Archive for category Big Data

IoT Telemetry Collection using Google Protocol Buffers, Google Cloud Functions, Cloud Pub/Sub, and MongoDB Atlas

Posted by Gary A. Stafford in Big Data, Cloud, GCP, Python, Serverless, Software Development on May 21, 2019

Business team meeting. Photo professional investor working new s
Collect IoT sensor telemetry using Google Protocol Buffers’ serialized binary format over HTTPS, serverless Google Cloud Functions, Google Cloud Pub/Sub, and MongoDB Atlas on GCP, as an alternative to integrated Cloud IoT platforms and standard IoT protocols. Aggregate, analyze, and build machine learning models with the data using tools such as MongoDB Compass, Jupyter Notebooks, and Google’s AI Platform Notebooks.

Introduction

Most of the dominant Cloud providers offer IoT (Internet of Things) and IIotT (Industrial IoT) integrated services. Amazon has AWS IoT, Microsoft Azure has multiple offering including IoT Central, IBM’s offering including IBM Watson IoT Platform, Alibaba Cloud has multiple IoT/IIoT solutions for different vertical markets, and Google offers Google Cloud IoT platform. All of these solutions are marketed as industrial-grade, highly-performant, scalable technology stacks. They are capable of scaling to tens-of-thousands of IoT devices or more and massive amounts of streaming telemetry.

In reality, not everyone needs a fully integrated IoT solution. Academic institutions, research labs, tech start-ups, and many commercial enterprises want to leverage the Cloud for IoT applications, but may not be ready for a fully-integrated IoT platform or are resistant to Cloud vendor platform lock-in.

Similarly, depending on the performance requirements and the type of application, organizations may not need or want to start out using IoT/IIOT industry standard data and transport protocols, such as MQTT (Message Queue Telemetry Transport) or CoAP (Constrained Application Protocol), over UDP (User Datagram Protocol). They may prefer to transmit telemetry over HTTP using TCP, or securely, using HTTPS (HTTP over TLS).

Demonstration

In this demonstration, we will collect environmental sensor data from a number of IoT device sensors and stream that telemetry over the Internet to Google Cloud. Each IoT device is installed in a different physical location. The devices contain a variety of common sensors, including humidity and temperature, motion, and light intensity.

Prototype IoT Devices used in this Demonstration

We will transmit the sensor telemetry data as JSON over HTTP to serverless Google Cloud Function HTTPS endpoints. We will then switch to using Google’s Protocol Buffers to transmit binary data over HTTP. We should observe a reduction in the message size contained in the request payload as we move from JSON to Protobuf, which should reduce system latency and cost.

Data received by Cloud Functions over HTTP will be published asynchronously to Google Cloud Pub/Sub. A second Cloud Function will respond to all published events and push the messages to MongoDB Atlas on GCP. Once in Atlas, we will aggregate, transform, analyze, and build machine learning models with the data, using tools such as MongoDB Compass, Jupyter Notebooks, and Google’s AI Platform Notebooks.

For this demonstration, the architecture for JSON over HTTP will look as follows. All sensors will transmit data to a single Cloud Function HTTPS endpoint.

For Protobuf over HTTP, the architecture will look as follows in the demonstration. Each type of sensor will transmit data to a different Cloud Function HTTPS endpoint.

Although the Cloud Functions will automatically scale horizontally to accommodate additional load created by the volume of telemetry being received, there are also other options to scale the system. For example, we could create individual pipelines of functions and topic/subscriptions for each sensor type. We could also split the telemetry data across multiple MongoDB Atlas Collections, based on sensor type, instead of a single collection. In all cases, we will still benefit from the Cloud Function’s horizontal scaling capabilities.

Source Code

All source code is all available on GitHub. Use the following command to clone the project.

git clone \
  --branch master --single-branch --depth 1 --no-tags \
  https://github.com/garystafford/iot-protobuf-demo.git

You will need to adjust the project’s environment variables to fit your own development and Cloud environments. All source code for this post is written in Python. It is intended for Python 3 interpreters but has been tested using Python 2 interpreters. The project’s Jupyter Notebooks can be viewed from within the project on GitHub or using the free, online Jupyter nbviewer.

Technologies

Protocol Buffers

Image result for google developer According to Google, Protocol Buffers (aka Protobuf) are a language- and platform-neutral, efficient, extensible, automated mechanism for serializing structured data for use in communications protocols, data storage, and more. Protocol Buffers are 3 to 10 times smaller and 20 to 100 times faster than XML.

Each protocol buffer message is a small logical record of information, containing a series of strongly-typed name-value pairs. Once you have defined your messages, you run the protocol buffer compiler for your application’s language on your .proto file to generate data access classes.

Google Cloud Functions

According to Google, Cloud Functions is Google’s event-driven, serverless compute platform. Key features of Cloud Functions include automatic scaling, high-availability, fault-tolerance,
no servers to provision, manage, patch or update, only
pay while your code runs, and they easily connect and extend other cloud services. Cloud Functions natively support multiple event-types, including HTTP, Cloud Pub/Sub, Cloud Storage, and Firebase. Current language support includes Python, Go, and Node.

Google Cloud Pub/Sub

According to Google, Cloud Pub/Sub is an enterprise message-oriented middleware for the Cloud. It is a scalable, durable event ingestion and delivery system. By providing many-to-many, asynchronous messaging that decouples senders and receivers, it allows for secure and highly available communication among independent applications. Cloud Pub/Sub delivers low-latency, durable messaging that integrates with systems hosted on the Google Cloud Platform and externally.

MongoDB Atlas

mongodb MongoDB Atlas is a fully-managed MongoDB-as-a-Service, available on AWS, Azure, and GCP. Atlas, a mature SaaS product, offers high-availability, uptime service-level agreements, elastic scalability, cross-region replication, enterprise-grade security, LDAP integration, BI Connector, and much more.

MongoDB Atlas currently offers four pricing plans, Free, Basic, Pro, and Enterprise. Plans range from the smallest, free M0-sized MongoDB cluster, with shared RAM and 512 MB storage, up to the massive M400 MongoDB cluster, with 488 GB of RAM and 3 TB of storage.

Cost Effectiveness of Cloud Functions

At true IIoT scale, Google Cloud Functions may not be the most efficient or cost-effective method of ingesting telemetry data. Based on Google’s pricing model, you get two million free function invocations per month, with each additional million invocations costing USD $0.40. The total cost also includes memory usage, total compute time, and outbound data transfer. If your system is comprised of tens or hundreds of IoT devices, Cloud Functions may prove cost-effective.

However, with thousands of devices or more, each transmitting data multiple times per minutes, you could quickly outgrow the cost-effectiveness of Google Functions. In that case, you might look to Google’s Google Cloud IoT platform. Alternately, you can build your own platform with Google products such as Knative, letting you choose to run your containers either fully managed with the newly-released Cloud Run, or in your Google Kubernetes Engine cluster with Cloud Run on GKE.

Sensor Scripts

For each sensor type, I have developed separate Python scripts, which run on each IoT device. There are two versions of each script, one for JSON over HTTP and one for Protobuf over HTTP.

JSON over HTTPS

Below we see the script, dht_sensor_http_json.py, used to transmit humidity and temperature data via JSON over HTTP to a Google Cloud Function running on GCP. The JSON request payload contains a timestamp, IoT device ID, device type, and the temperature and humidity sensor readings. The URL for the Google Cloud Function is stored as an environment variable, local to the IoT devices, and set when the script is deployed.

import json
import logging
import os
import socket
import sys
import time

import Adafruit_DHT
import requests

URL = os.environ.get('GCF_URL')
JWT = os.environ.get('JWT')
SENSOR = Adafruit_DHT.DHT22
TYPE = 'DHT22'
PIN = 18
FREQUENCY = 15


def main():
    if not URL or not JWT:
        sys.exit("Are the Environment Variables set?")
    get_sensor_data(socket.gethostname())


def get_sensor_data(device_id):
    while True:
        humidity, temperature = Adafruit_DHT.read_retry(SENSOR, PIN)
        payload = {'device': device_id,
                   'type': TYPE,
                   'timestamp': time.time(),
                   'data': {'temperature': temperature,
                            'humidity': humidity}}
        post_data(payload)
        time.sleep(FREQUENCY)


def post_data(payload):
    payload = json.dumps(payload)

    headers = {
        'Content-Type': 'application/json; charset=utf-8',
        'Authorization': JWT
    }

    try:
        requests.post(URL, json=payload, headers=headers)
    except requests.exceptions.ConnectionError:
        logging.error('Error posting data to Cloud Function!')
    except requests.exceptions.MissingSchema:
        logging.error('Error posting data to Cloud Function! Are Environment Variables set?')


if __name__ == '__main__':
    sys.exit(main())

Telemetry Frequency

Although the sensors are capable of producing data many times per minute, for this demonstration, sensor telemetry is intentionally limited to only being transmitted every 15 seconds. To reduce system complexity, potential latency, back-pressure, and cost, in my opinion, you should only produce telemetry data at the frequency your requirements dictate.

JSON Web Tokens

For security, in addition to the HTTPS endpoints exposed by the Google Cloud Functions, I have incorporated the use of a JSON Web Token (JWT). JSON Web Tokens are an open, industry standard RFC 7519 method for representing claims securely between two parties. In this case, the JWT is used to verify the identity of the sensor scripts sending telemetry to the Cloud Functions. The JWT contains an id, password, and expiration, all encrypted with a secret key, which is known to each Cloud Function, in order to verify the IoT device’s identity. Without the correct JWT being passed in the Authorization header, the request to the Cloud Function will fail with an HTTP status code of 401 Unauthorized. Below is an example of the JWT’s payload data.

{
  "sub": "IoT Protobuf Serverless Demo",
  "id": "iot-demo-key",
  "password": "t7J2gaQHCFcxMD6584XEpXyzWhZwRrNJ",
  "iat": 1557407124,
  "exp": 1564664724
}

For this demonstration, I created a temporary JWT using jwt.io. The HTTP Functions are using PyJWT, a Python library which allows you to encode and decode the JWT. The PyJWT library allows the Function to decode and validate the JWT (Bearer Token) from the incoming request’s Authorization header. The JWT token is stored as an environment variable. Deployment instructions are included in the GitHub project.

JSON Payload

Below is a typical JSON request payload (pretty-printed), containing DHT sensor data. This particular message is 148 bytes in size. The message format is intentionally reader-friendly. We could certainly shorten the message’s key fields, to reduce the payload size by an additional 15-20 bytes.

{
  "device": "rp829c7e0e",
  "type": "DHT22",
  "timestamp": 1557585090.476025,
  "data": {
    "temperature": 17.100000381469727,
    "humidity": 68.0999984741211
  }
}

Protocol Buffers

For the demonstration, I have built a Protocol Buffers file, sensors.proto, to support the data output by three sensor types: digital humidity and temperature (DHT), passive infrared sensor (PIR), and digital light intensity (DLI). I am using the newer proto3 version of the protocol buffers language. I have created a common Protobuf sensor message schema, with the variable sensor telemetry stored in the nested data object, within each message type.

It is important to use the correct Protobuf Scalar Value Type to maintain numeric precision in the language you compile for. For simplicity, I am using a double to represent the timestamp, as well as the numeric humidity and temperature readings. Alternately, you could choose Google’s Protobuf WellKnownTypes, Timestamp to store timestamp.

syntax = "proto3";

package sensors;

// DHT22
message SensorDHT {
    string device = 1;
    string type = 2;
    double timestamp = 3;
    DataDHT data = 4;
}

message DataDHT {
    double temperature = 1;
    double humidity = 2;
}

// Onyehn_PIR
message SensorPIR {
    string device = 1;
    string type = 2;
    double timestamp = 3;
    DataPIR data = 4;
}

message DataPIR {
    bool motion = 1;
}

// Anmbest_MD46N
message SensorDLI {
    string device = 1;
    string type = 2;
    double timestamp = 3;
    DataDLI data = 4;
}

message DataDLI {
    bool light = 1;
}

Since the sensor data will be captured with scripts written in Python 3, the Protocol Buffers file is compiled for Python, resulting in the file, sensors_pb2.py.

protoc --python_out=. sensors.proto

Protocol Buffers over HTTPS

Below we see the alternate DHT sensor script, dht_sensor_http_pb.py, which transmits a Protocol Buffers-based binary request payload over HTTPS to a Google Cloud Function running on GCP. Note the request’s Content-Type header has been changed from application/json to application/x-protobuf. In this case, instead of JSON, the same data fields are stored in an instance of the Protobuf’s SensorDHT message type (sensors_pb2.SensorDHT()). Note the import sensors_pb2 statement. This statement imports the compiled Protocol Buffers file, which is stored locally to the script on the IoT device.

import logging
import os
import socket
import sys
import time

import Adafruit_DHT
import requests
import sensors_pb2

URL = os.environ.get('GCF_DHT_URL')
JWT = os.environ.get('JWT')
SENSOR = Adafruit_DHT.DHT22
TYPE = 'DHT22'
PIN = 18
FREQUENCY = 15


def main():
    if not URL or not JWT:
        sys.exit("Are the Environment Variables set?")
    get_sensor_data(socket.gethostname())


def get_sensor_data(device_id):
    while True:
        try:
            humidity, temperature = Adafruit_DHT.read_retry(SENSOR, PIN)

            sensor_dht = sensors_pb2.SensorDHT()
            sensor_dht.device = device_id
            sensor_dht.type = TYPE
            sensor_dht.timestamp = time.time()
            sensor_dht.data.temperature = temperature
            sensor_dht.data.humidity = humidity

            payload = sensor_dht.SerializeToString()

            post_data(payload)
            time.sleep(FREQUENCY)
        except TypeError:
            logging.error('Error getting sensor data!')


def post_data(payload):
    headers = {
        'Content-Type': 'application/x-protobuf',
        'Authorization': JWT
    }

    try:
        requests.post(URL, data=payload, headers=headers)
    except requests.exceptions.ConnectionError:
        logging.error('Error posting data to Cloud Function!')
    except requests.exceptions.MissingSchema:
        logging.error('Error posting data to Cloud Function! Are Environment Variables set?')


if __name__ == '__main__':
    sys.exit(main())

Protobuf Binary Payload

To understand the binary Protocol Buffers-based payload, we can write a sample SensorDHT message to a file on disk as a byte array.

message = sensorDHT.SerializeToString()

binary_file_output = open("./data_binary.txt", "wb")
file_byte_array = bytearray(message)
binary_file_output.write(file_byte_array)

Then, using the hexdump command, we can view a representation of the binary data file.

> hexdump -C data_binary.txt
00000000  0a 08 38 32 39 63 37 65  30 65 12 05 44 48 54 32  |..829c7e0e..DHT2|
00000010  32 1d 05 a0 b9 4e 22 0a  0d ec 51 b2 41 15 cd cc  |2....N"...Q.A...|
00000020  38 42                                             |8B|
00000022

The binary data file size is 48 bytes on disk, as compared to the equivalent JSON file size of 148 bytes on disk (32% the size). As a test, we could then send that binary data file as the payload of a POST to the Cloud Function, as shown below using Postman. Postman will serialize the binary data file’s contents to a binary string before transmitting.

Similarly, we can serialize the same binary Protocol Buffers-based SensorDHT message to a binary string using the SerializeToString method.

message = sensorDHT.SerializeToString()
print(message)

The resulting binary string resembles the following.

b'\n\nrp829c7e0e\x12\x05DHT22\x19c\xee\xbcg\xf5\x8e\xccA"\x12\t\x00\x00\x00\xa0\x99\x191@\x11\x00\x00\x00`f\x06Q@'

The binary string length of the serialized message, and therefore the request payload sent by Postman and received by the Cloud Function for this particular message, is 111 bytes, as compared to the JSON payload size of 148 bytes (75% the size).

Validate Protobuf Payload

To validate the data contained in the Protobuf payload is identical to the JSON payload, we can parse the payload from the serialized binary string using the Protobuf ParseFromString method. We then convert it to JSON using the Protobuf MessageToJson method.

message = sensorDHT.SerializeToString() 
message_parsed = sensors_pb2.SensorDHT()
message_parsed.ParseFromString(message)
print(MessageToJson(message_parsed))

The resulting JSON object is identical to the JSON payload sent using JSON over HTTPS, earlier in the demonstration.

{
  "device": "rp829c7e0e",
  "type": "DHT22",
  "timestamp": 1557585090.476025,
  "data": {
    "temperature": 17.100000381469727,
    "humidity": 68.0999984741211
  }
}

Google Cloud Functions

There are a series of Google Cloud Functions, specifically four HTTP Functions, which accept the sensor data over HTTP from the IoT devices. Each function exposes an HTTPS endpoint. According to Google, you use HTTP functions when you want to invoke your function via an HTTP(S) request. To allow for HTTP semantics, HTTP function signatures accept HTTP-specific arguments.

Below, I have deployed a single function that accepts JSON sensor telemetry from all sensor types, and three functions for Protobuf, one for each sensor type: DHT, PIR, and DLI.

JSON Message Processing

Below, we see the Cloud Function, main.py, which processes the incoming JSON over HTTPS payload from all sensor types. Once the request’s JWT is validated, the JSON message payload is serialized to a byte string and sent to a common Google Cloud Pub/Sub Topic. Note the JWT secret key, id, and password, and the Google Cloud Pub/Sub Topic are all stored as environment variables, local to the Cloud Functions. In my tests, the JSON-based HTTP Functions took an average of 9–18 ms to execute successfully.

import logging
import os

import jwt
from flask import make_response, jsonify
from flask_api import status
from google.cloud import pubsub_v1

TOPIC = os.environ.get('TOPIC')
SECRET_KEY = os.getenv('SECRET_KEY')
ID = os.getenv('ID')
PASSWORD = os.getenv('PASSWORD')


def incoming_message(request):
    if not validate_token(request):
        return make_response(jsonify({'success': False}),
                             status.HTTP_401_UNAUTHORIZED,
                             {'ContentType': 'application/json'})

    request_json = request.get_json()
    if not request_json:
        return make_response(jsonify({'success': False}),
                             status.HTTP_400_BAD_REQUEST,
                             {'ContentType': 'application/json'})

    send_message(request_json)

    return make_response(jsonify({'success': True}),
                         status.HTTP_201_CREATED,
                         {'ContentType': 'application/json'})


def validate_token(request):
    auth_header = request.headers.get('Authorization')
    if not auth_header:
        return False
    auth_token = auth_header.split(" ")[1]

    if not auth_token:
        return False
    try:
        payload = jwt.decode(auth_token, SECRET_KEY)
        if payload['id'] == ID and payload['password'] == PASSWORD:
            return True
    except jwt.ExpiredSignatureError:
        return False
    except jwt.InvalidTokenError:
        return False


def send_message(message):
    publisher = pubsub_v1.PublisherClient()
    publisher.publish(topic=TOPIC, 
                      data=bytes(str(message), 'utf-8'))

The Cloud Functions are deployed to GCP using the gcloud functions deploy CLI command (I use Jenkins to automate the deployments). I have wrapped the deploy commands into bash scripts. The script also copies over a common environment variables YAML file, consumed by the Cloud Function. Each Function has a deployment script, included in the project.

# get latest env vars file
cp -f ./../env_vars_file/env.yaml .

# deploy function
gcloud functions deploy http_json_to_pubsub \
  --runtime python37 \
  --trigger-http \
  --region us-central1 \
  --memory 256 \
  --entry-point incoming_message \
  --env-vars-file env.yaml

Using a .gcloudignore file, the gcloud functions deploy CLI command deploys three files: the cloud function (main.py), required Python packages file (requirements.txt), the environment variables file (env.yaml). Google automatically installs dependencies using the requirements.txt file.

Protobuf Message Processing

Below, we see the Cloud Function, main.py, which processes the incoming Protobuf over HTTPS payload from DHT sensor types. Once the sensor data Protobuf message payload is received by the HTTP Function, it is deserialized to JSON and then serialized to a byte string. The byte string is then sent to a Google Cloud Pub/Sub Topic. In my tests, the Protobuf-based HTTP Functions took an average of 7–14 ms to execute successfully.

As before, note the import sensors_pb2 statement. This statement imports the compiled Protocol Buffers file, which is stored locally to the script on the IoT device. It is used to parse a serialized message into its original Protobuf’s SensorDHT message type.

import logging
import os

import jwt
import sensors_pb2
from flask import make_response, jsonify
from flask_api import status
from google.cloud import pubsub_v1
from google.protobuf.json_format import MessageToJson

TOPIC = os.environ.get('TOPIC')
SECRET_KEY = os.getenv('SECRET_KEY')
ID = os.getenv('ID')
PASSWORD = os.getenv('PASSWORD')


def incoming_message(request):
    if not validate_token(request):
        return make_response(jsonify({'success': False}),
                             status.HTTP_401_UNAUTHORIZED,
                             {'ContentType': 'application/json'})

    data = request.get_data()
    if not data:
        return make_response(jsonify({'success': False}),
                             status.HTTP_400_BAD_REQUEST,
                             {'ContentType': 'application/json'})

    sensor_pb = sensors_pb2.SensorDHT()
    sensor_pb.ParseFromString(data)
    sensor_json = MessageToJson(sensor_pb)
    send_message(sensor_json)

    return make_response(jsonify({'success': True}),
                         status.HTTP_201_CREATED,
                         {'ContentType': 'application/json'})


def validate_token(request):
    auth_header = request.headers.get('Authorization')
    if not auth_header:
        return False
    auth_token = auth_header.split(" ")[1]

    if not auth_token:
        return False
    try:
        payload = jwt.decode(auth_token, SECRET_KEY)
        if payload['id'] == ID and payload['password'] == PASSWORD:
            return True
    except jwt.ExpiredSignatureError:
        return False
    except jwt.InvalidTokenError:
        return False


def send_message(message):
    publisher = pubsub_v1.PublisherClient()
    publisher.publish(topic=TOPIC, data=bytes(message, 'utf-8'))

Cloud Pub/Sub Functions

In addition to HTTP Functions, the demonstration uses a function triggered by Google Cloud Pub/Sub Triggers. According to Google, Cloud Functions can be triggered by messages published to Cloud Pub/Sub Topics in the same GCP project as the function. The function automatically subscribes to the Topic. Below, we see that the function has automatically subscribed to iot-data-demo Cloud Pub/Sub Topic.

Sending Telemetry to MongoDB Atlas

The common Cloud Function, triggered by messages published to Cloud Pub/Sub, then sends the messages to MongoDB Atlas. There is a minimal amount of cleanup required to re-format the Cloud Pub/Sub messages to BSON (binary JSON). Interestingly, according to bsonspec.org, BSON can be compared to binary interchange formats, like Protocol Buffers. BSON is more schema-less than Protocol Buffers, which can give it an advantage in flexibility but also a slight disadvantage in space efficiency (BSON has overhead for field names within the serialized data).

The function uses the PyMongo to connect to MongoDB Atlas. According to their website, PyMongo is a Python distribution containing tools for working with MongoDB and is the recommended way to work with MongoDB from Python.

import base64
import json
import logging
import os
import pymongo

MONGODB_CONN = os.environ.get('MONGODB_CONN')
MONGODB_DB = os.environ.get('MONGODB_DB')
MONGODB_COL = os.environ.get('MONGODB_COL')


def read_message(event, context):
    message = base64.b64decode(event['data']).decode('utf-8')
    message = message.replace("'", '"')
    message = message.replace('True', 'true')
    message = json.loads(message)

    client = pymongo.MongoClient(MONGODB_CONN)
    db = client[MONGODB_DB]
    col = db[MONGODB_COL]
    col.insert_one(message)

The function responds to the published events and sends the messages to the MongoDB Atlas cluster, running in the same Region, us-central1, as the Cloud Functions and Pub/Sub Topic. Below, we see the current options available when provisioning an Atlas cluster.

MongoDB Atlas provides a rich, web-based UI for managing and monitoring MongoDB clusters, databases, collections, security, and performance.

Although Cloud Pub/Sub to Atlas function execution times are longer in duration than the HTTP functions, the latency is greatly reduced by locating the Cloud Pub/Sub Topic, Cloud Functions, and MongoDB Atlas cluster into the same GCP Region. Cross-region execution times were as high as 500-600 ms, while same-region execution times averaged 200-225 ms. Selecting a more performant Atlas cluster would likely result in even lower function execution times.

Aggregating Data with MongoDB Compass

MongoDB Compass is a free, convenient, desktop application for interacting with your MongoDB databases. You can view the collected sensor data, review message (document) schema, manage indexes, and build complex MongoDB aggregations.

When performing analytics or machine learning, I primarily use MongoDB Compass to preview the captured telemetry data and build aggregation pipelines. Aggregation operations process data records and returns computed results. This feature saves a ton of time, filtering and preparing data for further analysis, visualization, and machine learning with Jupyter Notebooks.

Aggregation pipelines can be directly exported to Java, Node, C#, and Python 3. The exported aggregation pipeline code can be placed directly into your Python applications and Jupyter Notebooks.

Below, the exported aggregation pipelines code from MongoDB Compass is used to load a resultset directly into a Pandas DataFrame. This particular aggregation returns time-series DHT sensor data from a specific IoT device over a 72-hour period.

DEVICE_1 = 'rp59adf374'
pipeline = [
    {
        '$match': {
            'type': 'DHT22', 
            'device': DEVICE_1, 
            'timestamp': {
                '$gt': 1557619200,
                '$lt': 1557792000
            }
        }
    }, {
        '$project': {
            '_id': 0,
            'timestamp': 1, 
            'temperature': '$data.temperature', 
            'humidity': '$data.humidity'
        }
    }, {
        '$sort': {
            'timestamp': 1
        }
    }
]
aggResult = iot_data.aggregate(pipeline)
df1 = pd.DataFrame(list(aggResult))

MongoDB Atlas Performance

In this demonstration, from Python3-based Jupyter Notebooks, I was able to consistently query a MongoDB Atlas collection of almost 70k documents for resultsets containing 3 days (72 hours) worth of digital temperature and humidity data, roughly 10.2k documents, in an average of 825 ms. That is round trip from my local development laptop to MongoDB Atlas running on GCP, in a different geographic region.

Query times on GCP are much faster, such as when running a Notebook in JupyterLab on Google’s AI Platform, or a PySpark job with Cloud Dataproc, against Atlas. Running the same Jupyter Notebook directly on Google’s AI Platform, the same MongoDB Atlas query took an average of 450 ms versus 825 ms (1.83x faster). This was across two different GCP Regions; same Region times should be even faster.

GCP Observability

There are several choices for observing the system’s Google Cloud Functions, Google Cloud Pub/Sub, and MongoDB Atlas. As shown above, the GCP Cloud Functions interface lets you see the individual function executions, execution times, memory usage, and active instances, over varying time intervals.

For a more detailed view of Google Cloud Functions and Google Cloud Pub/Sub, I built two custom dashboards using Stackdriver. According to Google, Stackdriver aggregates metrics, logs, and events from infrastructure, giving developers and operators a rich set of observable signals. I built a custom Stackdriver Cloud Functions dashboard (shown below) and a Cloud Pub/Sub Topics and Subscriptions dashboard.

For functions, I chose to display execution times, memory usage, the number of executions, and network egress, all in a single pane of glass, using four graphs. Below, I am using the 95th percentile average for monitoring. The 95th percentile asserts that 95% of the time, the observed values are below this amount and the remaining 5% of the time, the observed values are above that amount.

Data Analysis using Jupyter Notebooks

According to jupyter.org, the Jupyter Notebook is an open-source web application that allows you to create and share documents that contain live code, equations, visualizations, and narrative text. Uses include data cleaning and transformation, numerical simulation, statistical modeling, data visualization, machine learning, and much more. The widespread use of Jupyter Notebooks has grown significantly, as Big Data, AI, and ML have all experienced explosive growth.

PyCharm

JetBrains PyCharm, my favorite Python IDE, has direct integrations with Jupyter Notebooks. In fact, PyCharm’s most recent updates to the Professional Edition greatly enhanced those integrations. PyCharm offers round-trip editing in the IDE and the Jupyter Notebook web browser interface. PyCharm allows you to run and debug individual cells within the notebook. PyCharm automatically starts the Jupyter Server and appropriate kernel for the Notebook you have opened. And, one of my favorite features, PyCharm’s variable viewer tracks the current value of a variable, automatically.

Below, we see the example Analytics Notebook, included in the demonstration’s project, displayed in PyCharm 19.1.2 (Professional Edition). To effectively work with Notebooks in PyCharm really requires a full-size monitor. Working on a laptop with PyCharm’s crowded Notebook UI is workable, but certainly not as effective as on a larger monitor.

Jupyter Notebook Server

Below, we see the same Analytics Notebook, shown above in PyCharm, opened in Jupyter Notebook Server’s web-based client interface, running locally on the development workstation. The web browser-based interface also offers a rich set of features for Notebook development.

From within the Notebook, we are able to query the data from MongoDB Atlas, again using PyMongo, and load the resultsets into Panda DataFrames. As an alternative to hard-coded values and environment variables, with Notebooks, I use the python-dotenv Python package. This package allows me to place my environment variables in a common .env file and reference them from any Notebook. The package has many options for managing environment variables.

We can then analyze the data using a number of common frameworks, including Pandas, Matplotlib, SciPy, PySpark, and NumPy, to name but a few. Below, we see time series data from four different sensors, on the same IoT device. Viewing the data together, we can study the causal effect of one environment variable on another, such as the impact of light on temperature or humidity.

Below, we can use histograms to visualize temperature frequencies for
intervals, over time, for a given device location.

Machine Learning using Jupyter Notebooks

In addition to data analytics, we can use Jupyter Notebooks with tools such as scikit-learn to build machine learning models based on our sensor telemetry. Scikit-learn is a set of machine learning tools in Python, built on NumPy, SciPy, and matplotlib. Below, I have used JupyterLab on Google’s AI Platform and scikit-learn to build several models, based on the sensor data.

Using scikit-learn, we can build models to predict such things as which IoT device generated a specific temperature and humidity reading, or the temperature and humidity, given the time of day, device location, and external environment variables, or discover anomalies in the sensor telemetry.

Scikit-learn makes it easy to construct randomized training and test datasets, to build models, using data from multiple IoT devices, as shown below.

The project includes a Jupyter Notebook that demonstrates how to build several ML models using sensor data. Examples of supervised learning algorithms used to build the classification models in this demonstration include Support Vector Machine (SVM), k-nearest neighbors (k-NN), and Random Forest Classifier.

Having data from multiple sensors, we are able to enrich the ML models by adding additional categorical (discrete) features to our training data. For example, we could look at the effect of light, motion, and time of day on temperature and humidity.

Conclusion

Hopefully, this post has demonstrated how to efficiently collect telemetry data from IoT devices using Google Protocol Buffers over HTTPS, serverless Google Cloud Functions, Cloud Pub/Sub, and MongoDB Atlas, all on the Google Cloud Platform. Once captured, the telemetry data was easily aggregated and analyzed using common tools, such as MongoDB Compass and Jupyter Notebooks. Further, we used the data and tools to build machine learning models for prediction and anomaly detection.

All opinions expressed in this post are my own and not necessarily the views of my current or past employers or their clients.

Image: everythingpossible © 123RF.com

Cloud Function, GCP, Machine Learning, Messaging, ML, MongoDB, MongoDB Atlas, Protobuf, Protocol Buffers, PubSub, Serverless

Using the Google Cloud Dataproc WorkflowTemplates API to Automate Spark and Hadoop Workloads on GCP

Posted by Gary A. Stafford in Bash Scripting, Big Data, Cloud, Continuous Delivery, DevOps, GCP, Java Development, Python, Software Development on December 16, 2018

In the previous post, Big Data Analytics with Java and Python, using Cloud Dataproc, Google’s Fully-Managed Spark and Hadoop Service, we explored Google Cloud Dataproc using the Google Cloud Console as well as the Google Cloud SDK and Cloud Dataproc API. We created clusters, then uploaded and ran Spark and PySpark jobs, then deleted clusters, each as discrete tasks. Although each task could be done via the Dataproc API and therefore automatable, they were independent tasks, without awareness of the previous task’s state.

In this brief follow-up post, we will examine the Cloud Dataproc WorkflowTemplates API to more efficiently and effectively automate Spark and Hadoop workloads. According to Google, the Cloud Dataproc WorkflowTemplates API provides a flexible and easy-to-use mechanism for managing and executing Dataproc workflows. A Workflow Template is a reusable workflow configuration. It defines a graph of jobs with information on where to run those jobs. A Workflow is an operation that runs a Directed Acyclic Graph (DAG) of jobs on a cluster. Shown below, we see one of the Workflows that will be demonstrated in this post, displayed in Spark History Server Web UI.

Here we see a four-stage DAG of one of the three jobs in the workflow, displayed in Spark History Server Web UI.

Workflows are ideal for automating large batches of dynamic Spark and Hadoop jobs, and for long-running and unattended job execution, such as overnight.

Demonstration

Using the Python and Java projects from the previous post, we will first create workflow templates using the just the WorkflowTemplates API. We will create the template, set a managed cluster, add jobs to the template, and instantiate the workflow. Next, we will further optimize and simplify our workflow by using a YAML-based workflow template file. The YAML-based template file eliminates the need to make API calls to set the template’s cluster and add the jobs to the template. Finally, to further enhance the workflow and promote re-use of the template, we will incorporate parameterization. Parameters will allow us to pass parameters (key/value) pairs from the command line to workflow template, and on to the Python script as input arguments.

It is not necessary to use the Google Cloud Console for this post. All steps will be done using Google Cloud SDK shell commands. This means all steps may be automated using CI/CD DevOps tools, like Jenkins and Spinnaker on GKE.

Source Code

All open-sourced code for this post can be found on GitHub within three repositories: dataproc-java-demo, dataproc-python-demo, and dataproc-workflow-templates. Source code samples are displayed as GitHub Gists, which may not display correctly on all mobile and social media browsers.

WorkflowTemplates API

Always start by ensuring you have the latest Google Cloud SDK updates and are working within the correct Google Cloud project.

gcloud components update

export PROJECT_ID=your-project-id 
gcloud config set project $PROJECT

Set the following variables based on your Google environment. The variables will be reused throughout the post for multiple commands.

export REGION=your-region
export ZONE=your-zone
export BUCKET_NAME=your-bucket

The post assumes you still have the Cloud Storage bucket we created in the previous post. In the bucket, you will need the two Kaggle IBRD CSV files, available on Kaggle, the compiled Java JAR file from the dataproc-java-demo project, and a new Python script, international_loans_dataproc.py, from the dataproc-python-demo project.

Use gsutil with the copy (cp) command to upload the four files to your Storage bucket.

gsutil cp data/ibrd-statement-of-loans-*.csv $BUCKET_NAME
gsutil cp build/libs/dataprocJavaDemo-1.0-SNAPSHOT.jar $BUCKET_NAME
gsutil cp international_loans_dataproc.py $BUCKET_NAME

Following Google’s suggested process, we create a workflow template using the workflow-templates create command.

export TEMPLATE_ID=template-demo-1
  
gcloud dataproc workflow-templates create \
  $TEMPLATE_ID --region $REGION

Adding a Cluster

Next, we need to set a cluster for the workflow to use, in order to run the jobs. Cloud Dataproc will create and use a Managed Cluster for your workflow or use an existing cluster. If the workflow uses a managed cluster, it creates the cluster, runs the jobs, and then deletes the cluster when the jobs are finished. This means, for many use cases, there is no need to maintain long-lived clusters, they become just an ephemeral part of the workflow.

We set a managed cluster for our Workflow using the workflow-templates set-managed-cluster command. We will re-use the same cluster specifications we used in the previous post, the Standard, 1 master node and 2 worker nodes, cluster type.

gcloud dataproc workflow-templates set-managed-cluster \
  $TEMPLATE_ID \
  --region $REGION \
  --zone $ZONE \
  --cluster-name three-node-cluster \
  --master-machine-type n1-standard-4 \
  --master-boot-disk-size 500 \
  --worker-machine-type n1-standard-4 \
  --worker-boot-disk-size 500 \
  --num-workers 2 \
  --image-version 1.3-deb9

Alternatively, if we already had an existing cluster, we would use the workflow-templates set-cluster-selector command, to associate that cluster with the workflow template.

gcloud dataproc workflow-templates set-cluster-selector \
  $TEMPLATE_ID \
  --region $REGION \
  --cluster-labels goog-dataproc-cluster-uuid=$CLUSTER_UUID

To get the existing cluster’s UUID label value, you could use a command similar to the following.

CLUSTER_UUID=$(gcloud dataproc clusters describe $CLUSTER_2 \
  --region $REGION \
  | grep 'goog-dataproc-cluster-uuid:' \
  | sed 's/.* //')

echo $CLUSTER_UUID

1c27efd2-f296-466e-b14e-c4263d0d7e19

Adding Jobs

Next, we add the jobs we want to run to the template. Each job is considered a step in the template, each step requires a unique step id. We will add three jobs to the template, two Java-based Spark jobs from the previous post, and a new Python-based PySpark job.

First, we add the two Java-based Spark jobs, using the workflow-templates add-job spark command. This command’s flags are nearly identical to the dataproc jobs submit spark command, used in the previous post.

export STEP_ID=ibrd-small-spark
  
gcloud dataproc workflow-templates add-job spark \
  --region $REGION \
  --step-id $STEP_ID \
  --workflow-template $TEMPLATE_ID \
  --class org.example.dataproc.InternationalLoansAppDataprocSmall \
  --jars $BUCKET_NAME/dataprocJavaDemo-1.0-SNAPSHOT.jar

export STEP_ID=ibrd-large-spark
  
gcloud dataproc workflow-templates add-job spark \
  --region $REGION \
  --step-id $STEP_ID \
  --workflow-template $TEMPLATE_ID \
  --class org.example.dataproc.InternationalLoansAppDataprocLarge \
  --jars $BUCKET_NAME/dataprocJavaDemo-1.0-SNAPSHOT.jar

Next, we add the Python-based PySpark job, international_loans_dataproc.py, as the second job in the template. This Python script requires three input arguments, on lines 15–17, which are the bucket where the data is located and the and results are placed, the name of the data file, and the directory in the bucket where the results will be placed (gist).

We pass the arguments to the Python script as part of the PySpark job, using the workflow-templates add-job pyspark command.

export STEP_ID=ibrd-large-pyspark
  
gcloud dataproc workflow-templates add-job pyspark \
  $BUCKET_NAME/international_loans_dataproc.py \
  --step-id $STEP_ID \
  --workflow-template $TEMPLATE_ID \
  --region $REGION \
  -- $BUCKET_NAME \
     ibrd-statement-of-loans-historical-data.csv \
     ibrd-summary-large-python

That’s it, we have created our first Cloud Dataproc Workflow Template using the Dataproc WorkflowTemplate API. To view our template we can use the following two commands. First, use the workflow-templates list command to display a list of available templates. The list command output displays the version of the workflow template and how many jobs are in the template.

gcloud dataproc workflow-templates list --region $REGION
  
ID               JOBS  UPDATE_TIME               VERSION
template-demo-1  3     2018-12-15T16:32:06.508Z  5

Then, we use the workflow-templates describe command to show the details of a specific template.

gcloud dataproc workflow-templates describe \
  $TEMPLATE_ID --region $REGION

Using the workflow-templates describe command, we should see output similar to the following (gist).

In the template description, notice the template’s id, the managed cluster in the placement section, and the three jobs, all which we added using the above series of workflow-templates commands. Also, notice the creation and update timestamps and version number, which were automatically generated by Dataproc. Lastly, notice the name, which refers to the GCP project and region where this copy of the template is located. Had we used an existing cluster with our workflow, as opposed to a managed cluster, the placement section would have looked as follows.

placement:
  clusterSelector:
    clusterLabels:
      goog-dataproc-cluster-uuid: your_clusters_uuid_label_value

To instantiate the workflow, we use the workflow-templates instantiate command. This command will create the managed cluster, run all the steps (jobs), then delete the cluster. I have added the time command to see how fast the workflow will take to complete.

time gcloud dataproc workflow-templates instantiate \
  $TEMPLATE_ID --region $REGION #--async

We can observe the progress from the Google Cloud Dataproc Console, or from the command line by omitting the --async flag. Below we see the three jobs completed successfully on the managed cluster.

Waiting on operation [projects/dataproc-demo-224523/regions/us-east1/operations/e720bb96-9c87-330e-b1cd-efa4612b3c57].
WorkflowTemplate [template-demo-1] RUNNING
Creating cluster: Operation ID [projects/dataproc-demo-224523/regions/us-east1/operations/e1fe53de-92f2-4f8c-8b3a-fda5e13829b6].
Created cluster: three-node-cluster-ugdo4ygpl52bo.
Job ID ibrd-small-spark-ugdo4ygpl52bo RUNNING
Job ID ibrd-large-spark-ugdo4ygpl52bo RUNNING
Job ID ibrd-large-pyspark-ugdo4ygpl52bo RUNNING
Job ID ibrd-small-spark-ugdo4ygpl52bo COMPLETED
Job ID ibrd-large-spark-ugdo4ygpl52bo COMPLETED
Job ID ibrd-large-pyspark-ugdo4ygpl52bo COMPLETED
Deleting cluster: Operation ID [projects/dataproc-demo-224523/regions/us-east1/operations/f2a40c33-3cdf-47f5-92d6-345463fbd404].
WorkflowTemplate [template-demo-1] DONE
Deleted cluster: three-node-cluster-ugdo4ygpl52bo.

1.02s user 0.35s system 0% cpu 5:03.55 total

In the output, you see the creation of the cluster, the three jobs running and completing successfully, and finally the cluster deletion. The entire workflow took approximately 5 minutes to complete. Below is the view of the workflow’s results from the Dataproc Clusters Console Jobs tab.

Below we see the output from the PySpark job, run as part of the workflow template, shown in the Dataproc Clusters Console Output tab. Notice the three input arguments we passed to the Python script from the workflow template, listed in the output.

We see the arguments passed to the job, from the Jobs Configuration tab.

Examining the Google Cloud Dataproc Jobs Console, we will observe that the WorkflowTemplate API automatically adds a unique alphanumeric extension to both the name of the managed clusters we create, as well as to the name of each job that is run. The extension on the cluster name matches the extension on the jobs ran on that cluster.

YAML-based Workflow Template

Although, the above WorkflowTemplates API-based workflow was certainly more convenient than using the individual Cloud Dataproc API commands. At a minimum, we don’t have to remember to delete our cluster when the jobs are complete, as I often do. To further optimize the workflow, we will introduce YAML-based Workflow Template. According to Google, you can define a workflow template in a YAML file, then instantiate the template to run the workflow. You can also import and export a workflow template YAML file to create and update a Cloud Dataproc workflow template resource.

We can export our first workflow template to create our YAML-based template file.

gcloud dataproc workflow-templates export template-demo-1 \
  --destination template-demo-2.yaml \
  --region $REGION

Below is our first YAML-based template, template-demo-2.yaml. You will need to replace the values in the template with your own values, based on your environment (gist).

Note the template looks almost similar to the template we just created previously using the WorkflowTemplates API. The YAML-based template requires the placement and jobs fields. All the available fields are detailed, here.

To run the template we use the workflow-templates instantiate-from-file command. Again, I will use the time command to measure performance.

time gcloud dataproc workflow-templates instantiate-from-file \
  --file template-demo-2.yaml \
  --region $REGION

Running the workflow-templates instantiate-from-file command will run a workflow, nearly identical to the workflow we ran in the previous example, with a similar timing. Below we see the three jobs completed successfully on the managed cluster, in approximately the same time as the previous workflow.

Waiting on operation [projects/dataproc-demo-224523/regions/us-east1/operations/7ba3c28e-ebfa-32e7-9dd6-d938a1cfe23b].
WorkflowTemplate RUNNING
Creating cluster: Operation ID [projects/dataproc-demo-224523/regions/us-east1/operations/8d05199f-ed36-4787-8a28-ae784c5bc8ae].
Created cluster: three-node-cluster-5k3bdmmvnna2y.
Job ID ibrd-small-spark-5k3bdmmvnna2y RUNNING
Job ID ibrd-large-spark-5k3bdmmvnna2y RUNNING
Job ID ibrd-large-pyspark-5k3bdmmvnna2y RUNNING
Job ID ibrd-small-spark-5k3bdmmvnna2y COMPLETED
Job ID ibrd-large-spark-5k3bdmmvnna2y COMPLETED
Job ID ibrd-large-pyspark-5k3bdmmvnna2y COMPLETED
Deleting cluster: Operation ID [projects/dataproc-demo-224523/regions/us-east1/operations/a436ae82-f171-4b0a-9b36-5e16406c75d5].
WorkflowTemplate DONE
Deleted cluster: three-node-cluster-5k3bdmmvnna2y.

1.16s user 0.44s system 0% cpu 4:48.84 total

Parameterization of Templates

To further optimize the workflow template process for re-use, we have the option of passing parameters to our template. Imagine you now receive new loan snapshot data files every night. Imagine you need to run the same data analysis on the financial transactions of thousands of your customers, nightly. Parameterizing templates makes it more flexible and reusable. By removing hard-codes values, such as Storage bucket paths and data file names, a single template may be re-used for multiple variations of the same job. Parameterization allows you to automate hundreds or thousands of Spark and Hadoop jobs in a workflow or workflows, each with different parameters, programmatically.

To demonstrate the parameterization of a workflow template, we create another YAML-based template with just the Python/PySpark job, template-demo-3.yaml. If you recall from our first example, the Python script, international_loans_dataproc.py, requires three input arguments: the bucket where the data is located and the and results are placed, the name of the data file, and the directory in the bucket, where the results will be placed.

We will replace four of the values in the template with parameters. We will inject those parameter’s values when we instantiate the workflow. Below is the new parameterized template. The template now has a parameters section from lines 26–46. They define parameters that will be used to replace the four values on lines 3–7 (gist).

Note the PySpark job’s three arguments and the location of the Python script have been parameterized. Parameters may include validation. As an example of validation, the template uses regex to validate the format of the Storage bucket path. The regex follows Google’s RE2 regular expression library syntax. If you need help with regex, the Regex Tester – Golang website is a convenient way to test your parameter’s regex validations.

First, we import the new parameterized YAML-based workflow template, using the workflow-templates import command. Then, we instantiate the template using the workflow-templates instantiate command. The workflow-templates instantiate command will run the single PySpark job, analyzing the smaller IBRD data file, and placing the resulting Parquet-format file in a directory within the Storage bucket. We pass the Python script location, bucket link, smaller IBRD data file name, and output directory, as parameters to the template, and therefore indirectly, three of these, as input arguments to the Python script.

export TEMPLATE_ID=template-demo-3

gcloud dataproc workflow-templates import $TEMPLATE_ID \
   --region $REGION --source template-demo-3.yaml
  
gcloud dataproc workflow-templates instantiate \
  $TEMPLATE_ID --region $REGION --async \
  --parameters MAIN_PYTHON_FILE="$BUCKET_NAME/international_loans_dataproc.py",STORAGE_BUCKET=$BUCKET_NAME,IBRD_DATA_FILE="ibrd-statement-of-loans-latest-available-snapshot.csv",RESULTS_DIRECTORY="ibrd-summary-small-python"

Next, we will analyze the larger historic data file, using the same parameterized YAML-based workflow template, but changing two of the four parameters we are passing to the template with the workflow-templates instantiate command. This will run a single PySpark job on the larger IBRD data file and place the resulting Parquet-format file in a different directory within the Storage bucket.

time gcloud dataproc workflow-templates instantiate \
  $TEMPLATE_ID --region $REGION \
  --parameters MAIN_PYTHON_FILE="$BUCKET_NAME/international_loans_dataproc.py",STORAGE_BUCKET=$BUCKET_NAME,IBRD_DATA_FILE="ibrd-statement-of-loans-historical-data.csv",RESULTS_DIRECTORY="ibrd-summary-large-python"

This is the power of parameterization—one workflow template and one job script, but two different datasets and two different results.

Below we see the single PySpark job ran on the managed cluster.

Waiting on operation [projects/dataproc-demo-224523/regions/us-east1/operations/b3c5063f-e3cf-3833-b613-83db12b82f32].
WorkflowTemplate [template-demo-3] RUNNING
Creating cluster: Operation ID [projects/dataproc-demo-224523/regions/us-east1/operations/896b7922-da8e-49a9-bd80-b1ac3fda5105].
Created cluster: three-node-cluster-j6q2al2mkkqck.
Job ID ibrd-pyspark-j6q2al2mkkqck RUNNING
Job ID ibrd-pyspark-j6q2al2mkkqck COMPLETED
Deleting cluster: Operation ID [projects/dataproc-demo-224523/regions/us-east1/operations/fe4a263e-7c6d-466e-a6e2-52292cbbdc9b].
WorkflowTemplate [template-demo-3] DONE
Deleted cluster: three-node-cluster-j6q2al2mkkqck.

0.98s user 0.40s system 0% cpu 4:19.42 total

Using the workflow-templates list command again, should display a list of two workflow templates.

gcloud dataproc workflow-templates list --region $REGION
  
ID               JOBS  UPDATE_TIME               VERSION
template-demo-3  1     2018-12-15T17:04:39.064Z  2
template-demo-1  3     2018-12-15T16:32:06.508Z  5

Looking within the Google Cloud Storage bucket, we should now see four different folders, the results of the workflows.

Job Results and Testing

To check on the status of a job, we use the dataproc jobs wait command. This returns the standard output (stdout) and standard error (stderr) for that specific job.

export SET_ID=ibrd-large-dataset-pyspark-cxzzhr2ro3i54
  
gcloud dataproc jobs wait $SET_ID \
  --project $PROJECT_ID \
  --region $REGION

The dataproc jobs wait command is frequently used for automated testing of jobs, often within a CI/CD pipeline. Assume we have expected part of the job output that indicates success, such as a string, boolean, or numeric value. We could any number of test frameworks or other methods to confirm the existence of that expected value or values. Below is a simple example of using the grep command to check for the existence of the expected line ‘ state: FINISHED’ in the standard output of the dataproc jobs wait command.

command=$(gcloud dataproc jobs wait $SET_ID \
--project $PROJECT_ID \
--region $REGION) &>/dev/null

if grep -Fqx "  state: FINISHED" <<< $command &>/dev/null; then
  echo "Job Success!"
else
  echo "Job Failure?"
fi

# single line alternative
if grep -Fqx "  state: FINISHED" <<< $command &>/dev/null;then echo "Job Success!";else echo "Job Failure?";fi

Job Success!

Individual Operations

To view individual workflow operations, use the operations list and operations describe commands. The operations list command will list all operations.

Notice the three distinct series of operations within each workflow, shown with the operations list command: WORKFLOW, CREATE, and DELETE. In the example below, I’ve separated the operations by workflow, for better clarity.

gcloud dataproc operations list --region $REGION

NAME                                  TIMESTAMP                 TYPE      STATE  ERROR  WARNINGS
fe4a263e-7c6d-466e-a6e2-52292cbbdc9b  2018-12-15T17:11:45.178Z  DELETE    DONE
896b7922-da8e-49a9-bd80-b1ac3fda5105  2018-12-15T17:08:38.322Z  CREATE    DONE
b3c5063f-e3cf-3833-b613-83db12b82f32  2018-12-15T17:08:37.497Z  WORKFLOW  DONE
---
be0e5293-275f-46ad-b1f4-696ba44c222e  2018-12-15T17:07:26.305Z  DELETE    DONE
6784078c-cbe3-4c1e-a56e-217149f555a4  2018-12-15T17:04:40.613Z  CREATE    DONE
fcd8039e-a260-3ab3-ad31-01abc1a524b4  2018-12-15T17:04:40.007Z  WORKFLOW  DONE
---
b4b23ca6-9442-4ffb-8aaf-460bac144dd8  2018-12-15T17:02:16.744Z  DELETE    DONE
89ef9c7c-f3c9-4d01-9091-61ed9e1f085d  2018-12-15T17:01:45.514Z  CREATE    DONE
243fa7c1-502d-3d7a-aaee-b372fe317570  2018-12-15T17:01:44.895Z  WORKFLOW  DONE

We use the results of the operations list command to execute the operations describe command to describe a specific operation.

gcloud dataproc operations describe \
  projects/$PROJECT_ID/regions/$REGION/operations/896b7922-da8e-49a9-bd80-b1ac3fda5105

Each type of operation contains different details. Note the fine-grain of detail we get from Dataproc using the operations describe command for a CREATE operation (gist).

Conclusion

In this brief, follow-up post to the previous post, Big Data Analytics with Java and Python, using Cloud Dataproc, Google’s Fully-Managed Spark and Hadoop Service, we have seen how easy the WorkflowTemplates API and YAML-based workflow templates make automating our analytics jobs. This post only scraped the surface of the complete functionality of the WorkflowTemplates API and parameterization of templates.

In a future post, we leverage the automation capabilities of the Google Cloud Platform, the WorkflowTemplates API, YAML-based workflow templates, and parameterization, to develop a fully-automated DevOps for Big Data workflow, capable of running hundreds of Spark and Hadoop jobs.

All opinions expressed in this post are my own and not necessarily the views of my current or past employers or their clients.

Cloud Dataproc, Dataproc, GCP, Google Cloud Platform, Google Cloud Storage, Hadoop, HDFS, PySpark, Spark, YARN

1 Comment

Big Data Analytics with Java and Python, using Cloud Dataproc, Google’s Fully-Managed Spark and Hadoop Service

Posted by Gary A. Stafford in Big Data, Cloud, GCP, Java Development, Python, Software Development, Technology Consulting on December 11, 2018

However, installing, configuring, and managing the technologies that support big data analytics, data science, ML, and AI, at scale and in Production, often demands an advanced level of familiarity with Linux, distributed systems, cloud- and container-based platforms, databases, and data-streaming applications. The mere ability to manage terabytes and petabytes of transient data is beyond the capability of many enterprises, let alone performing analysis of that data.

To ease the burden of implementing these technologies, the three major cloud providers, AWS, Azure, and Google Cloud, all have multiple Big Data Analytics-, AI-, and ML-as-a-Service offerings. In this post, we will explore one such cloud-based service offering in the field of big data analytics, Google Cloud Dataproc. We will focus on Cloud Dataproc’s ability to quickly and efficiently run Spark jobs written in Java and Python, two widely adopted enterprise programming languages.

Featured Technologies

The following technologies are featured prominently in this post.

Google Cloud Dataproc

dataproc_logo According to Google, Cloud Dataproc is a fast, easy-to-use, fully-managed cloud service for running the Apache Spark and Apache Hadoop ecosystem on Google Cloud Platform. Dataproc is a complete platform for data processing, analytics, and machine learning. Dataproc offers per-second billing, so you only pay for exactly the resources you consume. Dataproc offers frequently updated and native versions of Apache Spark, Hadoop, Pig, and Hive, as well as other related applications. Dataproc has built-in integrations with other Google Cloud Platform (GCP) services, such as Cloud Storage, BigQuery, Bigtable, Stackdriver Logging, and Stackdriver Monitoring. Dataproc’s clusters are configurable and resizable from a three to hundreds of nodes, and each cluster action takes less than 90 seconds on average.

Similar Platform as a Service (PaaS) offerings to Dataproc, include Amazon Elastic MapReduce (EMR), Microsoft Azure HDInsight, and Qubole Data Service. Qubole is offered on AWS, Azure, and Oracle Cloud Infrastructure (Oracle OCI).

According to Google, Cloud Dataproc and Cloud Dataflow, both part of GCP’s Data Analytics/Big Data Product offerings, can both be used for data processing, and there’s overlap in their batch and streaming capabilities. Cloud Dataflow is a fully-managed service for transforming and enriching data in stream and batch modes. Dataflow uses the Apache Beam SDK to provide developers with Java and Python APIs, similar to Spark.

Apache Spark

spark_logo According to Apache, Spark is a unified analytics engine for large-scale data processing, used by well-known, modern enterprises, such as Netflix, Yahoo, and eBay. With in-memory speeds up to 100x faster than Hadoop, Apache Spark achieves high performance for static, batch, and streaming data, using a state-of-the-art DAG (Directed Acyclic Graph) scheduler, a query optimizer, and a physical execution engine.

According to a post by DataFlair, ‘the DAG in Apache Spark is a set of Vertices and Edges, where vertices represent the RDDs and the edges represent the Operation to be applied on RDD (Resilient Distributed Dataset). In Spark DAG, every edge directs from earlier to later in the sequence. On the calling of Action, the created DAG submits to DAG Scheduler which further splits the graph into the stages of the task.’ Below, we see a three-stage DAG visualization, displayed using the Spark History Server Web UI, from a job demonstrated in this post.

Spark’s polyglot programming model allows users to write applications in Scala, Java, Python, R, and SQL. Spark includes libraries for Spark SQL (DataFrames and Datasets), MLlib (Machine Learning), GraphX (Graph Processing), and DStreams (Spark Streaming). Spark may be run using its standalone cluster mode or on Apache Hadoop YARN, Mesos, and Kubernetes.

PySpark

pyspark_logo The Spark Python API, PySpark, exposes the Spark programming model to Python. PySpark is built on top of Spark’s Java API. Data is processed in Python and cached and shuffled in the JVM. According to Apache, Py4J enables Python programs running in a Python interpreter to dynamically access Java objects in a JVM.

Apache Hadoop

hadoop_logo1 According to Apache, the Apache Hadoop project develops open-source software for reliable, scalable, distributed computing. The Apache Hadoop software library is a framework that allows for the distributed processing of large data sets across clusters of computers using simple programming models. This is a rather modest description of such a significant and transformative project. When we talk about Hadoop, often it is in the context of the project’s well-known modules, which includes:

Hadoop Common: The common utilities that support the other Hadoop modules
Hadoop Distributed File System (HDFS): A distributed file system that provides high-throughput access to application data
Hadoop YARN (Yet Another Resource Negotiator): A framework for job scheduling and cluster resource management, also known as ‘Hadoop NextGen’
Hadoop MapReduce: A YARN-based system for parallel processing of large datasets
Hadoop Ozone: An object store for Hadoop

Based on the Powered by Apache Hadoop list, there are many well-known enterprises and academic institutions using Apache Hadoop, including Adobe, eBay, Facebook, Hulu, LinkedIn, and The New York Times.

Spark vs. Hadoop

There are many articles and posts that delve into the Spark versus Hadoop debate, this post is not one of them. Although both are mature technologies, Spark, the new kid on the block, reached version 1.0.0 in May 2014, whereas Hadoop reached version 1.0.0, earlier, in December 2011. According to Google Trends, interest in both technologies has remained relatively high over the last three years. However, interest in Spark, based on the volume of searches, has been steadily outpacing Hadoop for well over a year now. The in-memory speed of Spark over HDFS-based Hadoop and ease of Spark SQL for working with structured data are likely big differentiators for many users coming from a traditional relational database background and users with large or streaming datasets, requiring near real-time processing.

In this post, all examples are built to run on Spark. This is not meant to suggest Spark is necessarily superior or that Spark runs better on Dataproc than Hadoop. In fact, Dataproc’s implementation of Spark relies on Hadoop’s core HDFS and YARN technologies to run.

Demonstration

To show the capabilities of Cloud Dataproc, we will create both a single-node Dataproc cluster and three-node cluster, upload Java- and Python-based analytics jobs and data to Google Cloud Storage, and execute the jobs on the Spark cluster. Finally, we will enable monitoring and notifications for the Dataproc clusters and the jobs running on the clusters with Stackdriver. The post will demonstrate the use of the Google Cloud Console, as well as Google’s Cloud SDK’s command line tools, for all tasks.

In this post, we will be uploading and running individual jobs on the Dataproc Spark cluster, as opposed to using the Cloud Dataproc Workflow Templates. According to Google, Workflow Template is a reusable workflow configuration. It defines a graph of jobs with information on where to run those jobs. Workflow Templates are useful for automating your Datapoc workflows, however, automation is not the primary topic of this post.

Source Code

All open-sourced code for this post can be found on GitHub in two repositories, one for Java with Spark and one for Python with PySpark. Source code samples are displayed as GitHub Gists, which may not display correctly on all mobile and social media browsers.

Cost

Of course, there is a cost associated with provisioning cloud services. However, if you manage the Google Cloud Dataproc resources prudently, the costs are negligible. Regarding pricing, according to Google, Cloud Dataproc pricing is based on the size of Cloud Dataproc clusters and the duration of time that they run. The size of a cluster is based on the aggregate number of virtual CPUs (vCPUs) across the entire cluster, including the master and worker nodes. The duration of a cluster is the length of time, measured in minutes, between cluster creation and cluster deletion.

Over the course of writing the code for this post, as well as writing the post itself, the entire cost of all the related resources was a minuscule US$7.50. The cost includes creating, running, and deleting more than a dozen Dataproc clusters and uploading and executing approximately 75-100 Spark and PySpark jobs. Given the quick creation time of a cluster, 2 minutes on average or less in this demonstration, there is no reason to leave a cluster running longer than it takes to complete your workloads.

Kaggle Datasets

To explore the features of Dataproc, we will use a publicly-available dataset from Kaggle. Kaggle is a popular open-source resource for datasets used for big-data and ML applications. Their tagline is ‘Kaggle is the place to do data science projects’.

For this demonstration, I chose the IBRD Statement Of Loans Data dataset, from World Bank Financial Open Data, and available on Kaggle. The International Bank for Reconstruction and Development (IBRD) loans are public and publicly guaranteed debt extended by the World Bank Group. IBRD loans are made to, or guaranteed by, countries that are members of IBRD. This dataset contains historical snapshots of the Statement of Loans including the latest available snapshots.

There are two data files available. The ‘Statement of Loans’ latest available snapshots data file contains 8,713 rows of loan data (~3 MB), ideal for development and testing. The ‘Statement of Loans’ historic data file contains approximately 750,000 rows of data (~265 MB). Although not exactly ‘big data’, the historic dataset is large enough to sufficiently explore Dataproc. Both IBRD files have an identical schema with 33 columns of data (gist).

In this demonstration, both the Java and Python jobs will perform the same simple analysis of the larger historic dataset. For the analysis, we will ascertain the top 25 historic IBRD borrower, we will determine their total loan disbursements, current loan obligations, and the average interest rates they were charged for all loans. This simple analysis will be performed using Spark’s SQL capabilities. The results of the analysis, a Spark DataFrame containing 25 rows, will be saved as a CSV-format data file.

SELECT country, country_code,
       Format_number(total_disbursement, 0) AS total_disbursement,
       Format_number(total_obligation, 0) AS total_obligation,
       Format_number(avg_interest_rate, 2) AS avg_interest_rate
FROM   (SELECT country,
               country_code,
               Sum(disbursed) AS total_disbursement,
               Sum(obligation) AS total_obligation,
               Avg(interest_rate) AS avg_interest_rate
        FROM   loans
        GROUP  BY country, country_code
        ORDER  BY total_disbursement DESC
        LIMIT  25)

Google Cloud Storage

First, we need a location to store our Spark jobs, data files, and results, which will be accessible to Dataproc. Although there are a number of choices, the simplest and most convenient location for Dataproc is a Google Cloud Storage bucket. According to Google, Cloud Storage offers the highest level of availability and performance within a single region and is ideal for compute, analytics, and ML workloads in a particular region. Cloud Storage buckets are nearly identical to Amazon Simple Storage Service (Amazon S3), their object storage service.

Using the Google Cloud Console, Google’s Web Admin UI, create a new, uniquely named Cloud Storage bucket. Our Dataproc clusters will eventually be created in a single regional location. We need to ensure our new bucket is created in the same regional location as the clusters; I chose us-east1.

We will need the new bucket’s link, to use within the Java and Python code as well from the command line with gsutil. The gsutil tool is a Python application that lets you access Cloud Storage from the command line. The bucket’s link may be found on the Storage Browser Console’s Overview tab. A bucket’s link is always in the format, gs://bucket-name.

Alternatively, we may also create the Cloud Storage bucket using gsutil with the make buckets (mb) command, as follows:

# Always best practice since features are updated frequently
gcloud components update
  
export PROJECT=your_project_name
export REGION=us-east1
export BUCKET_NAME=gs://your_bucket_name
  
# Make sure you are creating resources in the correct project
gcloud config set project $PROJECT
  
gsutil mb -p $PROJECT -c regional -l $REGION $BUCKET_NAME

Cloud Dataproc Cluster

Next, we will create two different Cloud Dataproc clusters for demonstration purposes. If you have not used Cloud Dataproc previously in your GCP Project, you will first need to enable the API for Cloud Dataproc.

Single Node Cluster

We will start with a single node cluster with no worker nodes, suitable for development and testing Spark and Hadoop jobs, using small datasets. Create a single-node Dataproc cluster using the Single Node Cluster mode option. Create the cluster in the same region as the new Cloud Storage bucket. This will allow the Dataproc cluster access to the bucket without additional security or IAM configuration. I used the n1-standard-1 machine type, with 1 vCPU and 3.75 GB of memory. Observe the resources assigned to Hadoop YARN for Spark job scheduling and cluster resource management.

The new cluster, consisting of a single node and no worker nodes, should be ready for use in a few minutes or less.

Note the Image version, 1.3.16-deb9. According to Google, Dataproc uses image versions to bundle operating system, big data components, and Google Cloud Platform connectors into one package that is deployed on a cluster. This image, released in November 2018, is the latest available version at the time of this post. The image contains:

Apache Spark 2.3.1
Apache Hadoop 2.9.0
Apache Pig 0.17.0
Apache Hive 2.3.2
Apache Tez 0.9.0
Cloud Storage connector 1.9.9-hadoop2
Scala 2.11.8
Python 2.7

To avoid lots of troubleshooting, make sure your code is compatible with the image’s versions. It is important to note the image does not contain a version of Python 3. You will need to ensure your Python code is built to run with Python 2.7. Alternatively, use Dataproc’s --initialization-actions flag along with bootstrap and setup shell scripts to install Python 3 on the cluster using pip or conda. Tips for installing Python 3 on Datapoc be found on Stack Overflow and elsewhere on the Internet.

As as an alternative to the Google Cloud Console, we are able to create the cluster using a REST command. Google provides the Google Cloud Console’s equivalent REST request, as shown in the example below.

Additionally, we have the option of using the gcloud command line tool. This tool provides the primary command-line interface to Google Cloud Platform and is part of Google’s Cloud SDK, which also includes the aforementioned gsutil. Here again, Google provides the Google Cloud Console’s equivalent gcloud command. This is a great way to learn to use the command line.

Using the dataproc clusters create command, we are able to create the same cluster as shown above from the command line, as follows:

export PROJECT=your_project_name
export CLUSTER_1=your_single_node_cluster_name 
export REGION=us-east1
export ZONE=us-east1-b
export MACHINE_TYPE_SMALL=n1-standard-1
  
gcloud dataproc clusters create $CLUSTER_1 \
  --region $REGION \
  --zone $ZONE \
  --single-node \
  --master-machine-type $MACHINE_TYPE_SMALL \
  --master-boot-disk-size 500 \
  --image-version 1.3-deb9 \
  --project $PROJECT

There are a few useful commands to inspect your running Dataproc clusters. The dataproc clusters describe command, in particular, provides detailed information about all aspects of the cluster’s configuration and current state.

gcloud dataproc clusters list --region $REGION

gcloud dataproc clusters describe $CLUSTER_2 \
  --region $REGION --format json

Standard Cluster

In addition to the single node cluster, we will create a second three-node Dataproc cluster. We will compare the speed of a single-node cluster to that of a true cluster with multiple worker nodes. Create a new Dataproc cluster using the Standard Cluster mode option. Again, make sure to create the cluster in the same region as the new Storage bucket.

The second cluster contains a single master node and two worker nodes. All three nodes use the n1-standard-4 machine type, with 4 vCPU and 15 GB of memory. Although still considered a minimally-sized cluster, this cluster represents a significant increase in compute power over the first single-node cluster, which had a total of 2 vCPU, 3.75 GB of memory, and no worker nodes on which to distribute processing. Between the two workers in the second cluster, we have 8 vCPU and 30 GB of memory for computation.

Again, we have the option of using the gcloud command line tool to create the cluster:

export PROJECT=your_project_name
export CLUSTER_2=your_three_node_cluster_name 
export REGION=us-east1
export ZONE=us-east1-b
export NUM_WORKERS=2
export MACHINE_TYPE_LARGE=n1-standard-4
  
gcloud dataproc clusters create $CLUSTER_2 \
  --region $REGION \
  --zone $ZONE \
  --master-machine-type $MACHINE_TYPE_LARGE \
  --master-boot-disk-size 500 \
  --num-workers $NUM_WORKERS \
  --worker-machine-type $MACHINE_TYPE_LARGE \
  --worker-boot-disk-size 500 \
  --image-version 1.3-deb9 \
  --project $PROJECT

Cluster Creation Speed: Cloud Dataproc versus Amazon EMS?

In a series of rather unscientific tests, I found the three-node Dataproc cluster took less than two minutes on average to be created. Compare that time to a similar three-node cluster built with Amazon’s EMR service using their general purpose m4.4xlarge Amazon EC2 instance type. In a similar series of tests, I found the EMR cluster took seven minutes on average to be created. The EMR cluster took 3.5 times longer to create than the comparable Dataproc cluster. Again, although not a totally accurate comparison, since both services offer different features, it gives you a sense of the speed of Dataproc as compared to Amazon EMR.

Staging Buckets

According to Google, when you create a cluster, Cloud Dataproc creates a Cloud Storage staging bucket in your project or reuses an existing Cloud Dataproc-created bucket from a previous cluster creation request. Staging buckets are used to stage miscellaneous configuration and control files that are needed by your cluster. Below, we see the staging buckets created for the two Dataproc clusters.

Project Files

Before uploading the jobs and running them on the Cloud Dataproc clusters, we need to understand what is included in the two GitHub projects. If you recall from the Kaggle section of the post, both projects are basically the same but, written in different languages, Java and Python. The jobs they contain all perform the same basic analysis on the dataset.

Java Project

The dataproc-java-demo Java-based GitHub project contains three classes, each which are jobs to run by Spark. The InternationalLoansApp Java class is only intended to be run locally with the smaller 8.7K rows of data in the snapshot CSV file (gist).

On line 20, the Spark Session’s Master URL, .master("local[*]"), directs Spark to run locally with as many worker threads as logical cores on the machine. There are several options for setting the Master URL, detailed here.

On line 30, the path to the data file, and on line 84, the output path for the data file, is a local relative file path.

On lines 38–42, we do a bit of clean up on the column names, for only those columns we are interested in for the analysis. Be warned, the column names of the IBRD data are less than ideal for SQL-based analysis, containing mixed-cased characters, word spaces, and brackets.

On line 79, we call Spark DataFrame’s repartition method, dfDisbursement.repartition(1). The repartition method allows us to recombine the results of our analysis and output a single CSV file to the bucket. Ordinarily, Spark splits the data into partitions and executes computations on the partitions in parallel. Each partition’s data is written to separate CSV files when a DataFrame is written back to the bucket.

Using coalesce(1) or repartition(1) to recombine the resulting 25-Row DataFrame on a single node is okay for the sake of this demonstration, but is not practical for recombining partitions from larger DataFrames. There are more efficient and less costly ways to manage the results of computations, depending on the intended use of the resulting data.

The InternationalLoansAppDataprocSmall class is intended to be run on the Dataproc clusters, analyzing the same smaller CSV data file. The InternationalLoansAppDataprocLarge class is also intended to be run on the Dataproc clusters, however, it analyzes the larger 750K rows of data in the IRBD historic CSV file (gist).

On line 20, note the Spark Session’s Master URL, .master(yarn), directs Spark to connect to a YARN cluster in client or cluster mode depending on the value of --deploy-mode when submitting the job. The cluster location will be found based on the HADOOP_CONF_DIR or YARN_CONF_DIR variable. Recall, the Dataproc cluster runs Spark on YARN.

Also, note on line 30, the path to the data file, and on line 63, the output path for the data file, is to the Cloud Storage bucket we created earlier (.load("gs://your-bucket-name/your-data-file.csv"). Cloud Dataproc clusters automatically install the Cloud Storage connector. According to Google, there are a number of benefits to choosing Cloud Storage over traditional HDFS including data persistence, reliability, and performance.

These are the only two differences between the local version of the Spark job and the version of the Spark job intended for Dataproc. To build the project’s JAR file, which you will later upload to the Cloud Storage bucket, compile the Java project using the gradle build command from the root of the project. For convenience, the JAR file is also included in the GitHub repository.

Python Project

The dataproc-python-demo Python-based GitHub project contains two Python scripts to be run using PySpark for this post. The international_loans_local.py Python script is only intended to be run locally with the smaller 8.7K rows of data in the snapshot CSV file. It does a few different analysis with the smaller dataset. (gist).

Identical to the corresponding Java class, note on line 12, the Spark Session’s Master URL, .master("local[*]"), directs Spark to run locally with as many worker threads as logical cores on the machine.

Also identical to the corresponding Java class, note on line 26, the path to the data file, and on line 66, the output path for the resulting data file, is a local relative file path.

The international_loans_dataproc-large.py Python script is intended to be run on the Dataproc clusters, analyzing the larger 750K rows of data in the IRBD historic CSV file (gist).

On line 12, note the Spark Session’s Master URL, .master(yarn), directs Spark to connect to a YARN cluster.

Again, note on line 26, the path to the data file, and on line 59, the output path for the data file, is to the Cloud Storage bucket we created earlier (.load("gs://your-bucket-name/your-data-file.csv").

These are the only two differences between the local version of the PySpark job and the version of the PySpark job intended for Dataproc. With Python, there is no pre-compilation necessary. We will upload the second script, directly.

Uploading Job Resources to Cloud Storage

In total, we need to upload four items to the new Cloud Storage bucket we created previously. The items include the two Kaggle IBRD CSV files, the compiled Java JAR file from the dataproc-java-demo project, and the Python script from the dataproc-python-demo project. Using the Google Cloud Console, upload the four files to the new Google Storage bucket, as shown below. Make sure you unzip the two Kaggle IRBD CSV data files before uploading.

Like before, we also have the option of using gsutil with the copy (cp) command to upload the four files. The cp command accepts wildcards, as shown below.

export PROJECT=your_project_name
export BUCKET_NAME=gs://your_bucket_name
  
gsutil cp data/ibrd-statement-of-loans-*.csv $BUCKET_NAME
gsutil cp build/libs/dataprocJavaDemo-1.0-SNAPSHOT.jar $BUCKET_NAME
gsutil cp international_loans_dataproc_large.py $BUCKET_NAME

If our Java or Python jobs were larger, or more complex and required multiple files to run, we could also choose to upload ZIP or other common compression formatted archives using the --archives flag.

Running Jobs on Dataproc

The easiest way to run a job on the Dataproc cluster is by submitting a job through the Dataproc Jobs UI, part of the Google Cloud Console.

Dataproc has the capability of running multiple types of jobs, including:

Hadoop
Spark
SparkR
PySpark
Hive
SparkSql
Pig

We will be running both Spark and PySpark jobs as part of this demonstration.

Spark Jobs

To run a Spark job using the JAR file, select Job type Spark. The Region will match your Dataproc cluster and bucket locations, us-east-1 in my case. You should have a choice of both clusters in your chosen region. Run both jobs at least twice, once on both clusters, for a total of four jobs.

Lastly, you will need to input the main class and the path to the JAR file. The JAR location will be:

gs://your_bucket_name/dataprocJavaDemo-1.0-SNAPSHOT.jar

The main class for the smaller dataset will be:

org.example.dataproc.InternationalLoansAppDataprocSmall

The main class for the larger dataset will be:

org.example.dataproc.InternationalLoansAppDataprocLarge

During or after job execution, you may view details in the Output tab of the Dataproc Jobs console.

Like every other step in this demonstration, we can also use the gcloud command line tool, instead of the web console, to submit our Spark jobs to each cluster. Here, I am submitting the larger dataset Spark job to the three-node cluster.

export CLUSTER_2=your_three_node_cluster_name
export REGION=us-east1
export BUCKET_NAME=gs://your_bucket_name
  
gcloud dataproc jobs submit spark \
  --region $REGION \
  --cluster $CLUSTER_2 \
  --class org.example.dataproc.InternationalLoansAppDataprocLarge \
  --jars $BUCKET_NAME/dataprocJavaDemo-1.0-SNAPSHOT.jar \
  --async

PySpark Jobs

To run a Spark job using the Python script, select Job type PySpark. The Region will match your Dataproc cluster and bucket locations, us-east-1 in my case. You should have a choice of both clusters. Run the job at least twice, once on both clusters.

Lastly, you will need to input the main Python file path. There is only one Dataproc Python script, which analyzes the larger dataset. The script location will be:

gs://your_bucket_name/international_loans_dataproc_large.py

Like every other step in this demonstration, we can also use the gcloud command line tool instead of the web console to submit our PySpark jobs to each cluster. Below, I am submitting the PySpark job to the three-node cluster.

export CLUSTER_2=your_three_node_cluster_name
export REGION=us-east1
export BUCKET_NAME=gs://your_bucket_name
  
gcloud dataproc jobs submit pyspark \
  $BUCKET_NAME/international_loans_dataproc_large.py \
  --region $REGION \
  --cluster $CLUSTER_2 \
  --async

Including the optional --async flag with any of the dataproc jobs submit command, the job will be sent to the Dataproc cluster and immediately release the terminal back to the user. If you do not to use the --async flag, the terminal will be unavailable until the job is finished.

However, without the flag, we will get the standard output (stdout) and standard error (stderr) from Dataproc. The output includes some useful information, including different stages of the job execution lifecycle and execution times.

File Output

During development and testing, outputting results to the console is useful. However, in Production, the output from jobs is most often written to Apache Parquet, Apache Avro, CSV, JSON, or XML format files, persisted Apache Hive, SQL, or NoSQL database, or streamed to another system for post-processing, using technologies such as Apache Kafka.

Once both the Java and Python jobs have run successfully on the Dataproc cluster, you should observe the results have been saved back to the Storage bucket. Each script saves its results to a single CSV file in separate directories, as shown below.

The final dataset, written to the CSV file, contains the results of the analysis results (gist).

Cleaning Up

When you are finished, make sure to delete your running clusters. This may be done through the Google Cloud Console. Deletion of the three-node cluster took, on average, slightly more than one minute.

As usual, we can also use the gcloud command line tool instead of the web console to delete the Dataproc clusters.

export CLUSTER_1=your_single_node_cluster_name
export CLUSTER_2=your_three_node_cluster_name 
export REGION=us-east1
  
yes | gcloud dataproc clusters delete $CLUSTER_1 --region $REGION
yes | gcloud dataproc clusters delete $CLUSTER_2 --region $REGION

Results

Some observations, based on approximately 75 successful jobs. First, both the Python job and the Java jobs ran in nearly the same amount of time on the single-node cluster and then on the three-node cluster. This is beneficial since, although, a lot of big data analysis is performed with Python, Java is still the lingua franca of many large enterprises.

Consecutive Execution

Below are the average times for running the larger dataset on both clusters, in Java, and in Python. The jobs were all run consecutively as opposed to concurrently. The best time was 59 seconds on the three-node cluster compared to the best time of 150 seconds on the single-node cluster, a difference of 256%. Given the differences in the two clusters, this large variation is expected. The average difference between the two clusters for running the large dataset was 254%.

Concurrent Execution

It is important to understand the impact of concurrently running multiple jobs on the same Dataproc cluster. To demonstrate this, both the Java and Python jobs were also run concurrently. In one such test, ten copies of the Python job were run concurrently on the three-node cluster.

Observe that the execution times of the concurrent jobs increase in near-linear time. The first job completes in roughly the same time as the consecutively executed jobs, shown above, but each proceeding job’s execution time increases linearly.

According to Apache, when running on a cluster, each Spark application gets an independent set of executor JVMs that only run tasks and store data for that application. Each application is given a maximum amount of resources it can use and holds onto them for its whole duration. Note no tuning was done to the Dataproc clusters to optimize for concurrent execution.

Really Big Data?

Although there is no exact definition of ‘big data’, 750K rows of data at 265 MB is probably not generally considered big data. Likewise, the three-node cluster used in this demonstration is still pretty diminutive. Lastly, the SQL query was less than complex. To really test the abilities of Dataproc would require a multi-gigabyte or multi-terabyte-sized dataset, divided amongst multiple files, computed on a much beefier cluster with more workers nodes and more computer resources.

Monitoring and Instrumentation

In addition to viewing the results of running and completed jobs, there are a number of additional monitoring resources, including the Hadoop Yarn Resource Manager, HDFS NameNode, and Spark History Server Web UIs, and Google Stackdriver. I will only briefly introduce these resources, and not examine any of these interfaces in detail. You’re welcome to investigate the resources for your own clusters. Apache lists other Spark monitoring and instrumentation resources in their documentation.

To access the Hadoop Yarn Resource Manager, HDFS NameNode, and Spark History Server Web UIs, you must create an SSH tunnel and run Chrome through a proxy. Google Dataproc provides both commands and a link to documentation in the Dataproc Cluster tab, to connect.

Hadoop Yarn Resource Manager Web UI

Once you are connected to the Dataproc cluster, via the SSH tunnel and proxy, the Hadoop Yarn Resource Manager Web UI is accessed on port 8088. The UI allows you to view all aspects of the YARN cluster and the distributed applications running on the YARN system.

HDFS NameNode Web UI

Once you are connected to the Dataproc cluster, via the SSH tunnel and proxy, the HDFS NameNode Web UI may is accessed on port 9870. According to the Hadoop Wiki, the NameNode is the centerpiece of an HDFS file system. It keeps the directory tree of all files in the file system, and tracks were across the cluster the file data is kept. It does not store the data of these files itself.

Spark History Server Web UI

We can view the details of all completed jobs using the Spark History Server Web UI. Once you are connected to the cluster, via the SSH tunnel and proxy, the Spark History Server Web UI is accessed on port 18080. Of all the methods of reviewing aspects of a completed Spark job, the History Server provides the most detailed.

Using the History Server UI, we can drill into fine-grained details of each job, including the event timeline.

Also, using the History Server UI, we can see a visualization of the Spark job’s DAG (Directed Acyclic Graph). DataBricks published an excellent post on learning how to interpret the information visualized in the Spark UI.

Not only can view the DAG and drill into each Stage of the DAG, from the UI.

Stackdriver

We can also enable Google Stackdriver for monitoring and management of services, containers, applications, and infrastructure. Stackdriver offers an impressive array of services, including debugging, error reporting, monitoring, alerting, tracing, logging, and dashboards, to mention only a few Stackdriver features.

There are dozens of metrics available, which collectively, reflect the health of the Dataproc clusters. Below we see the states of one such metric, the YARN virtual cores (vcores). A YARN vcore, introduced in Hadoop 2.4, is a usage share of a host CPU. The number of YARN virtual cores is equivalent to the number of worker nodes (2) times the number of vCPUs per node (4), for a total of eight YARN virtual cores. Below, we see that at one point in time, 5 of the 8 vcores have been allocated, with 2 more available.

Next, we see the states of the YARN memory size. YARN memory size is calculated as the number of worker nodes (2) times the amount of memory on each node (15 GB) times the fraction given to YARN (0.8), for a total of 24 GB (2 x 15 GB x 0.8). Below, we see that at one point in time, 20 GB of RAM is allocated with 4 GB available. At that instant in time, the workload does not appear to be exhausting the cluster’s memory.

Notifications

Since no one actually watches dashboards all day, waiting for something to fail, how do know when we have an issue with Dataproc? Stackdrive offers integrations with most popular notification channels, including email, SMS, Slack, PagerDuty, HipChat, and Webhooks. With Stackdriver, we define a condition which describes when a service is considered unhealthy. When triggered, Stackdriver sends a notification to one or more channels.

Below is a preview of two alert notifications in Slack. I enabled Slack as a notification channel and created an alert which is triggered each time a Dataproc job fails. Whenever a job fails, such as the two examples below, I receive a Slack notification through the Slack Channel defined in Stackdriver.

Slack notifications contain a link, which routes you back to Stackdriver, to an incident which was opened on your behalf, due to the job failure.

For convenience, the incident also includes a pre-filtered link directly to the log entries at the time of the policy violation. Stackdriver logging offers advanced filtering capabilities to quickly find log entries, as shown below.

With Stackdriver, you get monitoring, logging, alerting, notification, and incident management as a service, with minimal cost and upfront configuration. Think about how much time and effort it takes the average enterprise to achieve this level of infrastructure observability on their own, most never do.

Conclusion

In this post, we have seen the ease-of-use, extensive feature-set, out-of-the-box integration ability with other cloud services, low cost, and speed of Google Cloud Dataproc, to run big data analytics workloads. Couple this with the ability of Stackdriver to provide monitoring, logging, alerting, notification, and incident management for Dataproc with minimal up-front configuration. In my opinion, based on these features, Google Cloud Dataproc leads other cloud competitors for fully-managed Spark and Hadoop Cluster management.

In future posts, we will examine the use of Cloud Dataproc Workflow Templates for process automation, the integration capabilities of Dataproc with services such as BigQuery, Bigtable, Cloud Dataflow, and Google Cloud Pub/Sub, and finally, DevOps for Big Data with Dataproc and tools like Spinnaker and Jenkins on GKE.

All opinions expressed in this post are my own and not necessarily the views of my current or past employers, their clients, nor Apache or Google.

Big Data, Cloud, Cloud Dataproc, Data Analytics, Dataproc, GCP, Google, Google Cloud, Google Cloud Platform, Hadoop, PySpark, Spark, Stackdriver

2 Comments

Getting Started with PySpark for Big Data Analytics using Jupyter Notebooks and Jupyter Docker Stacks

Posted by Gary A. Stafford in Big Data, Python, Software Development on November 19, 2018

There is little question, big data analytics, data science, artificial intelligence (AI), and machine learning (ML), a subcategory of AI, have all experienced a tremendous surge in popularity over the last few years. Behind the hype curves and marketing buzz, these technologies are having a significant influence on all aspects of our modern lives. Due to their popularity and potential benefits, academic institutions and commercial enterprises are rushing to train large numbers of Data Scientists and ML and AI Engineers.

Learning popular programming paradigms, such as Python, Scala, R, Apache Hadoop, Apache Spark, and Apache Kafka, requires the use of multiple complex technologies. Installing, configuring, and managing these technologies often demands an advanced level of familiarity with Linux, distributed systems, cloud- and container-based platforms, databases, and data-streaming applications. These barriers may prove a deterrent to Students, Mathematicians, Statisticians, and Data Scientists.

Driven by the explosive growth of these technologies and the need to train individuals, many commercial enterprises are lowering the barriers to entry, making it easier to get started. The three major cloud providers, AWS, Azure, and Google Cloud, all have multiple Big Data-, AI- and ML-as-a-Service offerings.

Similarly, many open-source projects are also lowering the barriers to entry into these technologies. An excellent example of an open-source project working on this challenge is Project Jupyter. Similar to the Spark Notebook and Apache Zeppelin projects, Jupyter Notebooks enables data-driven, interactive, and collaborative data analytics with Julia, Scala, Python, R, and SQL.

This post will demonstrate the creation of a containerized development environment, using Jupyter Docker Stacks. The environment will be suited for learning and developing applications for Apache Spark, using the Python, Scala, and R programming languages. This post is not intended to be a tutorial on Spark, PySpark, or Jupyter Notebooks.

Featured Technologies

The following technologies are featured prominently in this post.

Jupyter Notebooks

According to Project Jupyter, the Jupyter Notebook, formerly known as the IPython Notebook, is an open-source web application that allows users to create and share documents that contain live code, equations, visualizations, and narrative text. Uses include data cleaning and transformation, numerical simulation, statistical modeling, data visualization, machine learning, and much more. The word, Jupyter, is a loose acronym for Julia, Python, and R, but today, the Jupyter supports many programming languages. Interest in Jupyter Notebooks has grown dramatically.

Jupyter Docker Stacks

To enable quick and easy access to Jupyter Notebooks, Project Jupyter has created Jupyter Docker Stacks. The stacks are ready-to-run Docker images containing Jupyter applications, along with accompanying technologies. Currently, eight different Jupyter Docker Stacks focus on a particular area of practice. They include SciPy (Python-based mathematics, science, and engineering), TensorFlow, R Project for statistical computing, Data Science with Julia, and the main subject of this post, PySpark. The stacks also include a rich variety of well-known packages to extend their functionality, such as scikit-learn, pandas, Matplotlib, Bokeh, ipywidgets (interactive HTML widgets), and Facets.

Apache Spark

According to Apache, Spark is a unified analytics engine for large-scale data processing, used by well-known, modern enterprises, such as Netflix, Yahoo, and eBay. With speeds up to 100x faster than Hadoop, Apache Spark achieves high performance for static, batch, and streaming data, using a state-of-the-art DAG (Directed Acyclic Graph) scheduler, a query optimizer, and a physical execution engine.

Spark’s polyglot programming model allows users to write applications quickly in Scala, Java, Python, R, and SQL. Spark includes libraries for Spark SQL (DataFrames and Datasets), MLlib (Machine Learning), GraphX (Graph Processing), and DStreams (Spark Streaming). You can run Spark using its standalone cluster mode, on Amazon EC2, Apache Hadoop YARN, Mesos, or Kubernetes.

PySpark

The Spark Python API, PySpark, exposes the Spark programming model to Python. PySpark is built on top of Spark’s Java API. Data is processed in Python and cached and shuffled in the JVM. According to Apache, Py4J enables Python programs running in a Python interpreter to dynamically access Java objects in a JVM.

Docker

According to Docker, their technology developers and IT the freedom to build, manage and secure business-critical applications without the fear of technology or infrastructure lock-in. Although Kubernetes is now the leading open-source container orchestration platform, Docker is still the predominant underlying container engine technology. For this post, I am using Docker Desktop Community version for MacOS.

Docker Swarm

Current versions of Docker include both a Kubernetes and Swarm orchestrator for deploying and managing containers. We will choose Swarm for this demonstration. According to Docker, Swarm is the cluster management and orchestration features embedded in the Docker Engine are built using swarmkit. Swarmkit is a separate project which implements Docker’s orchestration layer and is used directly within Docker.

PostgreSQL

PostgreSQL is a powerful, open source object-relational database system. According to their website, PostgreSQL comes with many features aimed to help developers build applications, administrators to protect data integrity and build fault-tolerant environments, and help manage data no matter how big or small the dataset.

Demonstration

To show the capabilities of the Jupyter development environment, I will demonstrate a few typical use cases, such as executing Python scripts, submitting PySpark jobs, working with Jupyter Notebooks, and reading and writing data to and from different format files and to a database. We will be using the jupyter/all-spark-notebook Docker Image. This image includes Python, R, and Scala support for Apache Spark, using Apache Toree.

Architecture

As shown below, we will stand-up a Docker stack, consisting of Jupyter All-Spark-Notebook, PostgreSQL 10.5, and Adminer containers. The Docker stack will have local directories bind-mounted into the containers. Files from our GitHub project will be shared with the Jupyter application container through a bind-mounted directory. Our PostgreSQL data will also be persisted through a bind-mounted directory. This allows us to persist data external to the ephemeral containers.

Source Code

All open-sourced code for this post can be found on GitHub. Use the following command to clone the project. The post and project code was updated on 6/7/2019.

git clone \
  --branch master --single-branch --depth 1 --no-tags \
  https://github.com/garystafford/pyspark-setup-demo.git

Source code samples are displayed as GitHub Gists, which may not display correctly on some mobile and social media browsers.

Deploy Docker Stack

To start, create the $HOME/data/postgres directory to store PostgreSQL data files. This directory will be bind mounted into the PostgreSQL container on line 36 of the stack.yml file, $HOME/data/postgres:/var/lib/postgresql/data. The HOME environment variable assumes you are working on Linux or MacOS and is equivalent to HOMEPATH on Windows.

The Jupyter container’s working directory is set on line 10 of the stack.yml file, working_dir:/home/$USER/work. The local bind-mounted working directory is $PWD/work. This path is bind-mounted to the working directory in the Jupyter container, on line 24 of the stack.yml file, $PWD/work:/home/$USER/work. The PWD environment variable assumes you are working on Linux or MacOS (CD on Windows).

By default, the user within the Jupyter container is jovyan. Optionally, I have chosen to override that user with my own local host’s user account, as shown on line 16 of the stack.yml file, NB_USER: $USER. I have used the MacOS host’s USER environment variable value (equivalent to USERNAME on Windows). There are many options for configuring the Jupyter container, detailed here. Several of those options are shown on lines 12-18 of the stack.yml file (gist).

Assuming you have a recent version of Docker installed on your local development machine, and running in swarm mode, standing up the stack is as easy as running the following command from the root directory of the project:

docker stack deploy -c stack.yml pyspark

The Docker stack consists of a new overlay network, pyspark-net, and the three containers. To confirm the stack deployed, you can run the following command:

docker stack ps pyspark --no-trunc

Note the jupyter/all-spark-notebook container is quite large. Depending on your Internet connection, if this is the first time you have pulled this Docker image, the stack may take several minutes to enter a running state.

To access the Jupyter Notebook application, you need to obtain the Jupyter URL and access token (read more here). This information is output in the Jupyter container log, which can be accessed with the following command:

docker logs $(docker ps | grep pyspark_pyspark | awk '{print $NF}')

Using the URL and token shown in the log output, you will be able to access the Jupyter web-based user interface on localhost port 8888. Once there, from Jupyter dashboard landing page, you should see all the files in the project’s work/ directory.

Also shown below, note the types of files you are able to create from the dashboard, including Python 3, R, Scala (using Apache Toree or spylon-kernal), and text. You can also open a Jupyter Terminal or create a new Folder.

Running Python Scripts

Instead of worrying about installing and maintaining the latest version of Python and packages on your own development machine, we can run our Python scripts from the Jupyter container. At the time of this post update, the latest jupyter/all-spark-notebook Docker Image runs Python 3.7.3 and Conda 4.6.14. Let’s start with a simple example of the Jupyter container’s capabilities by running a Python script. I have included a sample Python script, 01_simple_script.py.

Run the script from within the Jupyter container, from a Jupyter Terminal window:

python ./01_simple_script.py

You should observe the following output.

Kaggle Datasets

To explore the features of the Jupyter Notebook container and PySpark, we will use a publically-available dataset from Kaggle. Kaggle is a fantastic open-source resource for datasets used for big-data and ML applications. Their tagline is ‘Kaggle is the place to do data science projects’.

For this demonstration, I chose the ‘Transactions from a Bakery’ dataset from Kaggle.

The dataset contains 21,294 rows, each with four columns of data. Although certainly nowhere near ‘big data’, the dataset is large enough to test out the Jupyter container functionality (gist).

Submitting Spark Jobs

We are not limited to Jupyter Notebooks to interact with Spark, we can also submit scripts directly to Spark from a Jupyter Terminal, or from our IDE. I have included a simple Python script, 02_bakery_dataframes.py. The script loads the Kaggle Bakery dataset from the CSV file into a Spark DataFrame. The script then prints out the top ten rows of data, along with a count of the total number of rows in the DataFrame.

Run the script directly from a Jupyter Terminal window:

python ./02_bakery_dataframes.py

An example of the output of the Spark job is shown below. At the time of this post update (6/7/2019), the latest jupyter/all-spark-notebook:latest Docker Image runs Spark 2.4.3, Scala 2.11.12, and Java 1.8.0_191 using the OpenJDK.

More typically, you would submit the Spark job, using the spark-submit command. Use a Jupyter Terminal window to run the following command:

$SPARK_HOME/bin/spark-submit 02_bakery_dataframes.py

Below, we see the beginning of the output from Spark, using the spark-submit command.

Below, we see the scheduled tasks executing and the output of the print statement, displaying the top 10 rows of bakery data.

Interacting with Databases

Often with Spark, you are loading data from one or more data sources (input). After performing operations and transformations on the data, the data is persisted or conveyed to another system for further processing (output).

To demonstrate the flexibility of the Jupyter Docker Stacks to work with databases, I have added PostgreSQL to the Docker Stack. We can read and write data from the Jupyter container to the PostgreSQL instance, running in a separate container.

To begin, we will run a SQL script, written in Python, to create our database schema and some test data in a new database table. To do so, we will need to install the psycopg2 package into our Jupyter container. You can use the docker exec command from your terminal. Alternatively, as a superuser, your user has administrative access to install Python packages within the Jupyter container using the Jupyter Terminal window. Both pip and conda are available to install packages, see details here.

Run the following command to install psycopg2:

# using pip
docker exec -it \
  $(docker ps | grep pyspark_pyspark | awk '{print $NF}') \
  pip install psycopg2-binary

These packages give Python the ability to interact with the PostgreSQL. The included Python script, 03_load_sql.py, will execute a set of SQL statements, contained in a SQL file, bakery_sample.sql, against the PostgreSQL container instance.

To execute the script, run the following command:

python ./03_load_sql.py

This should result in the following output, if successful.

To confirm the SQL script’s success, I have included Adminer. Adminer (formerly phpMinAdmin) is a full-featured database management tool written in PHP. Adminer natively recognizes PostgreSQL, MySQL, SQLite, and MongoDB, among other database engines.

Adminer should be available on localhost port 8080. The password credentials, shown below, are available in the stack.yml file. The server name, postgres, is the name of the PostgreSQL container. This is the domain name the Jupyter container will use to communicate with the PostgreSQL container.

Connecting to the demo database with Adminer, we should see the bakery_basket table. The table should contain three rows of data, as shown below.

Developing Jupyter NoteBooks

The true power of the Jupyter Docker Stacks containers is Jupyter Notebooks. According to the Jupyter Project, the notebook extends the console-based approach to interactive computing in a qualitatively new direction, providing a web-based application suitable for capturing the whole computation process: developing, documenting, and executing code, as well as communicating the results. Notebook documents contain the inputs and outputs of an interactive session as well as additional text that accompanies the code but is not meant for execution.

To see the power of Jupyter Notebooks, I have written a basic notebook document, 04_pyspark_demo_notebook.ipynb. The document performs some typical PySpark functions, such as loading data from a CSV file and from the PostgreSQL database, performing some basic data analytics with Spark SQL, graphing the data using BokehJS, and finally, saving data back to the database, as well as to the popular Apache Parquet file format. Below we see the notebook document, using the Jupyter Notebook user interface.

PostgreSQL Driver

The only notebook document dependency, not natively part of the Jupyter Image, is the PostgreSQL JDBC driver. The driver, postgresql-42.2.5.jar, is included in the project and referenced in the configuration of the notebook’s Spark Session. The JAR is added to the spark.driver.extraClassPath runtime environment property. This ensures the JAR is available to Spark (written in Scala) when the job is run.

PyCharm

Since the working directory for the project is shared with the container, you can also edit files, including notebook documents, in your favorite IDE, such as JetBrains PyCharm. PyCharm has built-in language support for Jupyter Notebooks, as shown below.

As mentioned earlier, a key feature of Jupyter Notebooks is their ability to save the output from each Cell as part of the notebook document. Below, we see the notebook document on GitHub. The output is saved, as part of the notebook document. Not only can you distribute the notebook document, but you can also preserve and share the output from each cell.

Using Additional Packages

As mentioned in the Introduction, the Jupyter Docker Stacks come ready-to-run, with a rich variety of Python packages to extend their functionality. To demonstrate the use of these packages, I have created a second Jupyter notebook document, 05_pyspark_demo_notebook.ipynb. This notebook document uses SciPy (Python-based mathematics, science, and engineering), NumPy (Python-based scientific computing), and the Plotly Python Graphing Library. While NumPy and SciPy are included on the Jupyter Docker Image, the Notebook used pip to install Plotly. Similar to Bokeh, shown previously, we can combine these libraries to create richly interactive data visualizations. To use Plotly, you will need to sign up for a free account and obtain a username and API key.

Shown below, we use Plotly to construct a bar chart of daily bakery items sold for the year 2017 based on the Kaggle dataset. The chart uses SciPy and NumPy to construct a linear fit (regression) and plot a line of best fit for the bakery data. The chart also uses SciPy’s Savitzky-Golay Filter to plot the second line, illustrating a smoothing of our bakery data.

Plotly also provides Chart Studio Online Chart Maker. Plotly describes Chart Studio as the world’s most sophisticated editor for creating d3.js and WebGL charts. Shown below, we have the ability to enhance, stylize, and share our bakery data visualization using the free version of Chart Studio Cloud.

nbviewer

Notebooks can also be viewed using Jupyter nbviewer, hosted on Rackspace. Below, we see the output of a cell from this project’s notebook document, showing a BokehJS chart, using nbviewer. You can view this project’s actual notebook document, using nbviewer, here.

Monitoring Spark Jobs

The Jupyter Docker container exposes Spark’s monitoring and instrumentation web user interface. We can observe each Spark Job in great detail.

We can review details of each stage of the Spark job, including a visualization of the DAG, which Spark constructs as part of the job execution plan, using the DAG Scheduler.

We can also review the timing of each event, occurring as part of the stages of the Spark job.

We can also use the Spark interface to review and confirm the runtime environment, including versions of Java, Scala, and Spark, as well as packages available on the Java classpath.

Spark Performance

Spark, running on a single node within the Jupyter container, on your development system, is not a substitute for a full Spark cluster, running on bare metal or robust virtualized hardware, with YARN, Mesos, or Kubernetes. In my opinion, you should adjust Docker to support an acceptable performance profile for the stack, running only a modest workload. You are not trying to replace the need to run real jobs on a Production Spark cluster.

We can use the docker stats command to examine the container’s CPU and memory metrics:

docker stats --format "table {{.Name}}\t{{.CPUPerc}}\t{{.MemUsage}}\t{{.MemPerc}}"

Below, we see the stats from the stack’s three containers immediately after being deployed, showing little or no activity. Here, Docker has been allocated 2 CPUs, 3GB of RAM, and 2 GB of swap space available, from the host machine.

Compare the stats above with the same three containers, while the example notebook document is running on Spark. The CPU shows a spike, but memory usage appears to be within acceptable ranges.

Linux top and htop

Another option to examine container performance metrics is with top. We can use the docker exec command to execute the top command within the Jupyter container, sorting processes by CPU usage:

docker exec -it \
  $(docker ps | grep pyspark_pyspark | awk '{print $NF}') \
  top -o %CPU

With top, we can observe the individual performance of each processes running in the Jupyter container.

Lastly, htop, an interactive process viewer for Unix, can be installed into the container and ran with the following set of bash commands, from a Jupyter Terminal window or using docker exec:

docker exec -it \
  $(docker ps | grep pyspark_pyspark | awk '{print $NF}') \
  sh -c "apt-get update && apt-get install htop && htop --sort-key PERCENT_CPU"

With htop, we can observe individual CPU activity. The two CPUs at the top left of the htop window are the two CPUs assigned to Docker. We get insight into the way Docker is using each CPU, as well as other basic performance metrics, like memory and swap.

Assuming your development machine host has them available, it is easy to allocate more compute resources to Docker if required. However, in my opinion, this stack is optimized for development and learning, using reasonably sized datasets for data analysis and ML. It should not be necessary to allocate excessive resources to Docker, possibly starving your host machine own compute capabilities.

Conclusion

In this brief post, we have seen how easy it is to get started learning and developing applications for big data analytics, using Python, Spark, and PySpark, thanks to the Jupyter Docker Stacks. We could use the same stack to learn and develop for machine learning, using Python, Scala, and R. Extending the stack’s capabilities is as simple as swapping out this Jupyter image for another, with a different set of tools, as well as adding additional containers to the stack, such as Apache Kafka or Apache Cassandra.

Search results courtesy GoogleTrends (https://trends.google.com)

All opinions expressed in this post are my own and not necessarily the views of my current or past employers, their clients.

Big Data, Bokeh, Data Analytics, docker, Jupyter, Jupyter Notebook, Kaggle, NumPy, Plotly, PySpark, Python, SciPy, Spark

Programmatic Ponderings

Archive for category Big Data

Getting Started with PySpark for Big Data Analytics using Jupyter Notebooks and Jupyter Docker Stacks

Featured Technologies

Jupyter Notebooks

Jupyter Docker Stacks

Apache Spark

PySpark

Docker

Docker Swarm

PostgreSQL

Demonstration

Architecture

Source Code

Deploy Docker Stack

Running Python Scripts

Kaggle Datasets

Submitting Spark Jobs

Interacting with Databases

Developing Jupyter NoteBooks

PostgreSQL Driver

PyCharm

Using Additional Packages

nbviewer

Monitoring Spark Jobs

Spark Performance

Linux top and htop

Conclusion

Gary Stafford

Recent Posts

Top Posts & Pages

Tag Cloud

Tweets