Cloud Databases in Research and Practice

transcript

Felix Gessert(29.04.2014)

Slides: baqend.com/nosql.pdf

About me

PhD student (database group university of hamburg)

About me

Research Project for PhD

About me

Research Project for PhD

Cloud Database Startup

Outline

• Categories of Cloud Databases

• Properties

What are Cloud Databases?

Cloud Databases in the wild

Research Perspectives

Wrap-up andliterature

2003: GFS (Google)

2006: BigTable (Google)

2007: Dynamo (Amazon)

2007: S3 (Amazon)

2008: SimpleDB (Amazon)

2010: SQL Azure

2011: Parse

2012: DynamoDB (Amazon)

2013: Redshift (Amazon)

Context: What kinds of Cloud databases are there?

Infrastructure-as-a-Service

Platform-as-a-Service

…Database-as-a-ServiceCloud SQL

Amazon RDS

SQL Azure

Cloudant

MongoHQ

Orestes

Google F1

DynamoDB BigQuery

EMR GCS

Amazon RDS

SQL Azure

Cloudant

MongoHQ

Orestes

Google F1

DynamoDB

ManagedRDBMSs

BigQuery

EMR GCS

Amazon RDS

SQL Azure

Cloudant

MongoHQ

Orestes

Google F1

DynamoDB

ManagedRDBMSs

Cloud-Deploymentof DBMSs

BigQuery

EMR GCS

Amazon RDS

SQL Azure

Cloudant

MongoHQ

Orestes

Google F1

DynamoDB

ManagedNoSQL DBs

ManagedRDBMSs

BigQuery

EMR GCS

Amazon RDS

SQL Azure

Cloudant

MongoHQ

Orestes

Google F1

DynamoDB

ManagedNoSQL DBs

ManagedRDBMSs

Cloud-onlyDBaaS-Systems

BigQuery

EMR GCS

Amazon RDS

SQL Azure

Cloudant

MongoHQ

Orestes

Google F1

DynamoDB

ManagedNoSQL DBs

ManagedRDBMSs

BigQuery

Analytics-as-a-Service

Amazon RDS

SQL Azure

Cloudant

MongoHQ

Orestes

Google F1

DynamoDB

ManagedNoSQL DBs

ManagedRDBMSs

BigQuery

ObjectStores

Amazon RDS

SQL Azure

Cloudant

MongoHQ

Orestes

Google F1

DynamoDB

ManagedNoSQL DBs

ManagedRDBMSs

Storage APIs

BigQuery

ObjectStores

Amazon RDS

SQL Azure

Cloudant

MongoHQ

Orestes

Google F1

DynamoDB

ManagedNoSQL DBs

ManagedRDBMSs

Backend-as-a-Service

Storage APIs

BigQuery

ObjectStores

Cloud-deployeddatabase

Data-Analytics-as-a-Service

Database-as-a-Service

ManagedRDBMS

ManagedDWH

ManagedNoSQL DB

Proprietary or

Polyglot

Files Object Store

Cloud Databases

ManagedRDBMS

ManagedDWH

ManagedNoSQL DB

Proprietary or

Polyglot

Files Object Store

Cloud Databases

Standard Interface

ManagedRDBMS

ManagedDWH

ManagedNoSQL DB

Proprietary or

Polyglot

Files Object Store

Cloud Databases

Standard Interface

Vendor-specific

Interface

Application Architecture <-> Cloud Database Category

Architectures

Applications

Data Warehouse

Operative Database

Reporting Data MiningAnalytics

Architectures

Applications

Data Warehouse

Operative Database

Architectures

Applications

Data Warehouse

Operative Database

DBaaSProbaby not.

Cloud-Deployed Database

IaaS-Cloud

Cloud-deploy yourfavourite database system

IaaS-Cloud

Cloud-deploy yourfavourite database system

Does not solve:Provisioning, Backups, Security, Scaling, Elasticity, Performance Tuning, Failover, Replication, ...

Managed RDBMS/DWH/NoSQL DB

IaaS-Cloud

DBaaS-Provider

IaaS-Cloud

RDBMS DWH NoSQL DB

DBaaS-Provider

Provisioning, Backups, Security, Scaling, Elasticity, Performance Tuning, Failover, Replication, ...

IaaS-Cloud

RDBMS DWH NoSQL DB

DBaaS-Provider

IaaS-Cloud

RDBMS DWH NoSQL DB

DBaaS-Provider

SQL Azure

Google

Cloud SQL

IaaS-Cloud

RDBMS DWH NoSQL DB

DBaaS-Provider

SQL Azure

Google

Cloud SQL

IaaS-Cloud

RDBMS DWH NoSQL DB

DBaaS-Provider

Amazon Redshift

SQL Azure

Google

Cloud SQL

Proprietary DB/Object Store

Black-Box Databaseor file system

Managed byCloud Provider

Provid

er‘sA

Provid

er‘sA

Amazon

SimpleDB

Google Cloud

DatastoreAzure Tables

Database.com

BigTable, Megastore, Spanner, F1, Dynamo,

PNuts, Relational Cloud, …

rietaryD

Provid

er‘sA

Amazon

SimpleDB

Google Cloud

Storage

Azure Blob

Storage

Google Cloud

DatastoreAzure Tables

Openstack

Database.com

BigTable, Megastore, Spanner, F1, Dynamo,

PNuts, Relational Cloud, …

rietaryD

Backend-as-a-Service, Polyglot PersistenceService

IaaS-Cloud

Backend API

Service-Layer

Data API

IaaS-Cloud

Backend API

Service-Layer

Data API

Authentication, Users, Validation,etc.

Maps to (different) databases

IaaS-Cloud

Backend API

Service-Layer

Data API

Realtim

IaaS-Cloud

Backend API

Service-Layer

Data API

Realtim

AppCelerator

Analytics Cluster

Provisioning, Data Ingest

Analytics Cluster

HDInsight

Amazon Elastic

MapReduce

Analytics Cluster

HDInsight

Google

BigQuery

Google

Prediction API

Amazon Elastic

MapReduce

DBaaS: Common Aspects

#1 Metric: Total Cost

Daniela Florescu and Donald Kossmann “Rethinking cost and performance of database systems”, SIGMOD Rec. 2009.

#1 Metric: Total Cost

Maximum utilization ofavailable hardware:

Multi-Tenancy

Daniela Florescu and Donald Kossmann “Rethinking cost and performance of database systems”, SIGMOD Rec. 2009.

Multi-Tenancy - four common approaches:

T. Kiefer, W. Lehner “Private table database virtualization for dbaas” UCC, 2011

Private OS Private Process/DB Private Schema Shared Schema

Private OS

Hardware Resources

Database Process

Database

Schema

Private Process/DB Private Schema Shared Schema

e.g. Amazon RDS

Private OS

Hardware Resources

Database Process

Database

Schema

Private Process/DB Private Schema

Hardware Resources

Database Process

Database

Schema

Shared Schema

e.g. Amazon RDS e.g. MongoHQ

Private OS

Hardware Resources

Database Process

Database

Schema

Hardware Resources

Database Process

Database

Schema

Hardware Resources

Database Process

Database

Schema

Shared Schema

e.g. Amazon RDS e.g. MongoHQ e.g. Google DataStore

Private OS

Hardware Resources

Database Process

Database

Schema

Hardware Resources

Database Process

Database

Schema

Hardware Resources

Database Process

Database

Schema

Shared Schema

Hardware Resources

Database Process

Database

Schema

Virtual Schema

e.g. Amazon RDS e.g. MongoHQ e.g. Google DataStore Most SaaS Apps

W. Lehner, U. Sattler “Web-scale Data Management for the Cloud” Springer, 2013

Private OS

Private Process/DB

Private Schema

Shared Schema

App.indep.

IsolationRessource

Util.Maintenance,Provisioning

Billing Models:

Account

Billing Models:

Account

Pay-per-useParameters: Network, Bandwidth, Storage, CPU, Requests, etc.Payment: Pre-Paid, Post-PaidVariants: On-Demand, Auction, Reserved

e.g. DynamoDB

Billing Models:

Account

End ofmonth

Plan-basedParameters: Allocated Plan (e.g. 2 instances + X GB storage)

e.g. MongoHQ

Billing Models:

Account

End ofmonth

Plan-basedParameters: Allocated Plan (e.g. 2 instances + X GB storage)

Free Tier: free plan or free initial account credit

e.g. MongoHQ

Database-a-

a-Service

Authentication

Authorization

Database-a-

a-Service

Authentication

Authorization

Authenticate

Internal Schemes External IdentityProvider

Federated Identity (SSO)

e.g. Amazon IAM e.g. OpenID e.g. SAML

Database-a-

a-Service

Authentication

Authorization

Authenticate

Used extensively

Database-a-

a-Service

Authentication

Authorization

Authenticate

Used extensively

Database-a-

a-Service

Authentication

Authorization

Authenticate

Used extensively

Database-a-

a-Service

Authentication

Authorization

Authenticate

Authenticated Request

Used extensively

User-based Access Control

Role-based Access Control

Policies

e.g. Amazon S3 ACLs e.g. Amazon IAM e.g. XACML

Database-a-

a-Service

Authentication

Authorization

Authenticate

Used extensively

Policies

Database-a-

a-Service

Authentication

Authorization

Authenticate

Response

Used extensively

Policies

Database-a-

a-Service

Authentication

Authorization

Authenticate

Response

Used extensively

Policies

Database-a-

a-Service

Authentication

Authorization

Authenticate

Response

Federated ACLs

M. Decat, B. Lagaisse, et al. “Toward efficient and confidentiality-aware federation of access control policies “, DOA Trusted Cloud 2013

• Imagine ACL: „Patient data can only beaccessed by treating physician“

• Idea: decompose policy and evaluate parts neardata owner

Service Level Agreements

Legal Part1. Fees2. Penalties

Technical Part1. SLO2. SLO3. SLO

Legal Part1. Fees2. Penalties

Technical Part1. SLO2. SLO3. SLO

Service Level Objectives:• Availability• Durability• Consistency/Staleness• Query Response Time

SLAs – achieved through Workload Management

Maximize:

QOS for NoSQL DBs

Y. Zhu et al. “Scheduling with Freshness and Performance Guarantees for Web Applications in the Cloud“, CRPIT

Old: Workload management in RDBMs (DB2 andOracle)

New:Use well-known scheduling algorithms for queriesin replicated DBs

Resource Provisionig

Goal: Resources ⇔ SLAs

T. Lorido-Botran, J. Miguel-Alonso et al.: “Auto-scaling Techniques forElastic Applications in Cloud Environments”. Technical Report, 2013

Resources

Expected Load

Resources

Expected Load

Provisioned Resources:• #No of Shard- or Replica

servers• Computing, Storage,

Network Capacities

Resources

Actual Load

Resources

Actual Load

Overprovisioning:• SLAs met• Excess Capacities

Resources

Actual Load

Underprovisioning:• SLAs violated• Usage maximized

Resources

Actual Load

Underprovisioning:• SLAs violated• Usage maximized

SmartSLA

P. Xiong: “Intelligent management of virtualized resources for database systems in cloud environment”, ICDE 2011

Solution: machine learning (regression + boosting) for prediction choose allocation that minimizesSLA penalties

Resourceallocation

Databaseperformance

LearnMapping

Functional Requirements

Scan-Querys

Conditional Updates

Transactions

Query by Example

Analytics

Elasticity

Consistency

Read-Latency

Write-Latency

Write-Throughput

Scalability of Data Volume

Read Scalability

Read-Availability

Write-Availability

Non-Functional Requirements

Durability

Write Scalability

DBaaS: General Considerations

Scan-Querys

Conditional Updates

Transactions

Query by Example

Analytics

Elasticity

Consistency

Read-Latency

Write-Latency

Write-Throughput

Read Scalability

Read-Availability

Write-Availability

Durability

Write Scalability

Scan-Querys

Conditional Updates

Transactions

Query by Example

Analytics

Elasticity

Consistency

Read-Latency

Write-Latency

Write-Throughput

Read Scalability

Read-Availability

Write-Availability

Durability

Write Scalability

Scan-Querys

Conditional Updates

Transactions

Query by Example

Analytics

Elasticity

Consistency

Read-Latency

Write-Latency

Write-Throughput

Read Scalability

Read-Availability

Write-Availability

Durability

Write Scalability

Scan-Querys

Conditional Updates

Transactions

Query by Example

Analytics

Elasticity

Consistency

Read-Latency

Write-Latency

Write-Throughput

Read Scalability

Read-Availability

Write-Availability

Durability

Write Scalability

Questions to ask: • Which requirements are met by the DB?• Which are met by the provider? SLAs

Outline

Examples of different clouddatabase systems:• Cloud-deployed• Managed DBMS

• SQL• NoSQL

• Proprietary• BaaS

Idea: Run (mostly) unmodified DB on IaaS

Cloud-Deployed DB

Method I: DIY

Method II: Deployment Tools

Method III: Marketplaces

Cloud-Deployed DB

Method I: DIY

1. Provision VM(s)

Cloud-Deployed DB

Method I: DIY

1. Provision VM(s) 2. Install DBMS (manual, script, Chef, Puppet)

Cloud-Deployed DB

Method I: DIY

> whirr launch-cluster --confighbase.properties

Login, cluster-size etc. Amazon EC2

Cloud-Deployed DB

Method I: DIY

> whirr launch-cluster --confighbase.properties

Login, cluster-size etc. Amazon EC2

Idea: Run preconfigured DB on IaaS

AWS Marketplace AWS

Marketplace

Model:

Cloud-Deployed

Pricing:

Instance +

Volume +

License

Underlying DB:

Choosable

DB-specific

AWS Marketplace AWS

Marketplace

Model:

Cloud-Deployed

Pricing:

Instance +

Volume +

License

Underlying DB:

Choosable

DB-specific

AWS Marketplace AWS

Marketplace

Model:

Cloud-Deployed

Pricing:

Instance +

Volume +

License

Underlying DB:

Choosable

DB-specific

AWS Marketplace AWS

Marketplace

Model:

Cloud-Deployed

Pricing:

Instance +

Volume +

License

Underlying DB:

Choosable

DB-specific

AWS Marketplace AWS

Marketplace

Model:

Cloud-Deployed

Pricing:

Instance +

Volume +

License

Underlying DB:

Choosable

DB-specific

AWS Marketplace AWS

Marketplace

Model:

Cloud-Deployed

Pricing:

Instance +

Volume +

License

Underlying DB:

Choosable

DB-specific

AWS Marketplace AWS

Marketplace

Model:

Cloud-Deployed

Pricing:

Instance +

Volume +

License

Underlying DB:

Choosable

DB-specific

AWS Marketplace AWS

Marketplace

Model:

Cloud-Deployed

Pricing:

Instance +

Volume +

License

Underlying DB:

Choosable

DB-specific

Bad: • No Clusters• Not managed (automatic Updates, Snapshots, etc.)• Private OS Multi-Tenancy Bad Resource Usage

• Easy to get started

Amazon Elastic MapReduce EMR

Model:

Analytics-aaS

Pricing:

Infrastructure

Hadoop

Amazon Elastic

MapReduce

Model:

Analytics-aaS

Pricing:

Infrastructure

Hadoop

Amazon Elastic

MapReduce

Provisions

Model:

Analytics-aaS

Pricing:

Infrastructure

Hadoop

Amazon Elastic

MapReduce

Job Tracker

Task Tracker + HDFS Data Node

Task Tracker

Provisions

Model:

Analytics-aaS

Pricing:

Infrastructure

Hadoop

Amazon Elastic

MapReduce

Data Source and Sink

Job Tracker

Task Tracker

Provisions

Model:

Analytics-aaS

Pricing:

Infrastructure

Hadoop

Amazon Elastic

MapReduce

Submits Hadoop Jobs as:• JAR• Streaming• Cascading• Pig• Hive• Impala

Job Tracker

Task Tracker

Provisions

Model:

Analytics-aaS

Pricing:

Infrastructure

Hadoop

Amazon Elastic

MapReduce

Submits Hadoop Jobs as:• JAR• Streaming• Cascading• Pig• Hive• Impala

Job Tracker

Task Tracker

Provisions

• No data locality with S3• AWS Import/Export: send your HDD• HBase Integration• Compatible with Spot and Reserved Instances• Similar: Azure HDInsight

Idea: Web-scale analysis of nested data

Google BigQuery BigQuery

Model:

Analytics-aaS

Pricing:

Storage + GBs

Processed

Google

BigQuery

Model:

Analytics-aaS

Pricing:

Storage + GBs

Processed

Google

BigQuery

Model:

Analytics-aaS

Pricing:

Storage + GBs

Processed

Google

BigQuery

Dremel

Melnik et al. “Dremel: Interactive analysis of web-scale datasets”, VLDB 2010

Idea:Multi-Level execution tree on nested columnar data format(≥100 nodes)

Model:

Analytics-aaS

Pricing:

Storage + GBs

Processed

Google

BigQuery

Dremel

Melnik et al. “Dremel: Interactive analysis of web-scale datasets”, VLDB 2010

Idea:Multi-Level execution tree on nested columnar data format(≥100 nodes)

• SLA: 99.9% uptime / month• Fundamentally different from relational DWHs

and MapReduce• Design copied by Apache Drill, Impala, Shark

Relational Database Service

Amazon RDS RDS

Model:

Managed RDBMS

Pricing:

Instance + Volume

+ License

Underlying DB:

MySQL, Postgres,

MSSQL, Oracle

DB-specific

Amazon RDS RDS

Model:

Managed RDBMS

Pricing:

Instance + Volume

+ License

Underlying DB:

MySQL, Postgres,

MSSQL, Oracle

DB-specific

Amazon RDS RDS

Model:

Managed RDBMS

Pricing:

Instance + Volume

+ License

Underlying DB:

MySQL, Postgres,

MSSQL, Oracle

DB-specific

• Synchronous Replication• Automatic Failover

Amazon RDS RDS

Model:

Managed RDBMS

Pricing:

Instance + Volume

+ License

Underlying DB:

MySQL, Postgres,

MSSQL, Oracle

DB-specific

99,95% uptime SLA

Amazon RDS RDS

Model:

Managed RDBMS

Pricing:

Instance + Volume

+ License

Underlying DB:

MySQL, Postgres,

MSSQL, Oracle

DB-specific

99,95% uptime SLA

Provisioned IOPS: access to EBS volumes network-optimized (up to 4000 IOPS)

Amazon RDS RDS

Model:

Managed RDBMS

Pricing:

Instance + Volume

+ License

Underlying DB:

MySQL, Postgres,

MSSQL, Oracle

DB-specific

Amazon RDS RDS

Model:

Managed RDBMS

Pricing:

Instance + Volume

+ License

Underlying DB:

MySQL, Postgres,

MSSQL, Oracle

DB-specific

EC2 instances: Up to 32 Cores, 244 GB RAM, 10 GbE

Amazon RDS RDS

Model:

Managed RDBMS

Pricing:

Instance + Volume

+ License

Underlying DB:

MySQL, Postgres,

MSSQL, Oracle

DB-specific

EC2 instances: Up to 32 Cores, 244 GB RAM, 10 GbE

Minor Version Upgrades are performed without downtime

Amazon RDS RDS

Model:

Managed RDBMS

Pricing:

Instance + Volume

+ License

Underlying DB:

MySQL, Postgres,

MSSQL, Oracle

DB-specific

Amazon RDS RDS

Model:

Managed RDBMS

Pricing:

Instance + Volume

+ License

Underlying DB:

MySQL, Postgres,

MSSQL, Oracle

DB-specific

Backups are automated and scheduled

Amazon RDS RDS

Model:

Managed RDBMS

Pricing:

Instance + Volume

+ License

Underlying DB:

MySQL, Postgres,

MSSQL, Oracle

DB-specific

Backups are automated and scheduled

• Support for (asynchronous) Read Replicas• Administration: Web-based or SDKs• Only RDBMSs• “Analytic Brother“ of RDS: RedShift (PDWH)

Similar to RDS

Microsoft SQL Azure SQL Azure

Model:

Managed RDBMS

Pricing:

Database size

Underlying DB:

MSSQL Server

T-SQL/TDS

SQL Azure

Similar to RDS

Model:

Managed RDBMS

Pricing:

Database size

Underlying DB:

MSSQL Server

T-SQL/TDS

SQL Azure

Similar to RDS

Model:

Managed RDBMS

Pricing:

Database size

Underlying DB:

MSSQL Server

T-SQL/TDS

SQL Azure

Similar to RDS

Model:

Managed RDBMS

Pricing:

Database size

Underlying DB:

MSSQL Server

T-SQL/TDS

SQL Azure

Similar to RDS

Model:

Managed RDBMS

Pricing:

Database size

Underlying DB:

MSSQL Server

T-SQL/TDS

SQL Azure

Similar to RDS

Model:

Managed RDBMS

Pricing:

Database size

Underlying DB:

MSSQL Server

T-SQL/TDS

SQL Azure

Cloud SQL Server

P. Bernstein et al. “Adapting Microsoft SQL server for cloud computing”, ICDE 2011

• Multi-Tenant MSSQL• Paxos-like commit protocol for

consistent replication

Similar to RDS

Model:

Managed RDBMS

Pricing:

Database size

Underlying DB:

MSSQL Server

T-SQL/TDS

SQL Azure

Keyless Table Group: regular databaseKeyed Table Group: partitioned by row key

Cloud SQL Server

Similar to RDS

Model:

Managed RDBMS

Pricing:

Database size

Underlying DB:

MSSQL Server

T-SQL/TDS

SQL Azure

Consistency unit (ACID boundary)

Cloud SQL Server

Similar to RDS

Model:

Managed RDBMS

Pricing:

Database size

Underlying DB:

MSSQL Server

T-SQL/TDS

SQL Azure

Automatic Partitioning for Keyed Table Groups

Cloud SQL Server

Similar to RDS

Model:

Managed RDBMS

Pricing:

Database size

Underlying DB:

MSSQL Server

T-SQL/TDS

SQL Azure

Automatic Partitioning for Keyed Table Groups

Cloud SQL Server

• SLA: 99.9% uptime / month• Usually Cheaper than RDS (Multi-Tenancy)• Smaller Databases (max. 150 GB)• Rich MSSQL server tooling• Keyed Table Group internal feature only

MySQL for Google App Engine PaaS

Support for: patching, replication, backup

SLA: 99,95 % uptime / month

Other RDBMS servicesGoogle

Cloud SQL Google Cloud SQL

Pricing:

Database size

Underlying DB:

Postgres for Heroku PaaS

Hosted on EC2

No SLAs

Pricing:

Database size

Underlying DB:

Heroku Postgres

Pricing:

Plan based

Underlying DB:

Postgres

Hosted on EC2

No SLAs

MySQL for OpenStack (Icehouse)

Under development (HP driven)

VM ⇔ Tenant

Pricing:

Database size

Underlying DB:

Heroku Postgres

Pricing:

Plan based

Underlying DB:

Postgres

Pricing:

Own Hardware

Underlying DB:

Hosted on EC2

No SLAs

MySQL for OpenStack (Icehouse)

Under development (HP driven)

VM ⇔ Tenant

Pricing:

Database size

Underlying DB:

Heroku Postgres

Pricing:

Plan based

Underlying DB:

Postgres

Pricing:

Own Hardware

Underlying DB:

Evaluation of Cloud RDBMSs

D. Kossmann,T. Kraska: An evaluation of alternative architectures for transaction processing in the cloud”, Sigmod 2010

TPC-W Benchmark (Online Shop), 2010, Emulated Browsers / RPS:

HBase Wide-Column

CP Over Row Key

~700 1/4 Apache

MongoDB Doc-ument

CP yes >100<500

4/4 GPL

Riak Key-Value

AP ~60 3/4 Apache

(Softlayer)

Cassandra Wide-Column

AP WithComp. Index

>300<1000

2/4 Apache

Redis Key-Value

CA Through Lists, etc.

manual N/A 4/4 BSD

Managed NoSQL services

Model CAP ScansSec.

IndicesLargestCluster

Lic.Lear-ning DBaaS

HBase Wide-Column

CP Over Row Key

~700 1/4 Apache

MongoDB Doc-ument

CP yes >100<500

4/4 GPL

Riak Key-Value

AP ~60 3/4 Apache

(Softlayer)

Cassandra Wide-Column

AP WithComp. Index

>300<1000

2/4 Apache

Redis Key-Value

CA Through Lists, etc.

manual N/A 4/4 BSD

Managed NoSQL services

Model CAP ScansSec.

IndicesLargestCluster

Lic.Lear-ning DBaaS

And there are many more:• CouchDB (e.g. Cloudant)• CouchBase (e.g. KuroBase Beta)• ElasticSearch(e.g. Bonsai)• Solr (e.g. WebSolr)• …

MongoHQ MongoHQ

Model:

Managed NoSQL

Pricing:

Plan-based

Underlying DB:

MongoDB

Mongo, REST (beta)

EC2-based MongoDB-as-a-Serivce

MongoHQ MongoHQ

Model:

Managed NoSQL

Pricing:

Plan-based

Underlying DB:

MongoDB

Mongo, REST (beta)

MongoHQ MongoHQ

Model:

Managed NoSQL

Pricing:

Plan-based

Underlying DB:

MongoDB

Mongo, REST (beta)

MongoHQ MongoHQ

Model:

Managed NoSQL

Pricing:

Plan-based

Underlying DB:

MongoDB

Mongo, REST (beta)

MongoHQ MongoHQ

Model:

Managed NoSQL

Pricing:

Plan-based

Underlying DB:

MongoDB

Mongo, REST (beta)

MongoHQ MongoHQ

Model:

Managed NoSQL

Pricing:

Plan-based

Underlying DB:

MongoDB

Mongo, REST (beta)

Private Process Multi-TenancyScale-Up Strategy: RAM, IOPs and CPU increased at runtimeMaximum Size: 1TB

Free Tier

MongoHQ MongoHQ

Model:

Managed NoSQL

Pricing:

Plan-based

Underlying DB:

MongoDB

Mongo, REST (beta)

Private Process Multi-TenancyScale-Up Strategy: RAM, IOPs and CPU increased at runtimeMaximum Size: 1TB

Free Tier

VM-Deployment (EC2):M1.large on EC2: $128.10on EC2 with 1-yr-Res.: $30With MongoHQ: $637

MongoHQ MongoHQ

Model:

Managed NoSQL

Pricing:

Plan-based

Underlying DB:

MongoDB

Mongo, REST (beta)

#1 thing that should never happen:

MongoHQ

Problem: Scalability

Client

configconfigconfig

mongos

Replica Set

Master

mongos

MongoHQ

Client

configconfigconfig

mongos

Replica Set

Master

mongos

What if Writes / second or data volume become bottleneck?

MongoHQ

Client

configconfigconfig

mongos

Replica Set

Master

mongos

Sharding (Scale Out)• Dynamic Scaling: Tenant adds

Replica Set• Elastic Scaling: Provider

adds/removes Replica SetMongoHQ: Only manual shardingon “contact us” basis

MongoHQ

Client

configconfigconfig

mongos

Replica Set

Master

mongos

Sharding (Scale Out)• Dynamic Scaling: Tenant adds

Replica Set• Elastic Scaling: Provider

adds/removes Replica SetMongoHQ: Only manual shardingon “contact us” basis

• Bad: no SLAs, no horizontal scaling

• Good: Solid Dashboard, Backups, Replication, integration with Heroku

• Competitors: MongoLab, ObjectRocket, MongoSoup

ElastiCache ElastiCache

Model:

Managed NoSQL

Pricing:

Infrastructure

Underlying DB:

Memcache, Redis

DB, REST

(management)

„RDS for Memcache and Redis“

Model:

Managed NoSQL

Pricing:

Infrastructure

Underlying DB:

Memcache, Redis

DB, REST

(management)

Memcache or RedisMemcache can be run as a cluster (client-side sharding)

Limited choice of instance types

Model:

Managed NoSQL

Pricing:

Infrastructure

Underlying DB:

Memcache, Redis

DB, REST

(management)

„RDS for Memcache and Redis“Announced last Saturday: Snapshots

Model:

Managed NoSQL

Pricing:

Infrastructure

Underlying DB:

Memcache, Redis

DB, REST

(management)

Model:

Managed NoSQL

Pricing:

Infrastructure

Underlying DB:

Memcache, Redis

DB, REST

(management)

$ elasticache-modify-cache-parameter-group xy

AWS CLI tools and SDKsoffer a superset of Dashboard functions

Many Hosted NoSQLDbaaS Providers represented

Heroku DBaaS Addons

Many Hosted NoSQLDbaaS Providers represented

And Search

Heroku DBaaS Addons

Heroku Redis2Go example Redis2Go

Model:

Managed NoSQL

Pricing:

Plan-based

Underlying DB:

Create Heroku App:

Model:

Managed NoSQL

Pricing:

Plan-based

Underlying DB:

Create Heroku App:

Add Redis2Go Addon:

Model:

Managed NoSQL

Pricing:

Plan-based

Underlying DB:

Create Heroku App:

Add Redis2Go Addon:

Use Connection URL (environment variable):

Model:

Managed NoSQL

Pricing:

Plan-based

Underlying DB:

Create Heroku App:

Add Redis2Go Addon:

Deploy:

Model:

Managed NoSQL

Pricing:

Plan-based

Underlying DB:

Create Heroku App:

Add Redis2Go Addon:

Deploy:• Very simple• Only suited for small to medium

applications (no SLAs, limited control)

SimpleDB Table-Store

CP Yes (asqueries)

Auto-matic

SQL-like(no joins, groups, …)

REST + SDKs

Dynamo-DB

Table-Store

CP By rangekey / index

Local Sec.Global Sec.

Key+Cond. On Range Key(s)

REST + SDKs

Automaticover Prim. Key

AzureTables

Table-Store

CP By rangekey

Key+Cond. On Range Key

REST + SDKs

Automaticover Part. Key

99.9% uptime

AE/CloudDataStore

Entity-Group

CP Yes (asqueries)

Auto-matic

Conjunct.of Eq. Predicates

REST/SDK, JDO,JPA

Automaticover EntityGroups

S3, Az. Blob, GCS

Blob-Store

AP REST + SDKs

Automaticover key

99.9% uptime(S3)

Proprietary Database services

Model CAP ScansSec.

IndicesQueries API SLA

Scale-out

SimpleDB Table-Store

CP Yes (asqueries)

Auto-matic

SQL-like(no joins, groups, …)

REST + SDKs

Dynamo-DB

Table-Store

CP By rangekey / index

Local Sec.Global Sec.

Key+Cond. On Range Key(s)

REST + SDKs

Automaticover Prim. Key

AzureTables

Table-Store

CP By rangekey

Key+Cond. On Range Key

REST + SDKs

Automaticover Part. Key

99.9% uptime

AE/CloudDataStore

Entity-Group

CP Yes (asqueries)

Auto-matic

Conjunct.of Eq. Predicates

REST/SDK, JDO,JPA

Automaticover EntityGroups

S3, Az. Blob, GCS

Blob-Store

AP REST + SDKs

Automaticover key

99.9% uptime(S3)

Proprietary Database services

Model CAP ScansSec.

IndicesQueries API SLA

Scale-out

There are many more object stores (HP, Rackspace, etc.)…but no comparable Table Stores

Azure Storage

Load-Balancing-System

Application

HTTP/HTTPS

Worker-Role

Web-Role

IIS-Webserver

Virtual Machines

OtherServices

Not Allowed

Virtual Machines

Azure Storage

Load-Balancing-System

Application

HTTP/HTTPS

Worker-Role

Web-Role

IIS-Webserver

Virtual Machines

OtherServices

Not Allowed

Virtual Machines

Tables

Queues

Table Service example: Azure Tables

Partition Key

Row Key (sortiert)

Timestamp(autom.)

Property1 Propertyn

intro.pdf v1.1 14/6/2013 … …

intro.pdf v1.2 15/6/2013 …

präs.pptx v0.0 11/6/2013 …

Partition

Partition Key

Row Key (sortiert)

Timestamp(autom.)

Property1 Propertyn

intro.pdf v1.1 14/6/2013 … …

intro.pdf v1.2 15/6/2013 …

präs.pptx v0.0 11/6/2013 …

Partition

SparseHash-distributed toparition servers

No Index: Lookup only (!) by full table scanAtomic "Entity-Group Batch Transaction" possible

Similar to Amazon SimpleDB and DynamoDB

Partition Key

Row Key (sortiert)

Timestamp(autom.)

Property1 Propertyn

intro.pdf v1.1 14/6/2013 … …

intro.pdf v1.2 15/6/2013 …

präs.pptx v0.0 11/6/2013 …

Partition

• Indexes all attributes• Rich(er) queries• Many Limits (size, RPS, etc.)

• Provisioned Throughput• On SSDs („single digit latency“)• Optional Indexes

Azure Table Storage Azure Tables

Model:

Propriertary

Pricing:

Requests + Storage

+ Network

Underlying DB:

Custom System

Challenges:

Single partition and range key (modelling) Very “basic“ queries

Azure Tables

Model:

Propriertary

Pricing:

Requests + Storage

+ Network

Underlying DB:

Custom System

Challenges:

Azure Tables

Model:

Propriertary

Pricing:

Requests + Storage

+ Network

Underlying DB:

Custom System

Challenges:

Good:Automatic DistributionReplicated 3x locally + 1x async. geo-replica

Azure Tables

Model:

Propriertary

Pricing:

Requests + Storage

+ Network

Underlying DB:

Custom System

Challenges:

Very good:SLA (99.9% uptime)Internal architecture published

Azure Tables

Model:

Propriertary

Pricing:

Requests + Storage

+ Network

Underlying DB:

Custom System

Challenges:

Azure Tables

Windows Azure Storage

B. Calder, et al. "Windows Azure Storage: a highly available cloud storage service with strong consistency." , SOSP 2011

Idea:• Layered storage infrastructure for Blobs and Tables• Use research results (GFS, BigTable, Paxos, LSM, Erasure Coding)

Model:

Propriertary

Pricing:

Requests + Storage

+ Network

Underlying DB:

Custom System

Challenges:

Azure Tables

Windows Azure Storage

B. Calder, et al. "Windows Azure Storage: a highly available cloud storage service with strong consistency." , SOSP 2011

Idea:• Layered storage infrastructure for Blobs and Tables• Use research results (GFS, BigTable, Paxos, LSM, Erasure Coding)

Think: GFS/HDFSThink: BigTable/HBase

Think: Chubby/ZooKeeper

DynamoDB DynamoDB

Model:

Propriertary

Pricing:

Provisioned

Throughput +

Network

Underlying DB:

Custom System

Successor to SimpleDB

Limitations: • Slow (~50-100 RPS)• 10 GB per Domain• Query result max. 2500 records or

1 MB• Max. 1K-sized attributes

DynamoDB DynamoDB

Model:

Propriertary

Pricing:

Provisioned

Throughput +

Network

Underlying DB:

Custom System

dom:com.cnn content : "<html>…"

Primary Key Attribute (scalar or set)

page:index

Range Key

DynamoDB DynamoDB

Model:

Propriertary

Pricing:

Provisioned

Throughput +

Network

Underlying DB:

Custom System

Querying Options:

• GetItem: Key Lookup

• Query: Primary Key + Condition on Range Key

• Scan: Full Table Scan with filter

• EMR: Hive queries (for analytics)

page:index

Range Key

DynamoDB DynamoDB

Model:

Propriertary

Pricing:

Provisioned

Throughput +

Network

Underlying DB:

Custom System

Consistency: Strongly (2x price) or Eventually Consistent Reads

Atomic (Conditional) Updates per Item

Indexing Options:◦ Local Sec. Index: consistent additional Range Key

◦ Global Sec. Index: eventually consistent index-table (Primary Key)

page:index

Range Key

DynamoDB DynamoDB

Model:

Propriertary

Pricing:

Provisioned

Throughput +

Network

Underlying DB:

Custom System

DynamoDB DynamoDB

Model:

Propriertary

Pricing:

Provisioned

Throughput +

Network

Underlying DB:

Custom System

DynamoDB DynamoDB

Model:

Propriertary

Pricing:

Provisioned

Throughput +

Network

Underlying DB:

Custom System

DynamoDB DynamoDB

Model:

Propriertary

Pricing:

Provisioned

Throughput +

Network

Underlying DB:

Custom System

DynamoDB DynamoDB

Model:

Propriertary

Pricing:

Provisioned

Throughput +

Network

Underlying DB:

Custom System

Unit of Billing

DynamoDB DynamoDB

Model:

Propriertary

Pricing:

Provisioned

Throughput +

Network

Underlying DB:

Custom System

Unit of BillingGood:• Low Latency (SSD)• Data partitioning and AZ-replicationBad:• Scaling not elastic (Capacity Units)• No SLAs, no internals published (≠ Dynamo!)• No built-in backups ( AWS data pipeline)• Vendor Lock-in

AE/Cloud DataStore DataStore

Model:

Propriertary

Pricing:

CPU + Storage +

Network

Underlying DB:

MegaStore

SDK, JPA, JDO

Google Cloud

Datastore

Structured Storage System for App Engine

Based on:◦ Megastore BigTable Colossus

Model:

Propriertary

Pricing:

CPU + Storage +

Network

Underlying DB:

MegaStore

SDK, JPA, JDO

Google Cloud

Datastore

Model:

Propriertary

Pricing:

CPU + Storage +

Network

Underlying DB:

MegaStore

SDK, JPA, JDO

Google Cloud

Datastore

Model:

Propriertary

Pricing:

CPU + Storage +

Network

Underlying DB:

MegaStore

SDK, JPA, JDO

Google Cloud

Datastore

Schemafree Entity Group (EG) data model:

IDName

IDUserURL

Root Table Child Table

Model:

Propriertary

Pricing:

CPU + Storage +

Network

Underlying DB:

MegaStore

SDK, JPA, JDO

Google Cloud

Datastore

IDName

IDUserURL

EG: User + n Photos• Unit of ACID transactions/

consistency• Fields autoindexed

(eventually consistent)

Model:

Propriertary

Pricing:

CPU + Storage +

Network

Underlying DB:

MegaStore

SDK, JPA, JDO

Google Cloud

Datastore

IDName

IDUserURL

EG: User + n Photos• Unit of ACID transactions/

consistency• Fields autoindexed

(eventually consistent)

SELECT * FROM photos

WHERE ANCESTOR IS :34 AND name = „sunset“

ORDER BY date ASC

LIMIT 10

OFFSET 10

AE/Cloud DataStore

Internally:

Entity Groups define partitions

SynchronousPaxos-based

replication

ACID per EG. Maximum of1 Write/s to an EG.

Eventual Consistencyacross groups

Stored in BigTable

AE/Cloud DataStore

Internally:

replication

Stored in BigTable

AE/Cloud DataStore

Internally:

MegaStore

J. Baker, et al. "Megastore: Providing Scalable, Highly Available Storage for Interactive Services." CIDR 2011.

• Paxos-based replication andtransactions

• 100 Google applicationsProblems: Slow Writes, PredefinedEntity Groups

replication

Stored in BigTable

AE/Cloud DataStore

Internally:

MegaStore

replication

Stored in BigTable

Spanner

J. Corbett et al. "Spanner: Google’s globally distributed database." TOCS 2013

Idea:• Autosharded Entity Groups• Not based on BigTableImplementation:• TrueTime API (GPS + atomic

clocks) commit timestampsof 2PL-SI transactions

• Paxos-replication per Shard

AE/Cloud DataStore

Internally:

MegaStore

replication

Stored in BigTable

Spanner

J. Shute, et al. "F1: A distributed SQL database that scales.“, VLDB 2013

Idea:• Full SQL relational database built on

Spanner, powers AdWordsImplementation:• 5-way replication• Data Model: relational + hierarchy

(customercampaignAdGroup)• Transactions: Snapshot-Read-Only,

Pessimistic (Spanner), optimistic• Distributed SQL Engine

AE/Cloud DataStore

Internally:

MegaStore

replication

Stored in BigTable

Spanner

J. Shute, et al. "F1: A distributed SQL database that scales.“, VLDB 2013

Idea:• Full SQL relational database built on

Spanner, powers AdWordsImplementation:• 5-way replication• Data Model: relational + hierarchy

(customercampaignAdGroup)• Transactions: Snapshot-Read-Only,

Pessimistic (Spanner), optimistic• Distributed SQL Engine

Good:• Transactions (though limited)• Good scalability of data volumeBad:• Entity Groups hard to define• Bad Scale-Out for write and reads (Kossmann et

al.) no advantage over RDBMS for small datavolumes

A Spanner/F1 based DBaaS?

Idea: Blobs with RESTful CRUD Interface

Distribution: 𝐵𝑢𝑐𝑘𝑒𝑡, 𝐾𝑒𝑦 → 𝑆𝑒𝑟𝑣𝑒𝑟 + 𝑅𝑒𝑝𝑙𝑖𝑐𝑎𝑠

CDN Integration (Azure CDN, Cloudfront)

S3: > 2 Trillion (2 ⋅ 1012) objects

Azure Blobs, Amazon S3, Google Cloud Storage

Container

Blobcontains

Block 1 (up to 4 MB)

Block 2

Block n

REST-Request

Azure Blobs:

Block Blob: 4MB blocks large filesPage Blob: 512B blocks random IO

Automatic Versioning

Pricing: per request (10.000 ~ 1c) and storage(1GB/month ~ 3c) and network (1GB ~ 10c)

S3 BlobDELETE /puppy.jpg HTTP/1.1

Host: mybucket.s3.amazonaws.comAuthorization: AWS AKIAIO...

AWS Ireland DC

Replicas

Reduced Redundancy: only 1 replicaGlacier: tape-disk archivalAmazon S3:

AWS Ireland DC

Replicas

S3 Consistency

D. Bermbach,, S. Tai. "Eventual consistency: How soon is eventual?”, MW4SOC 11

Findings:• Inconsistency window varies

from 2-11 seconds• Monotonic Read Consistency

is often violated

AWS Ireland DC

Replicas

S3 Consistency

D. Bermbach,, S. Tai. "Eventual consistency: How soon is eventual?”, MW4SOC 11

Findings:• Inconsistency window varies

from 2-11 seconds• Monotonic Read Consistency

is often violated

Building a database on S3

M. Brantner, et al. "Building a database on S3." Sigmod 2008

Idea:• Use S3 as the persistent storage of a

database

Implementation:• Buffer Pool and Log Manager on S3• No transaction or query support

Founded June, 2011

Acquired by Facebook April, 2013

Pricing: ◦ Free

◦ Pro (199$)

◦ Enterprise

Parse - MBaaS Parse

Model:

Backend-aas

Pricing:

Plan-based

Underlying DB:

Mainly MongoDB

SDKs, REST

Founded June, 2011

Pricing: ◦ Free

◦ Pro (199$)

◦ Enterprise

Parse - MBaaS Parse

Model:

Backend-aas

Pricing:

Plan-based

Underlying DB:

Mainly MongoDB

SDKs, REST

Parse Core

Founded June, 2011

Pricing: ◦ Free

◦ Pro (199$)

◦ Enterprise

Parse - MBaaS Parse

Model:

Backend-aas

Pricing:

Plan-based

Underlying DB:

Mainly MongoDB

SDKs, REST

Parse Core Parse Analytics

Founded June, 2011

Pricing: ◦ Free

◦ Pro (199$)

◦ Enterprise

Parse - MBaaS Parse

Model:

Backend-aas

Pricing:

Plan-based

Underlying DB:

Mainly MongoDB

SDKs, REST

Parse Core Parse Analytics Parse Push

Parse - MBaaS

AuthenticationUser + PasswordOAuth: Facebook, Twitter

Parse - MBaaS

Query Cinemas(new Parse.Query('Cinemas'))

.withinKilometers(...)

.fetch()

Parse - MBaaS

Query Cinemas(new Parse.Query('Cinemas'))

.withinKilometers(...)

.fetch()

Query Movies(new Parse.Query('Movies'))

.greaterThan('startAt‘, now)

.notEqualTo('cinemas', cId)

.fetch()

Meteor Meteor

Model:

Backend-aaS

Pricing:

No yet revealed

Underlying DB:

MongoDB

WebSockets

Idea: Full-Stack JavaScript with Node.js, MongoDB and WebSockets

Web Browser Node.js + MongoDB(Single Server)

WebSocket

Meteor Meteor

Model:

Backend-aaS

Pricing:

No yet revealed

Underlying DB:

MongoDB

WebSockets

<div class="player {{selected}}"><span class="name">{{name}}</span><span class="score">{{score}}</span>

</div>

WebSocket

Meteor Meteor

Model:

Backend-aaS

Pricing:

No yet revealed

Underlying DB:

MongoDB

WebSockets

Players = new Meteor.

Collection("players");

if (Meteor.isServer) {

</div>

WebSocket

Meteor Meteor

Model:

Backend-aaS

Pricing:

No yet revealed

Underlying DB:

MongoDB

WebSockets

if (Meteor.isClient) {

Template.leaderboard.players = function () {

return Players.find({},

{sort: {score: -1, name: 1}});

</div>

WebSocket

Meteor Meteor

Model:

Backend-aaS

Pricing:

No yet revealed

Underlying DB:

MongoDB

WebSockets

</div>

WebSocket

$ meteor deploy abc.meteor.com

Meteor Meteor

Model:

Backend-aaS

Pricing:

No yet revealed

Underlying DB:

MongoDB

WebSockets

</div>

WebSocket

$ meteor deploy abc.meteor.com

Very productive for very small projects

Fundamentally limited scalability:• Server tails Mongo‘s oplog• And holds the fetched data state of

every client

Outline

• Hot Topics• Orestes: a scalable, low-

latency architecture• Baqend: putting it into

practice

Example: CryptDB

Idea: Only decrypt as much as neccessary

Encrypted Databases: Research

SQL-Proxy

Encrypts and decrypts, rewrites queries

Example: CryptDB

SQL-Proxy

Example: CryptDB

SQL-Proxy

Relational Cloud

C. Curino, et al. "Relational cloud: A database-as-a-service for the cloud.“, CIDR 2011

DBaaS Architecture:• Encrypted with CryptDB• Multi-Tenancy through live

migration• Workload-aware partitioning

(graph-based)

Example: CryptDB

SQL-Proxy

Relational Cloud

C. Curino, et al. "Relational cloud: A database-as-a-service for the cloud.“, CIDR 2011

DBaaS Architecture:• Encrypted with CryptDB• Multi-Tenancy through live

migration• Workload-aware partitioning

(graph-based)

• Early approach• Not adopted in practice, yet

Dream solution:Full Homorphic Encryption

Transactions/Consistency: Research

Dynamo Eventual None 1 RT -

Yahoo PNuts Timeline per key Single Key 1 RT possible

COPS Causality Multi-Record 1 RT possible

MySQL (async) Serializable Static Partition 1 RT possible

Megastore Serializable Static Partition 2 RT -

Spanner/F1 Snapshot Isolation Partition 2 RT -

MDCC Read-Commited Multi-Record 1 RT -

Consistency Transactional UnitCommit Latency

Data Loss?

Multi-Data Center Consistency

T. Kraska et al. "MDCC: Multi-data center consistency." EuroSys, 2013.

Idea:• Multi-Data center commit protocol with

single round-trip

Implementation:• Optimistic Commit Protocol• Fast, Generalized Multi-Paxos

Result: almost as fast as Dynamo-style

Data Loss?

Multi-Data Center Consistency

T. Kraska et al. "MDCC: Multi-data center consistency." EuroSys, 2013.

Idea:• Multi-Data center commit protocol with

single round-trip

Implementation:• Optimistic Commit Protocol• Fast, Generalized Multi-Paxos

Result: almost as fast as Dynamo-style

Currently no NoSQL DB implementsconsistent Multi-DC replication

YCSB (Yahoo Cloud Serving Benchmark)

Benchmarking: Research

Data Store

Client

rkload

erator

ggable

interface

Data Store

Threads

Client

rkload

erator

ggable

interface

Workload:

1. Operation Mix

2. Record Size

3. Popularity Distribution

Runtime Parameters:

DB host name,

threads, etc.

Data Store

Threads

Client

rkload

erator

ggable

interface

Workload:

1. Operation Mix

2. Record Size

Runtime Parameters:

DB host name,

threads, etc.

Read()Insert()Update()Delete()Scan()

Data Store

Threads

DB protocol

Client

rkload

erator

ggable

interface

Workload:

1. Operation Mix

2. Record Size

Runtime Parameters:

DB host name,

threads, etc.

Read()Insert()Update()Delete()Scan()

Data Store

Threads

DB protocol

Workload Operation Mix Distribution Example

A – Update Heavy Read: 50%Update: 50%

Zipfian Session Store

B – Read Heavy Read: 95%Update: 5%

Zipfian Photo Tagging

C – Read Only Read: 100% Zipfian User Profile Cache

D – Read Latest Read: 95%Insert: 5%

Latest User Status Updates

E – Short Ranges Scan: 95%Insert: 5%

Zipfian/Uniform

Threaded Conversations

Example Result

(Read Heavy):

Example Result

(Read Heavy):

Weaknesses:• Single client can be a

bottleneck• No consistency &

availability measurement

Example Result

(Read Heavy):

Benchmarking: ResearchYCSB++

S. Patil, M. Polte, et al.„Ycsb++: benchmarking and performance debugging advanced features in scalable table stores“, SOCC 2011

• Clients coordinate throughZookeeper

• Simple Read-After-Write Checks• Evaluation: Hbase & Accumulo

Example Result

(Read Heavy):

• No Transaction Support

YCSB+T

A. Dey et al. “YCSB+T: Benchmarking Web-Scale Transactional Databases”, CloudDB 2014

• New workload: TransactionalBank Account

• Simple anomaly detection forLost Updates

• No comparison of systems

Example Result

(Read Heavy):

• No Transaction Support

YCSB+T

A. Dey et al. “YCSB+T: Benchmarking Web-Scale Transactional Databases”, CloudDB 2014

• New workload: TransactionalBank Account

• Simple anomaly detection forLost Updates

• No comparison of systems

No specific applicationCloudStone, CARE, TPC

extensions?

Benchmarking: state of the art

Tests latency and throughputof IaaS-Providers, CDNs and

Object Stores

Benchmarking: state of the art

Tests latency and throughputof IaaS-Providers, CDNs and

Object Stores

Does not test: Cloud Databases

VisionYCSB Harmony

YCSBClient

YCSBClient YCSB

Client

YCSBClient

Idea:1. Periodically

launch clients

VisionYCSB Harmony

YCSBClient

YCSBClient YCSB

Client

YCSBClient

Idea:1. Periodically

launch clients2. Benchmark

Cloud DB3. Publish results

VisionYCSB Harmony

System Availablity Reads/s Writes/s Avg. Latency 95th perc. Latency

DynamoDB 99.9% 23411 34534 3.2 ms 9 ms

RDS 99.8% 2342 2455 30 ms 80 ms

Azure Table 99.5% 22343 23442 12 ms 20 ms

GoogleDataStore

99.5% 3000 2000 30 ms 300 ms

Database research at theUniversity of Hamburg

Motivation

Classic 3-Tier-ArchitectureThree-Tier Architecture

Motivation

ClientApplicationDatabase

Web Server

Motivation

Web Server

Motivation

Web Server

High Latency

Motivation

Web Server

Average (2014):90 HTTP Requests per

page load

High Latency

Motivation

Web Server

page load

High Latency

Classic 3-Tier-ArchitectureThree-Tier ArchitectureWith every 100ms of additional page load time, revenue decreases by 1%.

Study by Amazon

Motivation

Web Server

page load

High Latency

Study by Amazon

When increasing load time of search results by 500ms, traffic decreases by 20%.

Study by Google

Motivation

Web Server

page load

High Latency

Study by Amazon

When increasing load time of search results by 500ms, traffic decreases by 20%.

Study by Google

Rich ClientApplicationDatabase

Web Server

Single-Page Applications

High Latency

(Rich Client, Smart Client)

Data (e.g. JSON)

(AngularJS, Backbone,

Ember.JS, etc.)

ORESTES

Rich ClientOrestesDB-Cluster

RESTServer

Low LatencyCloud

RESTServer

ORESTES

RESTServer

Web-Caches Low Latency

RESTServer

CacheableDatabase Objects

ORESTES

RESTServer

ScalableNoSQL DBs

RESTServer

ORESTES

RESTServer

ScalableNoSQL DBs

Exposes the DB via REST API and handles:• Scaling• Cache Consistency• Transactions• Schema

RESTServer

Middleware:◦ Scalable REST API for aggregate-oriented

persistence, queries, schema management and transactions.

Caching:◦ Consistent web caching of database objects for

read scalability and latency reduction.

Transactions:◦ Optimistic cache-aware transaction model.

ORESTES: Components

Java/JDO

persist

createQueryJavaScript/JPA Port

REST/HTTP API

others

Application

Server

Browser or

Mobile Device

ApplicationLayer

PersistenceAPI

DataStore

Java/JDO

persist

REST/HTTP API

others

Application

Server

Browser or

Mobile Device

ApplicationLayer

PersistenceAPI

DataStore

HTTP Server

Trans-

actionsQuerys

Object

Persist.Schema

uments

DBaaS & BaaS Layer

Transaction Validation

HTTP Server

Config-

urationPartial

Updates

Access Control

Multi-Tenancy Schema Management

Workload Management

Cache Coherence

Autoscaling

Database-independentConcerns

Database -specific Wrappers

HTTP Server

Java/JDO

persist

REST/HTTP API

others

Application

Server

Browser or

Mobile Device

ApplicationLayer

PersistenceAPI

DataStore

Forward-Proxy Caches

ISP Caches

Reverse-Proxy Caches and

Load Balancers

CDN Caches

Content DeliveryNetworks

HTTP Server

Trans-

actionsQuerys

Object

Persist.Schema

uments

DBaaS & BaaS Layer

HTTP Server

Config-

urationPartial

Updates

Access Control

Workload Management

Cache Coherence

Autoscaling

HTTP Server

Java/JDO

persist

REST/HTTP API

others

Application

Server

Browser or

Mobile Device

ApplicationLayer

PersistenceAPI

DataStore

ISP Caches

Load Balancers

CDN Caches

HTTP Server

Trans-

actionsQuerys

Object

Persist.Schema

uments

DBaaS & BaaS Layer

HTTP Server

Config-

urationPartial

Updates

Access Control

Workload Management

Cache Coherence

Autoscaling

HTTP Server

Redis (Replicated)

10201040

10101010Counting

Bloom Filter

delete

Node.JS (local to Server)

Stored Procedures

Custom Validation

Java/JDO

persist

REST/HTTP API

others

Application

Server

Browser or

Mobile Device

ApplicationLayer

PersistenceAPI

DataStore

ISP Caches

Load Balancers

CDN Caches

HTTP Server

Trans-

actionsQuerys

Object

Persist.Schema

uments

DBaaS & BaaS Layer

HTTP Server

Config-

urationPartial

Updates

Access Control

Workload Management

Cache Coherence

Autoscaling

HTTP Server

Redis (Replicated)

10201040

10101010Counting

Bloom Filter

delete

Node.JS (local to Server)

Stored Procedures

Custom Validation

GET /db/{bucket}/{class}/{id}

200 OKCache-Control: public, max-age=6000ETag: "3"JSON Object

Caching

Orestes

Client-

(Browser-)

CachesISP

CachesCDN

Caches

Reverse-

Caches

MissMiss

P(Cache-Hit)

0 ms 1 ms 10 ms 20 ms 50-500 ms 50-500 ms

Caching

Orestes

Client-

(Browser-)

CachesISP

CachesCDN

Caches

Reverse-

Caches

MissMiss

P(Cache-Hit)

0 ms 1 ms 10 ms 20 ms 50-500 ms 50-500 ms

em.find(id) JavaScript

Caching

Orestes

Client-

(Browser-)

CachesISP

CachesCDN

Caches

Reverse-

Caches

MissMiss

P(Cache-Hit)

0 ms 1 ms 10 ms 20 ms 50-500 ms 50-500 ms

GET /db/posts/{id} HTTP

Caching

Orestes

Client-

(Browser-)

CachesISP

CachesCDN

Caches

Reverse-

Caches

MissMiss

P(Cache-Hit)

0 ms 1 ms 10 ms 20 ms 50-500 ms 50-500 ms

Cache-Hit: Return ObjectCache-Miss: Forward Request

Caching

Orestes

Client-

(Browser-)

CachesISP

CachesCDN

Caches

Reverse-

Caches

MissMiss

P(Cache-Hit)

0 ms 1 ms 10 ms 20 ms 50-500 ms 50-500 ms

Fetch object from DB and returnit with caching information

Caching

Orestes

Client-

(Browser-)

CachesISP

CachesCDN

Caches

Reverse-

Caches

MissMiss

P(Cache-Hit)

0 ms 1 ms 10 ms 20 ms 50-500 ms 50-500 ms

Scalability and Cache-Hits

Caching

Orestes

Client-

(Browser-)

CachesISP

CachesCDN

Caches

Reverse-

Caches

MissMiss

P(Cache-Hit)

0 ms 1 ms 10 ms 20 ms 50-500 ms 50-500 ms

Scalability and Cache-Hits

Latency Benefit

BFB: Bloom Filter Bounded Staleness

1 4 020

How to prevent stalereads (inconsistency)?

1 4 020

purge(obj)

hashB(oid)hashA(oid)

1 4 020

purge(obj)

131 1 110Flat(Counting Bloomfilter)

1 4 020

purge(obj)

131 1 110

1 4 020

purge(obj)

131 1 110

1 4 020

purge(obj)

131 1 110

1 4 020

1 1 110

1 4 020

1 1 110

𝑓 ≈ 1 − 𝑒−𝑘𝑛𝑚

𝑘 = ln 2 ⋅ (𝑛

False-Positive

Functions:

With 10.000 Updates per 10 minutes and 1% error rate: 12 KByte

1 4 020

1 1 110

SCOT: Scalable Cache-Aware Optimistic Transactions

REST-Server

Coordinator

Client

REST-Server

Coordinator

Client

Begin Transaction

Bloom Filter1

REST-Server

Coordinator

Client

Begin Transaction

Bloom Filter1

Writes2

Writes

(Hidden)

REST-Server

Coordinator

Client

Begin Transaction

Bloom Filter1

Writes2

Commit: read- & write-set versions

Committed OR aborted + stale objects3

Writes

(Hidden)

REST-Server

Coordinator

Client

Begin Transaction

Bloom Filter1

Writes2

Writes

(Hidden)

validation 4

E.g. Redis or ZooKeeper

prevent conflicting

validations

REST-Server

Coordinator

Client

Begin Transaction

Bloom Filter1

Writes2

Writes

(Hidden)

validation 4

5Writes (Public)

Read all

prevent conflicting

validations

REST-Server

Coordinator

Client

Begin Transaction

Bloom Filter1

Writes2

Writes

(Hidden)

validation 4

5Writes (Public)

Read all

prevent conflicting

validations

Caching → Shorter transactionduration → less aborts

PolyglotPersistence

application

Orestes servers

REST/HTTPprotocol

Redis MongoDB db4o

meta data contains SLA

parse SLA & route data

manage materialisation

resolve mapping

Polyglot Persistence Mediator

PolyglotPersistence

application

Orestes servers

REST/HTTPprotocol

Redis MongoDB db4o

resolve mapping

PolyglotPersistence

application

Orestes servers

REST/HTTPprotocol

Redis MongoDB db4o

resolve mapping

PolyglotPersistence

application

Orestes servers

REST/HTTPprotocol

Redis MongoDB db4o

resolve mapping

Results:Article-Objects with Impression Count

Article

IDTitle…

Imp.ID

MongoDB Redis Sorted Set

Speedup with PPM:• 50-1000%• 66% performance of Varnish

Cloud Evaluation of ORESTES

Client Machine

50 ...

CacheOrestes

Server

Versant

Amazon EC2 Ireland EC2 USA165 ms

Client Machine

50 ...

CacheOrestes

Server

Versant

Client Machine

30 000 Objekte

500 Anfragen/

Client

30 000 Objects500 Req./Clients10/1 Read/Write

Client Machine

50 ...

CacheOrestes

Server

Versant

Client Machine

Stateless Scale-out REST middleware for DBaaS

Generic Schema, Authentication, Multi-Tenancy, etc.

SCOT (Scalable Optimistic Cache-Aware Transactions):◦ Optimistic Database-indepdent ACID transactions

BFB (Bloom Filter Bounded Staleness):◦ Allows static caching with tunable consistency guarantee

PPM (Polyglot Persistence Mediator):◦ SLAs appropriate database

Orestes: Summary

From Research toPractice

Orestes as a startup

Baqend

Internet

REST-API Transactions Schema Management Cache Consistency

Auto-Scaling Multi-Tenancy Security and Access Control Provisioning

Baqend

Baqend in Action

Baqend in ActionGET /app.html

db.find(Menu, 'main').done(...);

db.find(Page, 'hero').done(...);

db.query(Page, 'top3').done(...);

Baqend in ActionGET /app.html

db.find(Menu, 'main').done(...);

db.find(Page, 'hero').done(...);

db.query(Page, 'top3').done(...);

GET /img/pic005.jpgGET /img/pic017.jpgGET /img/pic022.jpg

More: Baqend.com

Wrap-up & book recommendations

How to choose a cloud database:

Wrap-up

Managed RDBMS

Managed DWH

Managed NoSQL DB

Proprietary Serivce

Object Store

Wrap-up

Define your functionalrequirements

Define your non-functionalrequirements

Managed RDBMS

Managed DWH

Managed NoSQL DB

Proprietary Serivce

Object Store

Wrap-up

Managed RDBMS

Managed DWH

Managed NoSQL DB

Proprietary Serivce

Object Store

1. Underyling DB2. Docs & books3. Your own tests

Evaluate by:

Wrap-up

Managed RDBMS

Managed DWH

Managed NoSQL DB

Proprietary Serivce

Object Store

4. SLAs5. Docs & books6. Experience Reports

Evaluate by:

Wrap-up

Managed RDBMS

Managed DWH

Managed NoSQL DB

Proprietary Serivce

Object Store

4. SLAs5. Docs & books6. Experience Reports

Evaluate by:

Try it

Book recommendations

• (Non-scientific) literature is rare• Some books cover specific DBaaS and

cloud platforms

http://nosql.mypopescu.com/

http://www.dzone.com/mz/nosql

http://www.dbms2.com/

http://hackingdistributed.com/

http://highscalability.com/

http://www.nosqlweekly.com/

VLDB (Very Large Databases)

SIGMOD (Special Interest Group on Management of Data)

ICDE (International Conference on Data Engineering)

CIDR (Conference on Innovative Data Systems Research)

SOCC (Symposium on Cloud Computing)

OSDI/SOSP (Operating Systems Design and

Implementation/ Symposium on Operating System Principles)

EuroSys

Top Scientific ConferencesD

atabase

Research

Distrib

System

esearch

VLDB (Very Large Databases)

SIGMOD (Special Interest Group on Management of Data)

ICDE (International Conference on Data Engineering)

CIDR (Conference on Innovative Data Systems Research)

SOCC (Symposium on Cloud Computing)

OSDI/SOSP (Operating Systems Design and

Implementation/ Symposium on Operating System Principles)

EuroSys

Top Scientific ConferencesD

atabase

Research

Distrib

System

esearch

This year probably in Washington D.C.Learn more: scdm2013.com

Thank you. Queries?

Contact:

felix.gessert@baqend.com

http://baqend.comhttp://orestes.infohttp://scdm2013.com

Cloud Databases in Research and Practice

Science