursor = snowflake_connect(PROCESSOR_ACCOUNT, PROCESSOR_USER, PROCESSOR_PASSWORD, 
PROCESSOR_WAREHOUSE)
       cursor.execute('USE DATABASE ' + PROCESSOR_DATABASE)
       cursor.execute('USE SCHEMA ' + PROCESSOR_SCHEMA)
       cursor.execute('COPY INTO @' + PROCESSOR_STAGE + '/raw/ FROM (select * FROM ' + 
OWNER_DATABASE + '.' + OWNER_SCHEMA +'.' + OWNER_TABLE + ')')

#Create Dataset 
def create_dataset():
        #forecast.delete_dataset(DatasetName=DATASETNAME)
       schema ={
          "Attributes":[
             {
                "AttributeName":"timestamp",
                "AttributeType":"timestamp"
             },
             {
                "AttributeName":"target_value",
                "AttributeType":"float"
             },
             {
                "AttributeName":"item_id",
                "AttributeType":"string"
             }
          ]
       }

    response = forecast.create_dataset( Domain="CUSTOM", DatasetType='TARGET_TIME_SERIES', 
DataFormat='CSV', DatasetName=DATASETNAME, DataFrequency="H", TimeStampFormat="yyyy-MM-dd 
hh:mm:ss", Schema=schema)

#Create dataset_group
def create_dataset_group():
    forecast.create_dataset_group(DatasetGroupName=DATASETGROUPNAME, RoleArn=ROLE_ARN, 
DatasetNames=[DATASETNAME])

Next, we get the data set ready to be sent to Amazon Forecast and handle potential conflicts. Here we are also creating the Predictor, which will be used to generate the forecast and the recipe that is used to train the model. (Note that several recipes are available depending on your needs; for more information check the Amazon Forecast developer guide.)

#Import DataSet
def import_dataset():
       ds_import_job_response=forecast.create_dataset_import_job(DatasetName=DATASETNAME, 
Delimiter=',', DatasetGroupName=DATASETGROUPNAME, S3Uri=RAW_FILEPATH)
       ds_versionId=ds_import_job_response['VersionId']

       #Wait for File To Finish Loading
       while True:
               dataImportStatus = 
forecast.describe_dataset_import_job(DatasetName=DATASETNAME, 
VersionId=ds_versionId)['Status']
               if (dataImportStatus != 'ACTIVE') and dataImportStatus != 
'FAILED':
                       sleep(30)
               else:
                      break
                       

#Create Recipe
def create_predictor():
        createPredictorResponse = 
forecast.create_predictor(RecipeName='forecast_MQRNN', 
DatasetGroupName=DATASETGROUPNAME, PredictorName=PREDICTORNAME, ForecastHorizo
= 24)
       predictorVersionId=createPredictorResponse['VersionId']

 #Wait for Predictor To Be Created
 while True:
           predictorStatus = 
forecast.describe_predictor(PredictorName=PREDICTORNAME, 
         VersionId=predictorVersionId)['Status']
         if predictorStatus != 'ACTIVE' and predictorStatus != 'FAILED':
                       sleep(30)
               else:
                       break

And now we deploy the Predictor to handle the forecast operation, after which we’ll be able to query with Amazon’s API to retrieve the results.

                      #Deploy Predictor
def deploy_predictor():
        forecast.deploy_predictor(PredictorName=PREDICTORNAME)

 #Wait for Predictor To Be Deployed
       while True:
             deployedPredictorStatus = 
forecast.describe_deployed_predictor(PredictorName=PREDICTORNAME)['Status']
               if deployedPredictorStatus != 'ACTIVE' 
and deployedPredictorStatus != 'FAILED':
                      sleep(30)
               else:
                      break

We can now access and get the forecast, specifying a filter to return our specific data set.

def get_forecast():
       forecastReponse = forecastquery.get_forecast(
               PredictorName=PREDICTORNAME,
               Interval="hour",
               Filters={"item_id":"client_12"}
       )


And finally, we get the data set resulting from Amazon Forecast back to Amazon S3 and into Snowflake using a Forecast Export job from Amazon. Then the data in the Snowflake share will be available for the consumer to read and process.

def export_data():
        forecastInfoList = forecast.list_forecasts(PredictorName=PREDICTORNAME)
['ForecastInfoList']
        forecastId = forecastInfoList[0]['ForecastId']

 #Drop Data Back In S3
       outputPath=S3_BUCKET + "/output"
       forecastExportResponse = 
forecast.create_forecast_export_job(ForecastId=forecastId, OutputPath={"S3Uri": 
outputPath, "RoleArn":ROLE_ARN})
      forecastExportJobId = forecastExportResponse['ForecastExportJobId']

 #Wait for Forecast to be Unloaded
       while True:
             forecastExportStatus = 
forecast.describe_forecast_export_job(ForecastExportJobId=forecastExportJobId)
['Status']
               if forecastExportStatus != 'ACTIVE' and forecastExportStatus != 
'FAILED':
                      sleep(30)
               else:
                      break


def load_data_snowflake():
        cursor = snowflake_connect(PROCESSOR_ACCOUNT, PROCESSOR_USER, PROCESSOR_PASSWORD, 
PROCESSOR_WAREHOUSE)
       cursor.execute('USE DATABASE ' + PROCESSOR_DATABASE)
       cursor.execute('USE SCHEMA ' + PROCESSOR_SCHEMA)
       cursor.execute('COPY INTO ' + PROCESSOR_TABLE + ' FROM @' + 
PROCESSOR_STAGE + '/output file_format=(skip_header=1)')

setup_owner()
setup_processor()
unload_data_snowflake()
create_dataset()
create_dataset_group()
import_dataset()
create_predictor()
deploy_predictor()
get_forecast()
export_data()
load_data_snowflake()

In summary, with this Python script you can run a data set through Amazon Forecast, receive a baseline forecast, and then load the data back into Snowflake. The process is simple because Snowflake does not require preparations before creating and sharing database objects while granting granular access controls, keeping the original data secure, and allowing you to read specific data sets for forecasting operations and processing. You can find the complete Python script for Amazon Forecast on GitHub.

Next up, we’ll finalize this series by going through the final stage of re-forecasting with enriched data and comparing it to the original forecast.

Get the Complete Python Script Here