Proprietary + Confidential

Choosing the
Right Database
on Google Cloud

Data Cloud Summit

Solving for the future.
 Kelly Stirman Mohamed Nazeemudeen
Director, Outbound Product Director, Software Engineering
Management, Databases Macy's
Google Cloud
01

Why databases belong in the cloud

Managed services make sense for databases

Server Maintenance/
Upgrades
Database backups

Rack & Stack OS Patches Software patches Scaling

Power, HVAC, Network OS Installation Database installs High Availability

Monitoring
Databases are moving to cloud, fast

By 2022, 75% of all databases

75% will be deployed or migrated to

a cloud platform
in the cloud
- Gartner, 2020

Gartner, Magic Quadrant for Cloud Database Management Systems, November 23, 2020. Donald Feinberg, et al.
You know the business destination,
Google Cloud provides you a well-defined path

Modernize Transform Tools & Partners

DMS Datastream
Spanner Firestore

Bigtable
02

Modernize your existing database workloads

Unblock your operations teams, and get out of the data center business
Modernize with relational database services

Database Google Cloud Service

Oracle → Bare Metal Solution

SQL Server → Cloud SQL for SQL Server

PostgreSQL Cloud SQL for PostgreSQL

MySQL Cloud SQL for MySQL

Challenges Oracle customers face
in the cloud

Increased costs

Oracle requires 2x the licensing to run workloads in cloud

virtual environments, plus 18%-20% maintenance costs

Incomplete deployment support

Other cloud compute environments do not support Oracle

features like Real Application Clusters and older versions of
Oracle DB
Bare Metal Solution

Compute and storage A fully managed data

components are center, with enterprise
certified and optimized grade security and
to run Oracle workloads reliability

Run all versions of Availability across 11 Migration timelines

the Oracle database, Google Cloud regions, 1 traditionally between 8
packaged apps, and ms away from GCP to 10 weeks
hypervisor of your
choice
 Why Bare Metal Solution

Impact Bare Metal Solution with GCP Other cloud providers*

✕ No support for Real Application Clusters

Flexibility ✓ Continue to run any version and other DB options due to Oracle
license restrictions

✓ Up to 50% reduction ✕ Oracle may require 2x the licenses per

Cost advantages
in license costs vs. other clouds CPU core

Data sovereignty ✓ You control regional deployment ✕ No SAN Storage

✕ No support for Real Application Clusters

✓ High availability with RAC
✕ Some applications use database features
✓ Enterprise grade
not available from managed database
Business continuity deployment platform
services
✓ Works with any application,
✕ Older Oracle releases may not be
all Oracle versions
supported

* Will depend on the nature of your company’s commercial terms with Oracle Corporation.
Not long ago, Cloud SQL had many limitations
Recovery

Security Maintenance Cloud SQL Latest Versions

Cross Region
Replicas

Cloud SQL
Cloud SQL Cloud SQL Cloud SQL for SQL Cloud SQL
Maintenance Cloud SQL
CMEK VPC Control Server MySQL 8.0
Controls MySQL &
PostgreSQL
PITR
Cloud SQL
Cloud SQL
Cloud SQL Maintenance Cloud SQL
Access PostgreSQL 13
Data Residency Downtime PostgreSQL 12
Transparency
(>300s) Open

Cloud SQL User In-place major

Cloud SQL Cloud SQL IAM Postgres -
Defined Deny version
Audit logging authentication pglogical, CDC
Periods upgrades

Cloud SQL for Migrations Unique Differentiation

SQL Server AD
Integration

??? ??? ??? ???

Cloud SQL for your critical relational workloads
Better Recovery

Best Maint. GA
More Security Controls Cloud SQL Latest Versions & New Engines
Cross Region
Replicas
GA GA GA GA GA
Cloud SQL
Cloud SQL Cloud SQL GA Cloud SQL for SQL Cloud SQL
Maintenance Cloud SQL
CMEK VPC Control Server MySQL 8.0
Controls MySQL &
PostgreSQL
PITR
GA GA Cloud SQL GA GA GA
Cloud SQL Reduced
Cloud SQL Cloud SQL
Access Maintenance PostgreSQL 13
Data Residency PostgreSQL 12
Transparency Downtime Open
(<90s)

GA GA GA H2 2022
2021
Cloud SQL User In-place major
Cloud SQL Cloud SQL IAM Postgres -
Defined Deny version
Audit logging authentication pglogical, CDC
Periods upgrades

Public
GA
Preview
Cloud SQL for Easy Migrations Only Cloud SQL
SQL Server AD
Integration
GA GA GA GA
Database BigQuery to
Committed Use
Migration Cloud SQL SQL Insights
Discounts
Service Federation
Modernize with managed NoSQL database services

Database Google Cloud Service

Redis → Memorystore for Redis | Redis Labs

Memcached → Memorystore for Memcached

HBase → Bigtable

MongoDB → MongoDB Atlas

Cassandra → Datastax Astra

Neo4j → Neo4j Aura

InfluxDB → InfluxDB Cloud

Partner solutions integrated into console, billing, and support

Unified support

Unified billing

Unified experience

Find them in the Google Cloud

Marketplace
03

Modernize with Cloud Native

Iterate on your customer experiences more quickly
Questions worth asking

Consideration Questions

Availability ● How much downtime is acceptable for the application (higher availability comes at higher cost)?

● Do you need strong consistency?

Consistency
● Can eventual consistency meet your needs (it tends to provide better price/performance)?

● What are your latency requirements?

Performance
● Are they different for reads, writes, or batch operations?

● How many concurrent users for each workload? How will these factors change over time?
Scalability
● How will you manage sharding as your system grows?

● Do you need the rigor of a relational model, or the simplicity and performance of key-value?
Data model
● The combination of flexibility and functionality that the document data model provides?

● Do you need a serverless model with no instance planning or management?

Operations
● Do you have operations teams that need fine-grained controls for your resources?

● Are costs better optimized with a pay per-query model

Cost model
● Or are costs better optimized with a pay per-instance model?

● Where are your customers, so you can deploy “near” them in network time?
Multi-region options
● Which regional combinations make sense for your failover strategy?
Questions worth asking

Consideration Questions

Availability ● How much downtime is acceptable for the application (higher availability comes at higher cost)?

● Do you need strong consistency?

Consistency
● Can eventual consistency meet your needs (it tends to provide better price/performance)?

● What are your latency requirements?

Performance
● Are they different for reads, writes, or batch operations?

● How many concurrent users for each workload? How will these factors change over time?
Scalability
● How will you manage sharding as your system grows?

● Do you need the rigor of a relational model, or the simplicity and performance of key-value?
Data model
● The combination of flexibility and functionality that the document data model provides?

● Do you need a serverless model with no instance planning or management?

Operations
● Do you have operations teams that need fine-grained controls for your resources?

● Are costs better optimized with a pay per-query model

Cost model
● Or are costs better optimized with a pay per-instance model?

● Where are your customers, so you can deploy “near” them in network time?
Multi-region options
● Which regional combinations make sense for your failover strategy?
How cloud native databases improve your agility
Built into Google’s
Cloud Native databases

Automatic shard
99.999% uptime SLA
management

No scheduled maintenance Lock-free schema changes

Spanner Bigtable
Online, unlimited scaling Automatic failure recovery

Zero-touch global
Available in all regions
replication Firestore
When to choose Spanner

You need the strongest consistency guarantees Considerations

You need low-latency local writes

You need RPO and RTO of 0 for zonal and regional failures
across regions

Your schema is unknown or

You need robust transaction controls
highly sparse

You need a drop-in replacement

Your data model is relational
for your legacy database

You need strongly consistent secondary indexes

Superior price+performance vs other sharded relational databases Of course, with cloud native you get

Agility Unlimited scale 99.999% uptime

When to choose Bigtable

You need the predictable low latency provided by global eventual consistency Considerations

You need ACID and

High throughput, predictable, single-digit ms latency for reads and writes
strong consistency

You need flexibility for evolving, or sparse schemas You need secondary indexes

You need drop in replacement for

Large records (up to 256MB)
Cassandra or DynamoDB

Simple migrations from HBase, straightforward from Cassandra, DynamoDB

Superior price+performance vs DynamoDB, Cassandra, HBase Of course, with cloud native you get

Agility Unlimited scale 99.999% uptime

When to choose Firestore

You want to ship robust, scalable applications in days/weeks Considerations

You need real-time data syncing and offline data access You need capacity-based pricing

You want a document database with ACID transactions, external consistency Your data is relational in nature

You want a serverless model to avoid over/under provisioning You need customer-managed
encryption keys

Superior price+performance vs other document databases

Of course, with cloud native you get

Agility Unlimited scale 99.999% uptime

04

How Macy’s Chooses Databases on Google Cloud

Hearing from the customer
Macy’s, Inc. is one of the nation’s premier omnichannel
fashion retailers comprises of 3 iconic brands: Macy’s,
Bloomingdale’s, and Bluemercury.

1858 Founded   90,000 Employees 

727 Stores Top 25 companies for
women in Tech
43 States in U.S.
Top 50 companies
90,000 Employees  for diversity
Common Services – A Strategic Initiative at Macy’s
leverages Google Managed Service Databases
Objectives - GCP
Call
Digital Stores
1. Single source of truth for all clients across Center

all channels - Digital, Stores & Enterprise.

2. Reduce code and build redundancies across teams.
3. Improve speed to market and reduce operational cost. GraphQL – Assembling Item and Product representation from various Services

4. Offer unified customer experience by unlocking

siloed systems.
Catalog Price Inventory Shipping Location
Service Service Service Service Service

SLO Requirements -
1. Hot-hot, Multi-regional with 99.99% availability
2. > 30K TPS throughput with P95 latency as single Spanner BigTable Spanner Datastore Firestore

digit ms.
Price Common Service –
Google BigTable Use case
Macy’s Data Center

Price
Requirements - Authoring
MQ Data Loader

● Need a datastore that is consistently fast

delivering p95 with single digit ms Merchants
latency regardless of the size of dataset. Mainframe DB2

● It should be a fully managed service and

highly scalable.
Interconnect
Access Patterns -
● Given a Division, Location and UPC get
the item ticket price. Google Cloud Platform
Pub Sub

● Given a Division, Location, UPC and

Effective Timestamp (past, current,
future +/- 15 days ) get the item ticket Common Price
Data Flow
Meta Data
price. Service Processor

Clients –
Data Insights - Digital,
Stores, Big Table GBQ
● We have millions of UPCs that grows
Call Center
rapidly across 750 locations with 30 days
worth of price data.
Price Common Service –
Cloud BigTable Schema Design
Row Key Price Data (in protobuf) Metadata
Challenge: Anticipated Volume of records - <location>#<upc>
● Multimillion UPCs * 750 Stores * 30 (if timestamp is part of
12#5432#843572043705 [{"effTs":"2021-05-04T00:00:00","endTs":null, ………….
row key) => Billions of records (division is a low cardinality "priceTypeId":1,"ticketPrice":"876.00”}]
remove that even distribution of
● With each row ~400 Bytes, the total size = ~40 TB data)

Design Approach | Learnings:

10#492036388830 [{"effTs":"2021-05-04T00:00:00", ………….
● We have stored price data as a protobuf in a single column to "endTs":null”,ticketPrice":"1378.00”..etc.,},
{"effTs":"2021-02-09T00:00:00",
avoid using too many columns in any row or having timestamp as "endTs":"2021-05-03T23:59:59",
part of the rowkey which increases the storage size significantly. "ticketPrice":"1208.00”..etc.,}]
Although a table can have millions of columns, each cell adds an
overhead to the amount of data that is sent over the network. 7340#190828116523 ………………. ………….

● Deserializing latency cost for .proto is almost negligible while still

retaining the single digit ms.
● Since pricing system also involves heavy batch writes while
processing price adjustment instructions, we’ve isolated the read
and write workloads using appropriate app-profiles.
● The app profile is configured with “Multi-cluster” routing, so
BigTable handles failover automatically.
05

What’s next?
Moving forward with your modernization
Helping you get there safely and efficiently

Database Assessment Partners & Programs for Database Migration Service A+ Customer Experience
(Stratozone, migVisor) Migrations (DMS) Program (ACE)

Assess and prioritize your Partners and programs that Migrate workloads in just a Providing enterprise grade
workloads for migration to accelerate migrations to few clicks with a serverless reliability and improved
Google Cloud Google Cloud experience customer experience
Thank you.