Develop a Multi-Region Web Application

This version of CockroachDB is no longer supported. For more details, see the Release Support Policy.

This page walks you through developing a multi-region application. It is the fourth section of the Develop and Deploy a Multi-Region Web Application tutorial.


CockroachDB versions v21.1 and above support new multi-region capabilities, with different SQL syntax.

For the latest version of the application and database schema built on v21.1 multi-region features, see the movr-flask repository.

For the latest version of the tutorial, see the v21.1 docs.

Before you begin

Before you begin this section, complete the previous section of the tutorial, Set Up a Virtual Environment for Developing Multi-Region Applications.

Project structure

Our application needs to handle requests from clients, namely web browsers. To translate these kinds of requests into database transactions, the application stack consists of the following components:

  • A geo-partitioned database schema that defines the tables and indexes for user, vehicle, and ride data. The database schema is covered on a separate page, Create a Multi-Region Database Schema.
  • A multi-node, geo-distributed CockroachDB cluster, with each node's locality corresponding to cloud provider regions. Database deployment is covered on separate pages. For instructions on setting up a demo cluster, see Set Up a Virtual Environment for Developing Multi-Region Applications. For instructions on setting up a multi-region cluster, see Deploy a Multi-Region Web Application.
  • Python class definitions that map to the tables in our database. For details, see Mappings below.
  • Functions that wrap database transactions. For details, see Transactions below.
  • A backend API that defines the application's connection to the database. For details, see Database connection below.
  • A Flask server that handles requests from client mobile applications and web browsers. For details, see Web application below.
  • HTML files that the Flask server can render into web pages. For details, see User interface below.

In the sections that follow, we go over each of the files and folders in the project, with a focus on the backend and database components. Here is the application's project directory structure:

├── Dockerfile  ## Defines the Docker container build
├── LICENSE  ## A license file for your application
├── Pipfile ## Lists PyPi dependencies for pipenv
├── Pipfile.lock
├──  ## Contains instructions on running the application
├── dbinit.sql  ## Initializes the database, and partitions the tables by region
├── mcingress.yaml ## Defines the multi-cluster ingress K8s object, for multi-region deployment
├── movr
│   ├──
│   ├──  ## Defines classes that map to tables in the movr database
│   ├──  ## Defines the primary backend API
│   └── ## Defines transaction callback functions
├── movr.yaml  ## Defines service and deployment K8s objects for multi-region deployment
├── requirements.txt  ## Lists PyPi dependencies for Docker container
├──  ## Defines a Flask web application to handle requests from clients and render HTML files
├── static  ## Static resources for the web frontend
│   ├── movr-lockup.png
│   └── movr-symbol.png
├── templates  ## HTML templates for the web UI
│   ├── _nav.html
│   ├── home.html
│   ├── layout.html
│   ├── login.html
│   ├── register.html
│   ├── rides.html
│   ├── user.html
│   ├── users.html
│   ├── vehicles-add.html
│   └── vehicles.html
└── web
    ├──   ## Contains Flask configuration settings
    ├──  ## Defines FlaskForm classes
    ├──  ## Defines utility functions for the web application
    └──  ## Contains gunicorn configuration settings

SQLAlchemy with CockroachDB

Object Relational Mappers (ORMs) map classes to tables, class instances to rows, and class methods to transactions on the rows of a table. The sqlalchemy package includes some base classes and methods that you can use to connect to your database's server from a Python application, and then map tables in that database to Python classes.

In our example, we use SQLAlchemy's Declarative extension, which is built on the mapper() and Table data structures. We also use the sqlalchemy-cockroachdb Python package, which defines the CockroachDB SQLAlchemy dialect. The package includes some functions that help you handle transactions in a running CockroachDB cluster.


After completing the Create a Multi-Region Database Schema section, you should be familiar with the movr database and each of the tables in the database (users, vehicles, and rides).

Open movr/, and look at the first 10 lines of the file:

from sqlalchemy.ext.declarative import declarative_base
from sqlalchemy import Column, String, DateTime, Boolean, Interval, ForeignKey, PrimaryKeyConstraint
from sqlalchemy.types import DATE
from sqlalchemy.dialects.postgresql import UUID
import datetime
from import generate_password_hash
from flask_login import UserMixin

Base = declarative_base()

The first data structure that the file imports is declarative_base, a constructor for the declarative base class, built on SQLAlchemy's Declarative extension. All mapped objects inherit from this object. We assign a declarative base object to the Base variable right below the imports.

The file also imports some other standard SQLAlchemy data structures that represent database objects (like columns), data types, and constraints, in addition to a standard Python library to help with default values (i.e., datetime).

Since the application handles user logins and authentication, we also need to import some security libraries from the Flask ecosystem. We'll cover the web development libraries in more detail in the Web application section below.

The User class

Recall that each instance of a table class represents a row in the table, so we name our table classes as if they were individual rows of their parent table, since that's what they'll represent when we construct class objects.

Take a look at the User class definition:

class User(Base, UserMixin):
    Represents rows of the users table.

        Base {DeclaritiveMeta} -- Base class for declarative SQLAlchemy class definitions that produces appropriate `sqlalchemy.schema.Table` objects.
        UserMixin {UserMixin} -- Mixin object that provides default implementations for the methods that Flask-Login expects user objects to have.

        User -- Instance of the User class.
    __tablename__ = 'users'
    id = Column(UUID)
    city = Column(String)
    first_name = Column(String)
    last_name = Column(String)
    email = Column(String)
    username = Column(String, unique=True)
    password_hash = Column(String)
    is_owner = Column(Boolean)
    PrimaryKeyConstraint(city, id)

    def set_password(self, password):
        Hash the password set by the user at registration.
        self.password_hash = generate_password_hash(password)

    def __repr__(self):
        return "<User(city='{0}', id='{1}', name='{2}')>".format(
  ,, self.first_name + ' ' + self.last_name)

The User class has the following attributes:

  • __tablename__, which holds the stored name of the table in the database. SQLAlchemy requires this attribute for all classes that map to tables.
  • All of the other attributes of the User class (id, city, first_name, etc.), stored as Column objects. These attributes represent columns of the users table. The constructor for each Column takes the column data type as its first argument, and then any additional arguments. To help define column objects, SQLAlchemy also includes classes for SQL data types and column constraints. For the columns in this table, we use UUID and String data types.
  • PrimaryKeyConstraint(), which defines a composite primary key with the city and id columns. Note that the partitioned column (city) precedes the id column, per partition querying best practices.
  • The __repr__ function, which defines the string representation of the object.

The Vehicle class

Next, look at the Vehicle class definition:

class Vehicle(Base):
    Represents rows of the vehicles table.

        Base {DeclaritiveMeta} -- Base class for declarative SQLAlchemy class definitions that produces appropriate `sqlalchemy.schema.Table` objects.

        Vehicle -- Instance of the Vehicle class.
    __tablename__ = 'vehicles'
    id = Column(UUID)
    city = Column(String)
    type = Column(String)
    owner_id = Column(UUID, ForeignKey(''))
    date_added = Column(DATE,
    status = Column(String)
    last_location = Column(String)
    color = Column(String)
    brand = Column(String)
    PrimaryKeyConstraint(city, id)

    def __repr__(self):
        return "<Vehicle(city='{0}', id='{1}', type='{2}', status='{3}')>".format(
  ,, self.type, self.status)

Recall that the vehicles table contains more columns and data types than the users table. It also contains a foreign key constraint (on the users table), and a default value. These differences are reflected in the Vehicle class.

The Ride class

Lastly, there's the Ride class:

class Ride(Base):
    Represents rows of the rides table.

        Base {DeclaritiveMeta} -- Base class for declarative SQLAlchemy class definitions that produces appropriate `sqlalchemy.schema.Table` objects.

        Ride -- Instance of the Ride class.
    __tablename__ = 'rides'
    id = Column(UUID)
    city = Column(String, ForeignKey(''))
    rider_id = Column(UUID, ForeignKey(''))
    rider_city = Column(String, ForeignKey(''))
    vehicle_id = Column(UUID, ForeignKey(''))
    start_location = Column(String)
    end_location = Column(String)
    start_time = Column(DateTime)
    end_time = Column(DateTime)
    length = Column(Interval)
    PrimaryKeyConstraint(city, id)

    def __repr__(self):
        return "<Ride(city='{0}', id='{1}', rider_id='{2}', vehicle_id='{3}')>".format(
  ,, self.rider_id, self.vehicle_id)

The rides table has four foreign key constraints, two on the users table and two on the vehicles table.


After you create a class for each table in the database, you can start defining functions that bundle together common SQL operations as atomic transactions.

The SQLAlchemy CockroachDB dialect

The sqlalchemy-cockroachdb Python library handles transactions in SQLAlchemy with the run_transaction() function. This function takes a transactor, which can be an Engine, Connection, or sessionmaker object, and a callback function. It then uses the transactor to connect to the database, and executes the callback as a database transaction. For some examples of run_transaction() usage, see Transaction callback functions below.

run_transaction() abstracts the details of transaction retries away from your application code. Transactions may require retries if they experience deadlock or read/write contention with other concurrent transactions that cannot be resolved without allowing potential serializable anomalies. As a result, a CockroachDB transaction may have to be tried more than once before it can commit. This is part of how we ensure that our transaction ordering guarantees meet the ANSI SERIALIZABLE isolation level.

run_transaction() has the following additional benefits:

  • When passed a sqlalchemy.orm.session.sessionmaker object, it ensures that a new session is created exclusively for use by the callback, which protects you from accidentally reusing objects via any sessions created outside the transaction. Note that a sessionmaker objects is different from a session object, which is not an allowable transactor for run_transaction().
  • By abstracting transaction retry logic away from your application, it keeps your application code portable across different databases.

Because all callback functions are passed to run_transaction(), the Session method calls within those callback functions are written a little differently than the typical SQLAlchemy application. Most importantly, those functions must not change the session and/or transaction state. This is in line with the recommendations of the SQLAlchemy FAQs, which state (with emphasis added by the original author) that

As a general rule, the application should manage the lifecycle of the session externally to functions that deal with specific data. This is a fundamental separation of concerns which keeps data-specific operations agnostic of the context in which they access and manipulate that data.


Keep the lifecycle of the session (and usually the transaction) separate and external.

In keeping with the above recommendations from the official docs, we strongly recommend avoiding any explicit mutations of the transaction state inside the callback passed to run_transaction(). Specifically, we do not make calls to the following functions from inside run_transaction():

  • sqlalchemy.orm.Session.commit() (or other variants of commit()): This is not necessary because run_transaction() handles the savepoint/commit logic for you.
  • sqlalchemy.orm.Session.rollback() (or other variants of rollback()): This is not necessary because run_transaction() handles the commit/rollback logic for you.
  • Session.flush(): This will not work as expected with CockroachDB because CockroachDB does not support nested transactions, which are necessary for Session.flush() to work properly. If the call to Session.flush() encounters an error and aborts, it will try to rollback. This will not be allowed by the currently-executing CockroachDB transaction created by run_transaction(), and will result in an error message like the following: sqlalchemy.orm.exc.DetachedInstanceError: Instance <FooModel at 0x12345678> is not bound to a Session; attribute refresh operation cannot proceed (Background on this error at:

In the example application provided, all calls to run_transaction() are found within the methods of the MovR class (defined in movr/, which represents the connection to the running database. Requests to the web application frontend (defined in, are routed to the MovR class methods.

Transaction callback functions

To separate concerns, we define all callback functions passed to run_transaction() calls in a separate file, movr/ These callback functions wrap Session method calls, like session.query() and session.add(), to perform database operations within a transaction.


We recommend that you use a sessionmaker object, bound to an existing Engine, as the transactor that you pass to run_transaction(). This protects you from accidentally reusing objects via any sessions created outside the transaction. Every time run_transaction() is called, it uses the sessionmaker object to create a new Session object for the callback. If the sessionmaker is bound to an existing Engine, the same database connection can be reused. imports all of the table classes that we defined in movr/, in addition to some standard Python data structures needed to generate correctly-typed row values that the ORM can write to the database.

from movr.models import Vehicle, Ride, User
import datetime
import uuid


A common query that a client might want to run is a read of the rides table. The transaction callback function for this query is defined here as get_rides_txn():

def get_rides_txn(session, rider_id):
    Select the rows of the rides table for a specific user.

        session {.Session} -- The active session for the database connection.
        rider_id {UUID} -- The user's unique ID.

        List -- A list of dictionaries containing ride information.
    rides = session.query(Ride).filter(Ride.rider_id == rider_id).order_by(
    return list(
            lambda ride: {
                'vehicle_id': ride.vehicle_id,
                'start_time': ride.start_time,
                'end_time': ride.end_time,
                'rider_id': ride.rider_id,
                'length': ride.length
            }, rides))

The get_rides_txn() function takes a Session object and a rider_id string as its inputs, and it outputs a list of dictionaries containing the columns-value pairs of a row in the rides table. To retrieve the data from the database bound to a particular Session object, we use Session.query(), a method of the Session class. This method returns a Query object, with methods for filtering and ordering query results.

Note that get_rides_txn() gets all rides for a specific rider across multiple regions. It does not filter the rides by city, and therefore does not limit the query to data stored in a particular region. As discussed in MovR: An Example Multi-Region Use-Case, running queries on data in multiple regions of a multi-region deployment can lead to latency problems. Because this function defines a read operation, it requires fewer trips between replicated nodes than a write operation, but will likely be slower than a read operation on data constrained to a single region.

Another common query would be to read the registered vehicles in a particular city, to see which vehicles are available for riding. Unlike get_rides_txn(), the get_vehicles_txn() function takes the city string as an input.

def get_vehicles_txn(session, city):
    Select the rows of the vehicles table for a specific city.

        session {.Session} -- The active session for the database connection.
        city {String} -- The vehicle's city.

        List -- A list of dictionaries containing vehicle information.
    vehicles = session.query(Vehicle).filter( == city, Vehicle.status != 'removed').all()
    return list(
            lambda vehicle: {
                'owner_id': vehicle.owner_id,
                'type': vehicle.type,
                'last_location': vehicle.last_location + ', ' +,
                'status': vehicle.status,
                'date_added': vehicle.date_added,
                'color': vehicle.color,
                'brand': vehicle.brand
            }, vehicles))

This function filters the query on the city column. Because the data in the vehicles table is partitioned by city, the function only queries a specific partition of data, constrained to a particular region. This limits latency, as the query only needs to travel to database deployments in a single region.


There are two basic types of write operations: creating new rows and updating existing rows. In SQL terminology, these are INSERT/UPSERT statements and UPDATE statements. All transaction callback functions that update existing rows include a session.query() call. All functions adding new rows call session.add(). Some functions do both.

For example, start_ride_txn(), which is called when a user starts a ride, adds a new row to the rides table, and then updates a row in the vehicles table.

def start_ride_txn(session, city, rider_id, rider_city, vehicle_id):
    Insert a new row into the rides table and update a row of the vehicles table.

        session {.Session} -- The active session for the database connection.
        city {String} -- The vehicle's city.
        rider_id {UUID} -- The user's unique ID.
        rider_city {String} -- The city in which the rider is registered.
        vehicle_id {UUID} -- The vehicle's unique ID.
    v = session.query(Vehicle).filter( == city,
                             == vehicle_id).first()
    r = Ride(city=city,
    v.status = "unavailable"

The function takes the city string, rider_id UUID, rider_city string, and vehicle_id UUID as inputs. It queries the vehicles table for all vehicles of a specific ID, and in a specific city. It also creates a Ride object, representing a row of the rides table. To add the ride to the table in the database bound to the Session, the function calls Session.add(). To update a row in the vehicles table, it modifies the object attribute. start_ride_txn() is called by run_transaction(), which commits the transaction to the database.

Be sure to review the other callback functions in movr/ before moving on to the next section.

Now that we've covered the table classes and some transaction functions, we can look at the interface that connects web requests to a running CockroachDB cluster.

Database connection

The MovR class, defined in movr/, handles connections to CockroachDB using SQLAlchemy's Engine class.

Let's start with the beginning of the file:

from movr.transactions import start_ride_txn, end_ride_txn, add_user_txn, add_vehicle_txn, get_users_txn, get_user_txn, get_vehicles_txn, get_rides_txn, remove_user_txn, remove_vehicle_txn
from cockroachdb.sqlalchemy import run_transaction
from sqlalchemy import create_engine
from sqlalchemy.orm import sessionmaker
from sqlalchemy.dialects import registry
registry.register("cockroachdb", "cockroachdb.sqlalchemy.dialect",

class MovR:
    Wraps the database connection. The class methods wrap database transactions.
    def __init__(self, conn_string):
        Establish a connection to the database, creating Engine and Sessionmaker objects.

            conn_string {String} -- CockroachDB connection string.
        self.engine = create_engine(conn_string, convert_unicode=True)
        self.sessionmaker = sessionmaker(bind=self.engine) first imports the transaction callback functions that we defined in movr/ It then imports the run_transaction() function. The MovR class methods call run_transaction() to execute the transaction callback functions as transactions.

The file also imports some sqlalchemy libraries to create instances of the Engine and Session classes, and to register the CockroachDB as a dialect.

When called, the MovR class constructor creates an instance of the Engine class using an input connection string. This Engine object represents the connection to the database, and is used by the sessionmaker() function to create Session objects for each transaction.

Backend API

The MovR class methods function as the "backend API" for the application. Frontend requests get routed to the these methods.

We've already defined the transaction logic in the transaction callback functions, in movr/ We can now wrap calls to the run_transaction() function in the MovR class methods.

For example, look at the start_ride() method, the method to which frontend requests to start a ride are routed:

    def start_ride(self, city, rider_id, rider_city, vehicle_id):
        Wraps a `run_transaction` call that starts a ride.

            city {String} -- The ride's city.
            rider_id {UUID} -- The user's unique ID.
            rider_city {String} -- The user's city.
            vehicle_id {UUID} -- The vehicle's unique ID.
        return run_transaction(
            self.sessionmaker, lambda session: start_ride_txn(
                session, city, rider_id, rider_city, vehicle_id))

This method takes some keyword arguments, and then returns a run_transaction() call. It passes sessionmaker(bind=self.engine) as the first argument to run_transaction(), which creates a new Session object that binds to the Engine instance initialized by the MovR constructor. It also passes city, rider_id, rider_city, and vehicle_id, values passed from the frontend request, as inputs. These are the same keyword arguments, and should be of the same types, as the inputs for the transaction callback function start_ride_txn().

Be sure to review the other functions in movr/ before moving on to the next section.

Web application

The application needs a frontend and a user interface. We use Flask for the web server, routing, and forms. For the web UI, we use some basic, Bootstrapped HTML templates.

Most of the web application components are found in the web and templates folders, and in the file.


We store the Flask configuration settings in a Config class, defined in web/

# This file defines classes for flask configuration
import os

class Config:
    """Flask configuration class.
    DEBUG = os.environ['DEBUG']
    SECRET_KEY = os.environ['SECRET_KEY']
    API_KEY = os.environ['API_KEY']
    DB_URI = os.environ['DB_URI']
    PREFERRED_URL_SCHEME = ('https', 'http')[DEBUG == 'True']

This file imports the os library, which is used to read from environment variables. When debugging a local deployment, these environment variables are set by the .env file that Pipenv reads. In a multi-region deployment, the environment variables are set by the Dockerfile, and by some commands issued during the Kubernetes deployment.

The application sets a global Flask configuration variable (DEBUG), which it checks to determine whether or not to run against a demo cluster, or a real multi-region deployment.

Web forms

Forms make up an important part of most web application frontends. We define these in web/, using some data structures from the flask_wtf and wtforms libraries.

We will not go into much detail about these forms. The important thing to know is that they help handle POST requests in the web UI.

The CredentialForm class, for example, defines the fields of the login form that users interface with to send a login request to the server:

class CredentialForm(FlaskForm):
    """Login form class.
    username = StringField('Username: ', validators=[data_required()])
    password = PasswordField('Password: ', validators=[data_required()])
    submit = SubmitField('Sign In')

Most of the forms defined on this page take the inputs for database reads and writes.

VehicleForm, for example, defines the fields of the vehicle registration form. Users enter information about a vehicle they would like to register, and the data is routed to the add_vehicle() method defined in movr/

class VehicleForm(FlaskForm):
    """Vehicle registration form class.
    type = SelectField(label='Type',
                       choices=[('bike', 'Bike'), ('scooter', 'Scooter'),
                                ('skateboard', 'Skateboard')])
    color = StringField(label='Color', validators=[data_required()])
    brand = StringField(label='Brand')
    location = StringField(label='Current location: ',
    submit = SubmitField('Add vehicle')

Initialization defines the main process of the application: the web server. After initializing the database, we run python to start up the web server.

Let's look at the first ten lines of

from flask import Flask, render_template, session, redirect, flash, url_for, Markup, request, Response
from flask_bootstrap import Bootstrap, WebCDN
from flask_login import LoginManager, current_user, login_user, logout_user, login_required
from import check_password_hash
from movr.movr import MovR
from web.forms import CredentialForm, RegisterForm, VehicleForm, StartRideForm, EndRideForm, RemoveUserForm, RemoveVehicleForm
from web.config import Config
from web.geoutils import get_region
from sqlalchemy.exc import DBAPIError

The first line imports standard Flask libraries for connecting, routing, and rendering web pages. The next three lines import some libraries from the Flask ecosystem, for bootstrapping, authentication, and security.

The next four lines import other web resources that we've defined separately in our project:

  • The MovR class, which handles the connection and interaction with the running CockroachDB cluster.
  • Several FlaskForm superclasses, which define the structure of web forms.
  • The Config class, which holds configuration information for the Flask application.
  • get_region(), which matches a city to a cloud provider region, so the application knows which partition to query.

Finally, we import the DBAPIError type from sqlalchemy, for error handling.

The next five or so lines initialize the application:


# Initialize the app
app = Flask(__name__)
login = LoginManager(app)
protocol = ('https', 'http')[app.config.get('DEBUG') == 'True']

Calling the Flask() constructor initializes our Flask web server. By assigning this to a variable (app), we can then configure the application, store variables in its attributes, and call it with functions from other libraries, to add features and functionality to the application.

For example, we can bootstrap the application for enhanced HTML and form generation with Bootstrap(app). We can also add authentication with LoginManager(app), and then control the default routes, based on authentication, with the DEFAULT_ROUTE_AUTHENTICATED and DEFAULT_ROUTE_AUTHENTICATED constants.

Note that we also define a protocol variable that the application later uses to determine its protocol scheme. You should always use HTTPS for secure connections when building an application that accepts user login information.

After initializing the application, we can connect to the database:

conn_string = app.config.get('DB_URI')
movr = MovR(conn_string)

These two lines connect the application to the database. First the application retrieves the connection string that is stored as a configuration variable of the app. Then it calls the MovR() constructor to establish a connection to the database at the location we provided to the Config object.

User authentication

User authentication is handled with the Flask-Login library. This library manages user logins with the LoginManager, and some other functions that help verify if a user has been authenticated or not.

To control whether certain routes are accessible to a client session, we define a user_loader() function:

def load_user(user_id):
    return movr.get_user(user_id=user_id)

To restrict access to a certain page to users that are logged in with the LoginManager, we add the @login_required decorator function to the route. We'll go over some examples in the Routing section below.


We define all Flask routing functions directly in

Flask applications use @app.route() decorators to handle client requests to specific URLs. When a request is sent to a URL served by the Flask app instance, the server calls the function defined within the decorator (the routing function).

Flask provides a few useful callbacks to use within a routing function definition:

  • redirect(), which redirects a request to a different URL.
  • render_template(), which renders an HTML page, with Jinja2 templating, into a static webpage.
  • flash(), which sends messages from the application output to the webpage.

In addition to calling these functions, and some other standard Flask and Python libraries, the application's routing functions need to call some of the methods that we defined in movr/ (the "backend API").

For example, look at the login() route:

@app.route('/login', methods=['GET', 'POST'])
def login_page():
    if current_user.is_authenticated:
        return redirect(url_for(DEFAULT_ROUTE_AUTHENTICATED, _external=True, _scheme=protocol))
        form = CredentialForm()
        if form.validate_on_submit():
                user = movr.get_user(
                if user is None or not check_password_hash(
                            'Invalid user credentials.<br>If you aren\'t registered with MovR, go <a href="{0}">Sign Up</a>!'
                    return redirect(
                        url_for('login_page', _external=True,
                return redirect(
                    url_for(DEFAULT_ROUTE_AUTHENTICATED, _external=True, _scheme=protocol))
            except Exception as error:
                return redirect(
                    url_for('login_page', _external=True, _scheme=protocol))
        return render_template('login.html',
                               title='Log In',

Client Location

To optimize for latency in a global application, when a user arrives at the website, their request needs to be routed to the application deployment closest to the location from which they made the request. This step is handled outside of the application logic, by a cloud-hosted, global load balancer. In our example multi-region deployment, we use a GCP multi-cluster ingress.

Let's assume that the application has been deployed behind a global load balancer that routes requests to deployments, based on the location of the request. Even if the application request is routed to the correct regional deployment of the application, the application still needs to know where the client is located, so that requests can be translated into database operations on the partitions relevant to the client's location (city, in this case). For example, if a user logs onto your website in New York, you want them to only see the vehicles available in New York, and you want to them to only be able to start rides in New York.

All HTTP requests include standard HTTP headers. Client location information is not included in standard HTTP header fields, but some non-standard HTTP header fields include information that can be used to derive client location, using client-side logic. For example, you can reverse geo-locate an IP address delivered through the X-Forwarded-For field.

Most load-balancing services require you to enable header forwarding, or the original request headers will be lost behind the load balancer. Some services, like GCP Network Services, allow you to define non-standard HTTP headers to include location information directly with user-defined request headers, helping you avoid writing client-side logic for reverse geo-location.

We discuss the details of configuring custom request header fields in our example production deployment. For now, let's assume that you receive the location of your client from a request header field named "X-City".

Suppose that a user is coming to the website for the first time. They type in the address of the website, let's say, into their browser, and hit enter. The following example routing function gathers the client's location, stores it in a persistent Flask session variable, and then renders the home page:

@app.route('/', methods=['GET'])
def index():
    if app.config.get('DEBUG') == 'True':
        session['city'] = 'new york'
        session['region'] = 'gcp-us-east1'
            'Application is in DEBUG mode. Currently using New York as city and gcp-us-east1 as region.')
            place = request.headers.get("X-PLACE").split(",", 1)
            # This header attribute is passed by the HTTP load balancer, to its
            # configured backend. The header must be configured manually in the
            # cloud service provider's console to include this attribute. See
            # README for more details.
            session['city'] = place[0].lower()
            session['latlong'] = place[1]
            session['region'] = get_region(session['city'], session['latlong'])
        except Exception as error:
            session['city'] = 'new york'
            session['region'] = 'gcp-us-east1'
                '{0}\n Unable to retrieve client location information.\n Application is now assuming you are in New York.'
    session['riding'] = None
    if current_user.is_authenticated:
        return redirect(url_for(DEFAULT_ROUTE_AUTHENTICATED, _external=True, _scheme=protocol))
        return redirect(url_for(DEFAULT_ROUTE_NOT_AUTHENTICATED, _external=True, _scheme=protocol))

The function's control flow sequence evaluates the DEBUG configuration variable first. If the application is in DEBUG mode, it simplifies the work that the application needs to do with the request information by setting the city to new york. This variable should be set to True if you are running the application against a demo CockroachDB cluster.

If the application is not in DEBUG mode, then the request defines a new session variable, city, as the value in the request header field X-City. If the process fails for some reason (e.g., the X-City header field wasn't properly forwarded), the application sets city to new york, like the DEBUG case, and prints the error to the rendered webpage, using Flask's flash() function.

In all cases, the session variables city, region, and riding are initialized, and then the home.html web page is rendered, with the region and city variables passed to the HTML template as variables.

User interface

For the example application, we limit the web UI to some static HTML web pages, rendered using Flask's built-in Jinja-2 engine. We will not spend much time covering the web UI. Just note that the forms take input from the user, and that input is usually passed to the backend where it is translated into and executed as a database transaction.

We've also added some Bootstrap syntax and Google Maps, for UX purposes. As you can see, the Google Maps API requires a key. For debugging, you can define this key in the .env file. If you decide to use an embedded service like Google Maps in production, you should restrict your Google Maps API key to a specific hostname or IP address from within the cloud provider's console, as the API key will be publicly visible in the HTML. In Deploying a Multi-Region Web Application, we use Kubernetes secrets to the store the API keys.

Next Steps

After you finish developing and debugging your application, you can start deploying the application.

See also

Yes No

Yes No