Use Apache Beam and Bigtable to enrich data

Run in Google Colab

View source on GitHub

This notebook shows how to enrich data by using the Apache Beam enrichment transform with Bigtable. The enrichment transform is an Apache Beam turnkey transform that lets you enrich data by using a key-value lookup. This transform has the following features:

• The transform has a built-in Apache Beam handler that interacts with Bigtable to get data to use in the enrichment.
• The enrichment transform uses client-side throttling to manage rate limiting the requests. The requests are exponentially backed off with a default retry strategy. You can configure rate limiting to suit your use case.

This notebook demonstrates the following ecommerce use case:

A stream of online transaction from Pub/Sub contains the following fields: sale_id, product_id, customer_id, quantity, and price. Additional customer demographic data is stored in a separate Bigtable cluster. The demographic data is used to enrich the event stream from Pub/Sub. Then, the enriched data is used to predict the next product to recommended to a customer.

Before you begin

Set up your environment and download dependencies.

Install Apache Beam

To use the enrichment transform with the built-in Bigtable handler, install the Apache Beam SDK version 2.54.0 or later.

pip install torch
pip install apache_beam[interactive,gcp]==2.54.0 --quiet

import datetime
import json
import math

from typing import Any
from typing import Dict

import torch
from google.cloud import pubsub_v1
from google.cloud.bigtable import Client
from google.cloud.bigtable import column_family

import apache_beam as beam
import apache_beam.runners.interactive.interactive_beam as ib
from apache_beam.ml.inference.base import RunInference
from apache_beam.ml.inference.pytorch_inference import PytorchModelHandlerTensor
from apache_beam.options import pipeline_options
from apache_beam.runners.interactive.interactive_runner import InteractiveRunner
from apache_beam.transforms.enrichment import Enrichment
from apache_beam.transforms.enrichment_handlers.bigtable import BigTableEnrichmentHandler

Authenticate with Google Cloud

This notebook reads data from Pub/Sub and Bigtable. To use your Google Cloud account, authenticate this notebook. To prepare for this step, replace <PROJECT_ID>, <INSTANCE_ID>, and <TABLE_ID> with the appropriate values for your setup. These fields are used with Bigtable.

PROJECT_ID = "<PROJECT_ID>"
INSTANCE_ID = "<INSTANCE_ID>"
TABLE_ID = "<TABLE_ID>"

from google.colab import auth
auth.authenticate_user(project_id=PROJECT_ID)

Train the model

Create sample data by using the format [product_id, quantity, price, customer_id, customer_location, recommend_product_id].

data = [
    [3, 5, 127, 9, 'China', 7], [1, 6, 167, 5, 'Peru', 4], [5, 4, 91, 2, 'USA', 8], [7, 2, 52, 1, 'India', 4], [1, 8, 118, 3, 'UK', 8], [4, 6, 132, 8, 'Mexico', 2],
    [6, 3, 154, 6, 'Brazil', 3], [4, 7, 163, 1, 'India', 7], [5, 2, 80, 4, 'Egypt', 9], [9, 4, 107, 7, 'Bangladesh', 1], [2, 9, 192, 8, 'Mexico', 4], [4, 5, 116, 5, 'Peru', 8],
    [8, 1, 195, 1, 'India', 7], [8, 6, 153, 5, 'Peru', 1], [5, 3, 120, 6, 'Brazil', 2], [2, 7, 187, 7, 'Bangladesh', 4], [1, 8, 103, 6, 'Brazil', 8], [2, 9, 181, 1, 'India', 8],
    [6, 5, 166, 3, 'UK', 5], [3, 4, 115, 8, 'Mexico', 1], [4, 7, 170, 4, 'Egypt', 2], [9, 3, 141, 7, 'Bangladesh', 3], [9, 3, 157, 1, 'India', 2], [7, 6, 128, 9, 'China', 1],
    [1, 8, 102, 3, 'UK', 4], [5, 2, 107, 4, 'Egypt', 6], [6, 5, 164, 8, 'Mexico', 9], [4, 7, 188, 5, 'Peru', 1], [8, 1, 184, 1, 'India', 2], [8, 6, 198, 2, 'USA', 5],
    [5, 3, 105, 6, 'Brazil', 7], [2, 7, 162, 7, 'Bangladesh', 7], [1, 8, 133, 9, 'China', 3], [2, 9, 173, 1, 'India', 7], [6, 5, 183, 5, 'Peru', 8], [3, 4, 191, 3, 'UK', 6],
    [4, 7, 123, 2, 'USA', 5], [9, 3, 159, 8, 'Mexico', 2], [9, 3, 146, 4, 'Egypt', 8], [7, 6, 194, 1, 'India', 8], [3, 5, 112, 6, 'Brazil', 1], [4, 6, 101, 7, 'Bangladesh', 2],
    [8, 1, 192, 4, 'Egypt', 4], [7, 2, 196, 5, 'Peru', 6], [9, 4, 124, 9, 'China', 7], [3, 4, 129, 5, 'Peru', 6], [6, 3, 151, 8, 'Mexico', 9], [5, 7, 114, 7, 'Bangladesh', 4],
    [4, 7, 175, 6, 'Brazil', 5], [1, 8, 121, 1, 'India', 2], [4, 6, 187, 2, 'USA', 5], [6, 5, 144, 9, 'China', 9], [9, 4, 103, 5, 'Peru', 3], [5, 3, 84, 3, 'UK', 1],
    [3, 5, 193, 2, 'USA', 4], [4, 7, 135, 1, 'India', 1], [7, 6, 148, 8, 'Mexico', 8], [1, 6, 160, 5, 'Peru', 7], [8, 6, 155, 6, 'Brazil', 9], [5, 7, 183, 7, 'Bangladesh', 2],
    [2, 9, 125, 4, 'Egypt', 4], [6, 3, 111, 9, 'China', 9], [5, 2, 132, 3, 'UK', 3], [4, 5, 104, 7, 'Bangladesh', 7], [2, 7, 177, 8, 'Mexico', 7]]

countries_to_id = {'India': 1, 'USA': 2, 'UK': 3, 'Egypt': 4, 'Peru': 5,
                   'Brazil': 6, 'Bangladesh': 7, 'Mexico': 8, 'China': 9}

Preprocess the data:

1. Convert the lists to tensors.
2. Separate the features from the expected prediction.

X = [torch.tensor(item[:4]+[countries_to_id[item[4]]], dtype=torch.float) for item in data]
Y = [torch.tensor(item[-1], dtype=torch.float) for item in data]

Define a simple model that has five input features and predicts a single value.

def build_model(n_inputs, n_outputs):
  """build_model builds and returns a model that takes
  `n_inputs` features and predicts `n_outputs` value"""
  return torch.nn.Sequential(
      torch.nn.Linear(n_inputs, 8),
      torch.nn.ReLU(),
      torch.nn.Linear(8, 16),
      torch.nn.ReLU(),
      torch.nn.Linear(16, n_outputs))

Train the model.

model = build_model(n_inputs=5, n_outputs=1)

loss_fn = torch.nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters())

for epoch in range(1000):
  print(f'Epoch {epoch}: ---')
  optimizer.zero_grad()
  for i in range(len(X)):
    pred = model(X[i])
    loss = loss_fn(pred, Y[i])
    loss.backward()
  optimizer.step()

Save the model to the STATE_DICT_PATH variable.

STATE_DICT_PATH = './model.pth'

torch.save(model.state_dict(), STATE_DICT_PATH)

Set up the Bigtable table

Create a sample Bigtable table for this notebook.

# Connect to the Bigtable instance. If you don't have admin access, then drop `admin=True`.
client = Client(project=PROJECT_ID, admin=True)
instance = client.instance(INSTANCE_ID)

# Create a column family.
column_family_id = 'demograph'
max_versions_rule = column_family.MaxVersionsGCRule(2)
column_families = {column_family_id: max_versions_rule}

# Create a table.
table = instance.table(TABLE_ID)

# You need admin access to use `.exists()`. If you don't have the admin access, then
# comment out the if-else block.
if not table.exists():
  table.create(column_families=column_families)
else:
  print("Table %s already exists in %s:%s" % (TABLE_ID, PROJECT_ID, INSTANCE_ID))

Add rows to the table for the enrichment example.

# Define column names for the table.
customer_id = 'customer_id'
customer_name = 'customer_name'
customer_location = 'customer_location'

# The following data is sample data to insert into Bigtable.
customers = [
  {
    'customer_id': 1, 'customer_name': 'Sam', 'customer_location': 'India'
  },
  {
    'customer_id': 2, 'customer_name': 'John', 'customer_location': 'USA'
  },
  {
    'customer_id': 3, 'customer_name': 'Travis', 'customer_location': 'UK'
  },
]

for customer in customers:
  row_key = str(customer[customer_id]).encode()
  row = table.direct_row(row_key)
  row.set_cell(
    column_family_id,
    customer_id.encode(),
    str(customer[customer_id]),
    timestamp=datetime.datetime.utcnow())
  row.set_cell(
    column_family_id,
    customer_name.encode(),
    customer[customer_name],
    timestamp=datetime.datetime.utcnow())
  row.set_cell(
    column_family_id,
    customer_location.encode(),
    customer[customer_location],
    timestamp=datetime.datetime.utcnow())
  row.commit()
  print('Inserted row for key: %s' % customer[customer_id])

Inserted row for key: 1
Inserted row for key: 2
Inserted row for key: 3

Publish messages to Pub/Sub

Use the Pub/Sub Python client to publish messages.

# Replace <TOPIC_NAME> with the name of your Pub/Sub topic.
TOPIC = "<TOPIC_NAME>"

# Replace <SUBSCRIPTION_PATH> with the subscription for your topic.
SUBSCRIPTION = "<SUBSCRIPTION_PATH>"

messages = [
    {'sale_id': i, 'customer_id': i, 'product_id': i, 'quantity': i, 'price': i*100}
    for i in range(1,4)
  ]

publisher = pubsub_v1.PublisherClient()
topic_name = publisher.topic_path(PROJECT_ID, TOPIC)

for message in messages:
  data = json.dumps(message).encode('utf-8')
  publish_future = publisher.publish(topic_name, data)

Use the Bigtable enrichment handler

The BigTableEnrichmentHandler is a built-in handler included in the Apache Beam SDK versions 2.54.0 and later.

Configure the BigTableEnrichmentHandler handler with the following required parameters:

• project_id: the Google Cloud project ID for the Bigtable instance
• instance_id: the instance name of the Bigtable cluster
• table_id: the table ID of table containing relevant data
• row_key: The field name from the input row that contains the row key to use when querying Bigtable.

Optionally, you can use parameters to further configure the BigTableEnrichmentHandler handler. For more information about the available parameters, see the enrichment handler module documentation.

Note: When exceptions occur, by default, the logging severity is set to warning (ExceptionLevel.WARN). To configure the severity to raise exceptions, set exception_level to ExceptionLevel.RAISE. To ignore exceptions, set exception_level to ExceptionLevel.QUIET.

The following example demonstrates how to set the exception level in the BigTableEnrichmentHandler handler:

bigtable_handler = BigTableEnrichmentHandler(project_id=PROJECT_ID,
                                             instance_id=INSTANCE_ID,
                                             table_id=TABLE_ID,
                                             row_key=row_key,
                                             exception_level=ExceptionLevel.RAISE)

The row_key parameter represents the field in input schema (beam.Row) that contains the row key for a row in the table.

Starting with Apache Beam version 2.54.0, you can perform either of the following tasks when a table uses composite row keys:

• Modify the input schema to contain the row key in the format required by Bigtable.
• Use a custom enrichment handler. For more information, see the example handler with composite row key support.

row_key = 'customer_id'

bigtable_handler = BigTableEnrichmentHandler(project_id=PROJECT_ID,
                                             instance_id=INSTANCE_ID,
                                             table_id=TABLE_ID,
                                             row_key=row_key)

Use the enrichment transform

To use the enrichment transform, the enrichment handler parameter is the only required parameter.

The following example demonstrates the code needed to add this transform to your pipeline.

with beam.Pipeline() as p:
  output = (p
            ...
            | "Enrich with BigTable" >> Enrichment(bigtable_handler)
            | "RunInference" >> RunInference(model_handler)
            ...
            )

By default, the enrichment transform performs a cross_join. This join returns the enriched row with the following fields: sale_id, customer_id, product_id, quantity, price, and customer_location.

To make a prediction when running the ecommerce example, however, the trained model needs the following fields: product_id, quantity, price, customer_id, and customer_location.

Therefore, to get the required fields for the ecommerce example, design a custom join function that takes two dictionaries as input and returns an enriched row that include these fields.

def custom_join(left: Dict[str, Any], right: Dict[str, Any]):
  enriched = {}
  enriched['product_id'] = left['product_id']
  enriched['quantity'] = left['quantity']
  enriched['price'] = left['price']
  enriched['customer_id'] = left['customer_id']
  enriched['customer_location'] = right['demograph']['customer_location']
  return beam.Row(**enriched)

To provide a lambda function for using a custom join with the enrichment transform, see the following example.

with beam.Pipeline() as p:
  output = (p
            ...
            | "Enrich with BigTable" >> Enrichment(bigtable_handler, join_fn=custom_join)
            | "RunInference" >> RunInference(model_handler)
            ...
            )

Because the enrichment transform makes API calls to the remote service, use the timeout parameter to specify a timeout duration of 10 seconds:

with beam.Pipeline() as p:
  output = (p
            ...
            | "Enrich with BigTable" >> Enrichment(bigtable_handler, join_fn=custom_join, timeout=10)
            | "RunInference" >> RunInference(model_handler)
            ...
            )

Use the PyTorchModelHandlerTensor interface to run inference

Because the enrichment transform outputs data in the format beam.Row, to make it compatible with the PyTorchModelHandlerTensor interface, convert it to torch.tensor. Additionally, the enriched field customer_location is a string type, but the model requires a float type. Convert the customer_location field to a float type.

def convert_row_to_tensor(element: beam.Row):
  row_dict = element._asdict()
  row_dict['customer_location'] = countries_to_id[row_dict['customer_location']]
  return torch.tensor(list(row_dict.values()), dtype=torch.float)

Initialize the model handler with the preprocessing function.

model_handler = PytorchModelHandlerTensor(state_dict_path=STATE_DICT_PATH,
                                          model_class=build_model,
                                          model_params={'n_inputs':5, 'n_outputs':1}
                                          ).with_preprocess_fn(convert_row_to_tensor)

Define a DoFn to format the output.

class PostProcessor(beam.DoFn):
  def process(self, element, *args, **kwargs):
    print('Customer %d who bought product %d is recommended to buy product %d' % (element.example[3], element.example[0], math.ceil(element.inference[0])))

Run the pipeline

Configure the pipeline to run in streaming mode.

options = pipeline_options.PipelineOptions()
options.view_as(pipeline_options.StandardOptions).streaming = True # Streaming mode is set True

Pub/Sub sends the data in bytes. Convert the data to beam.Row objects by using a DoFn.

class DecodeBytes(beam.DoFn):
  """
  The DecodeBytes `DoFn` converts the data read from Pub/Sub to `beam.Row`.
  First, decode the encoded string. Convert the output to
  a `dict` with `json.loads()`, which is used to create a `beam.Row`.
  """
  def process(self, element, *args, **kwargs):
    element_dict = json.loads(element.decode('utf-8'))
    yield beam.Row(**element_dict)

Use the following code to run the pipeline.

Note: Because this pipeline is a streaming pipeline, you need to manually stop the cell. If you don't stop the cell, the pipeline continues to run.

with beam.Pipeline(options=options) as p:
  _ = (p
       | "Read from Pub/Sub" >> beam.io.ReadFromPubSub(subscription=SUBSCRIPTION)
       | "ConvertToRow" >> beam.ParDo(DecodeBytes())
       | "Enrichment" >> Enrichment(bigtable_handler, join_fn=custom_join, timeout=10)
       | "RunInference" >> RunInference(model_handler)
       | "Format Output" >> beam.ParDo(PostProcessor())
       )

Customer 1 who bought product 1 is recommended to buy product 3
Customer 2 who bought product 2 is recommended to buy product 5
Customer 3 who bought product 3 is recommended to buy product 7