Back to the Future with Oracle Database 12c

Back to the Future with Oracle Database 12c
An Oracle White Paper
April 2014
Back to the Future
with Oracle Database 12c
Back to the Future with Oracle Database 12c
Introduction ....................................................................................... 1
How Database Concepts Influence the Database Market .................. 2
Database concepts and fundamentals – a practical review ............ 3
Exemplary history: Cassandra, mongoDB & the Oracle RDBMS ... 7
The DBMS variety – a result & a choice, but not required .............. 8
How to Choose the Right DBMS for Your Business? ......................... 9
Operation (standard features and management) ........................... 9
Data Management (access and modification) .............................. 10
Migration (inter-solution data access and migration) .................... 11
Scalability (vertically and horizontally) ......................................... 13
High Availability (locally, remotely and long term) ........................ 14
Security (for data at rest and in transit) ........................................ 16
Conclusion ...................................................................................... 17
Appendix A – References ................................................................ 18
Back to the Future with Oracle Database 12c
The IT business evolves in circles with a tendency to form a spiral. Every once in a while an
idea or a technology moves back into the focus, which is not entirely new, but has evolved
since the first time it appeared.
In October 2013, Gartner published their latest Magic Quadrant (MQ) for Operational Database
Management Systems, formerly OLTP [Gartner 2013]. The number of players in the lower left
quadrant, the so-called “Niche Players”, may be surprising and may indicate that the IT
business is about to enter a new cycle introducing changes that could elevate and transform
the Operational Database Management Systems market significantly. It may, however, just
be a minor cycle leading to a broad but rather insignificant adjustment of the market.
Currently, mainly the increased number of Niche Players, amongst them names such as
Couchbase, VoltDB and mongoDB, surprises. All Niche Players are well behind the “Leaders”
considering their ability to execute and / or their vision according to Gartner. Current leaders
are IBM, Microsoft, SAP, and Oracle, which leads on both axes (ability to execute and vision).
Some of the players, mostly the Niche Players, only offer one specific product with a very
explicit set of features, while the Leaders can typically leverage a universal solution. The
latter is the key to success for those players and ensures the success of their customers.
This paper is intended for technically interested parties that are involved in the operation,
management as well as selection of a standard database management system. The paper
is split into two parts. Part one provides a practical review of some database concepts and
relates them to the current DBMS market as well as the history of three exemplary database
management systems. Part two looks into the reasons for the recent popularity of NoSQL
databases in order to determine their potential impact on the current DBMS market. It
discusses two exemplary NoSQL products and compares them to the solutions offered by
the leading RDBMS vendor: Oracle.
Back to the Future with Oracle Database 12c
How Database Concepts Influence the Database Market
In order to understand the rise of Niche Players in the Operational Database Management Systems
market as well as the recent increase in popularity of NoSQL (Not Only or No-Structured Query
Language) databases, one will have to review certain database concepts and see how they influenced
the current database market. The background and the focus of nearly all database management systems
(DBMS) available today is related to the requirements and technology that were available at the time
they were first invented. The following three examples of database management systems will be used
to illustrate this relation in the course of this discussion:
Apache Cassandra
Cassandra is an open-source project under the Apache Software Foundation license.
DataStax, which is part of Gartner’s MQ, offers professional services for Cassandra.
mongoDB is originally also an open-source database (see
 / mongoDB – the company – offers professional services
Oracle RDBMS
Oracle Relational Database Management System; current version: Oracle Database 12c
Apache Cassandra and mongoDB were chosen as examples for open-source database products. Both
offerings have found support by companies providing professional services for the respective solution.
Both database offerings also address “agility and scalability” as well as “fault tolerance, performance
and elasticity” in their own way, attributes that are typically also associated with the Oracle Database.
Both solutions represent a very different type of database compared to the Oracle Database, which is
an example of a relational database management system. Cassandra as well as mongoDB are nonACID, NoSQL, key-value pair databases using a sharding approach for horizontal scalability. In other
words, they are built on guiding principles that are the opposite of those guiding the Oracle RDBMS.
In “Cassandra’s words”: “On the contrary to the strong consistency used in most relational databases
(ACID for Atomicity Consistency Isolation Durability) Cassandra is at the other end of the spectrum (BASE
for Basically Available Soft-state Eventual consistency). Cassandra weak consistency comes in the form of
eventual consistency which means the database eventually reaches a consistent state.”1
Back to the Future with Oracle Database 12c
Database concepts and fundamentals – a practical review
In order to understand statements like the definition of “strong consistency” that Cassandra uses and
their practical impact, one needs to first understand the differences in architecture and technology that
are available and used for today’s DBMS solutions. For the purpose of this paper, the following three
concepts will be considered and discussed:
NoSQL databases vs. Relational Database Management Systems
The ACID consistency vs. the BASE consistency model
Shared Disk Clusters vs. Shared Nothing Clusters
The CAP theorem in the context of distributed database systems
Sharding as an evolution of the Shared Nothing approach
NoSQL databases vs. Relational Database Management Systems
According to Wikipedia, “a NoSQL database provides a mechanism for storage and retrieval of data
that is modeled in means other than the tabular relations used in relational databases.”2 Wikipedia also
mentions that “NoSQL databases are often highly optimized key–value stores intended primarily for
simple retrieval and appending operations, whereas an RDBMS is intended as a general purpose data
store.” In simple terms, this means, that the definition of NoSQL databases is based on the idea that
a NoSQL database uses a different approach than the “tabular relations” used by RDBMS.
The association of NoSQL databases with key–value stores, is implied, but in the absence of a NoSQL
standard not compelling. NoSQL databases can be implemented in many ways. currently lists over 150 NoSQL DBMS in more than ten categories. Those categories
include, but are not limited to: Wide Column Store / Column Families, Document Store, Key Value /
Tuple Store, Graph Databases, Multimodel Databases, Object Databases and quite a few more.
In addition, Wikipedia mentions that “NoSQL systems are also referred to as "Not only SQL" to
emphasize that they may in fact allow SQL-like query languages to be used.” An example for such a
DBMS is Cassandra, which provides CQL – the Cassandra Query Language.3 On the other hand, an
RDBMS can also provide a NoSQL-based way of accessing data. For example, VoltDB as well as the
Oracle RDBMS both support “SQL and the flexibility of JSON”4.
Wikipedia, the free encyclopedia, 20th March 2014, NoSQL:
3 The Cassandra Query Language:
4 For VoltDB’s support of SQL and JSON, see:; for Oracle Database 12c,
JSON support will be part of an upcoming release as discussed in this Oracle Open World 2013 Session:
“Storing and Querying JSON Data in Oracle Database 12c [CON9348]”, San Francisco, September 2013
Back to the Future with Oracle Database 12c
For the purpose of this discussion, the fundamental differences between NoSQL databases and
RDBMS do not represent a compelling theoretical reason to use either type. NoSQL databases are
typically “highly optimized key–value stores” and often provide a variety of ways for accessing data.
In contrast, RDBMS are typically designed as multi-purpose DBMS using the SQL standard to access
and modify the data. These differing access methods are in practice the most distinguishing aspects
that should be considered when choosing a standard (general purpose) DBMS.
The ACID consistency vs. the BASE consistency model
Another fundamental design difference distinguishing various DBMS implementations is their
consistency model. In computer science, two different consistency models are distinguished:
ACID: Atomicity, Consistency, Isolation, Durability
BASE: Basically Available Soft-state Eventual consistency
which basically describe contrary approaches to how reliable database transactions are processed by a
given DBMS. A DBMS that follows ACID design principles5 ensures the highest level of reliability for
database transactions, whereas DBMS that follow the BASE design principles follow what is often
referred to as an optimistic replication approach, which means that “reads eventually return the same
value”6 and the DBMS does not make safety guarantees:
An eventually consistent system can return any value before it converges. The verb “converge” in
this context refers to a particular aspect of the eventual consistency model pertaining to using this
consistency model in distributed computing. It basically describes an informal guarantee “that, if no
new updates are made to a given data item, eventually all accesses to that item will return the last
updated value… A system that has achieved eventual consistency is often said to have converged,
or achieved replica convergence.” [Petersen et. al. 1997)]
Most of the database management systems nowadays represent distributed DBMS since they use a
distributed approach for scaling. Hence, the consistency model is an important selection criterion
and should be considered carefully when choosing a standard DBMS.
Wikipedia, the free encyclopedia, 20th March 2014, ACID:
Wikipedia, the free encyclopedia, 20th March 2014, Eventual consistency:
Back to the Future with Oracle Database 12c
2a) The CAP theorem
In the context of discussing the consistency model, some DBMS vendors refer to the CAP theorem.
The CAP theorem by Brewer [Brewer 2000] basically says that a distributed shared-data system can at
the most have two of the following three properties: Consistency, Availability, and Partition Tolerance
Cassandra for example is very clear on their implementation of the CAP theorem: “Cassandra values
Availability and Partitioning tolerance (AP). Tradeoffs between consistency and latency are tunable in
Cassandra. You can get strong consistency with Cassandra (with an increased latency). But, you can't
get row locking: that is a definite win for HBase. Note: Hbase values Consistency and Partitioning
tolerance (CP)”7
The first publication of Brewer’s CAP theorem dates back to 2000, but he clarified some of his
positions in may 2012 [Brewer 2012], when he found that “In the decade since its introduction,
designers and researchers have used (and sometimes abused) the CAP theorem as a reason to
explore a wide variety of novel distributed systems.
The NoSQL movement also has applied it as an argument against traditional databases.” Brewer
clarified that “The "2 of 3" formulation was always misleading because it tended to oversimplify the
tensions among properties. Now such nuances matter. CAP prohibits only a tiny part of the design
space: perfect availability and consistency in the presence of partitions, which are rare.”
Practically, the nuances do not matter that much, since some DBMS vendors (especially, those,
classified as Niche Players by Gartner) still stick to the idea that choosing two properties is mostly
“sufficient”, whereas most of the RDBMS vendors (especially, those, classified as Leaders by Gartner)
have always striven to satisfy all three properties with their RDBMS products in various ways.
Shared Disk Clusters vs. Shared Nothing Clusters
One of the areas where the leading RDBMS vendors have traditionally been at odds is the way
storage is organized for the shared-data, distributed DBMS. In the ideal case, a client does not have
to be aware of the location of the data within the DBMS. As long as this assumption is fulfilled, it is
up to the DBMS (RDBMS or general DBMS) to decide on how the data is persisted on storage.
Traditionally, two approaches have been used by various DBMS: 1) Shared Disk 2) Shared Nothing
In a shared disk system, all servers that run a DBMS instance have (concurrent) access to all the data
on storage, whereas in case of a shared nothing system, the data is partitioned and each server that runs
an instance of the DBMS system accesses and manages only a subset of the data. Figure 1 illustrates
the differences between these two architectures from a managerial point of view.
Back to the Future with Oracle Database 12c
Stonebraker [Stonebraker 1986] was one of the first ones that mentioned and advocated the shared
nothing (SN) approach in his paper titled “The Case for Shared Nothing”. In very simple terms, he
concluded that the SN architecture “will have no apparent disadvantages compared to the other
alternatives. Hence the SN architecture adequately addresses the common case.” This basically means
that according to his study, the SN architecture is no worse than any other alternative, while he clearly
saw an availability advantage in the SN architecture.
De facto however, it has never incontrovertibly been proven that shared nothing always provides
better availability or scalability. As summarized by [Michalewicz 2003], it is the management layer of
the DBMS that can turn any DBMS into a potential single point of failure, not the shared disk based
approach by itself, at least not in case of a DBMS that also implements an ACID consistency model
and uses disk redundancy.
It is also the management layer and its ability to adapt to various use cases that mainly determines the
scalability of the DBMS, not the way the storage is organized for the shared-data, distributed DBMS.
Finally, it needs to be noted that in any system shown in figure 1 disk-protection (redundancy) is
required, as a failure of an unprotected disk would otherwise still lead to data loss. The amount of data
that is affected by the loss of a (single) disk and the impact on the overall system differs between the
three solutions. For the SN cluster, it may be assumed that there is a failover mechanism in place that
allows a server in the cluster to manage an additional subset of data should the primary server fail.
3a) Sharding as an evolution of the Shared Nothing approach
Quite a few DBMS vendors use a variation of the shared nothing approach called sharding. Sharding
combines two techniques: 1) Horizontal Partitioning & 2) Vertical Partitioning (illustrated in figure 1).
Figure 1: Shared Disk, Shared Nothing, and Sharding in comparison
Back to the Future with Oracle Database 12c
Unlike shared nothing, sharding allows for shards to be independent databases, while the data remains
to be partitioned. Sharding can therefore be considered an evolutionary step compared to shared
nothing, but introduces further administrative challenges which should be considered for production
databases. For example, if shards are distributed across data centers, the system as a whole needs to
be made aware of this separation, especially assuming that replication is used as illustrated in figure 1.
From the same illustration, it also becomes apparent why most shard-based DBMS implement a BASE
consistency model; it simplifies the recovery from partial failures in the system. On the other hand,
using a BASE consistency model will make it more difficult to perform a consistent backup across all
shards that constitute the sharded database.
From a practical point of view, sharding is simpler to implement as far as the DBMS is concerned and
will provide scalability for certain applications, such as read-mostly or strict key-value pair-based
applications. It also allows for flexible architectures, but requires a well thought-through system layout
as well as continuous maintenance and product-specific training.
Exemplary history: Cassandra, mongoDB & the Oracle RDBMS
The history of Cassandra, mongoDB and the Oracle RDBMS are exemplary in showing how the
requirements and technology that were available at the time when they were first invented influenced
the design. The Oracle RDBMS was designed to be a universal DBMS, while Cassandra and mongoDB
are purpose-built DBMS that were created to address certain data management requirements.
The history of the Oracle RDBMS is tightly coupled with the history of relational databases as first
introduced by E.F. Codd in his famous paper, titled “A Relational Model of Data for Large Shared
Data Banks”, published 1970 [Codd 1970]. The Oracle RDBMS was the first commercial
implementation of a Relational Database Management System:
“It all began when Larry Ellison saw an opportunity other companies missed: a description of a
working prototype for a relational database. No company had committed to commercializing the
technology, but Ellison and co-founders Bob Miner and Ed Oates realized the tremendous business
potential of the relational database model.”8
According to Oracle’s Interactive Timeline9, the first version of the “Oracle Database” was released in
1979: “Although RSI [author remark: Relational Software Inc., former company name] has supplied
several government agencies with early versions of its RDBMS product, the company releases its
relational database product, "Oracle," for the first time commercially in July [author remark: 1979].”
Oracle's History: Innovation, Leadership, Results:
Oracle Interactive Timeline:
Back to the Future with Oracle Database 12c
Cassandra and mongoDB have a far shorter history. According to DataStax, Cassandra “was created
for solving the problem of inbox search at Facebook. It combines Amazon Dynamo’s [DeCandia et. al.
2004] fully distributed design with Google Bigtable’s [Chang et. al. 2006] column-oriented data model.
Facebook open-sourced Cassandra in 2008 and it became an Apache Incubator project. In early 2010,
Cassandra became a top-level Apache project.”10
As Kristina Chodorow in her “Snail in a Turtleneck”-Blog11 describes, “mongoDB was created by the
founders of DoubleClick. Since leaving DoubleClick, they founded a number of startups and kept
running into the same scaling problems over and over. They decided to try to create an application
stack that would scale out easily, as companies everywhere seemed to be running into the same issues.”
mongoDB was originally developed as the database backend for “an application platform for the
cloud, similar to Google App Engine12”. “The problem was, no one cared about Google App Engine
and certainly no one cared about 10gen’s [the company that was subsequently founded by the founders
of DoubleClick] app engine. ... After a year of work and practically no users, we ripped the database
out of the app engine and open sourced them.”
The DBMS variety – a result & a choice, but not required
There is no database type that is simply the best one. Various DBMS are available on the market and
as history shows, the traditional RDBMS have been designed as universal database solutions and hence
can be used with a variety of applications and for various use cases.
Most of the upcoming NoSQL DBMS solutions were originally designed to address a certain data
management requirement, mostly driven by a specific application. For applications that are similar to
the one for which those solutions were designed, these DBMS represent a good solution. For other
applications they will fall short in certain areas, as their optimization for one use case almost always
means that compromises were made in other areas.
As the practical review of various database concepts shows, it is theoretically possible to solve most,
if not all, data management requirements with one type of DBMS depending on the capabilities of
DBMS layer used as part of a solution and hence, a rich variety of niche DBMS types is probably not
required to satisfy market requirements.
DATASTAX, The History of Cassandra,
Kristina Chodorow Blog:
12 Google App Engine:
Back to the Future with Oracle Database 12c
How to Choose the Right DBMS for Your Business?
Many technical details can be reviewed when it comes to choosing the right DBMS for your business.
However, one aspect will be the one influencing the decision the most: costs. The price points at
which most of the niche DBMS vendors offer a supported version of their solution is understandably
attractive at first glance, while the question remains whether they provide the lowest Total Cost of
Ownership (TCO) in the long term. In other words:”Do you want to pay now or later?”
Speaking to various customers, one once told Oracle that over the last year, they have deployed over
180 new NoSQL database projects, which means that this customer has deployed an average of 15
new NoSQL projects per month. The same customer, however, admitted later that if they had known
that there will be 180 new NoSQL DB installations after a year, they would have re-considered this
approach in the first place, as the number of NoSQL DBMS deployments gets unmanageable.
This is an example for the cost challenges associated with using purpose-built DBMS. In case of
traditional RDBMS the license and support costs are typically the main cost contributors to the total
cost of ownership. With NoSQL or purpose-built database solutions, the TCO is mostly influenced by
operational expenditures such as administration costs and secondary solution costs, such as additional
administrative tools and support that needs to be purchased to efficiently operate those databases.
The subsequent part of the paper will review the six most relevant areas that you should take into
consideration when choosing a standard DBMS in order to determine whether the DBMS solution
of choice will be able provide the lowest TCO for your data center in the long term.
Operation (standard features and management)
Cassandra, mongoDB and other niche solutions often do not provide features that are generally
considered “standard features” for traditional RDBMS, such as joins or aggregation. Providing such
functionality is not a strict requirement, but has proven to be useful, as the lack of such features will
lead to cost intensive workarounds. For example:
Storing similar data more than once.
Using additional software external to the database to provide missing functionality.
Running multiple queries to get a sub-set of data and letting the application perform the joins.
Similar holds true for administration. For production environments the installation, maintenance and
patching needs to be orchestrated and ideally automated. While automation is not a strict requirement,
it has proven to be useful, as the lack of such features will lead to cost intensive workarounds such as:
DBAs write installation and patching scripts.
Orchestration software is purchased to monitor and maintain the DB deployments.
Back to the Future with Oracle Database 12c
Part of the value proposition of most of the purpose-built DBMS solutions is that deployments are
rpm-based, making it simple and fast to deploy an initial database. Replication and replication groups
can be enabled and managed by the user to allow for maximum flexibility.
For a production environment this type of flexibility can be expensive, while the question arises of
whether or not the initial deployment time is relevant, as it should be an infrequent task. Configuring
and maintaining a cluster for scalability can be a rather challenging and ongoing task. It is therefore
desirable that the DBMS of choice provides an easy way of managing the system as a whole.
In a simple Cassandra cluster with six servers13 (clustered servers are referred to as “nodes”) across
two data centers for example, the setup not only foresees potential restarts and data clearance, it also
requires maintaining a configuration file (the Cassandra.yaml configuration file) in order to define the
nodes in the cluster as well as their roles (e.g. seed nodes). In a 6-node cluster, the Cassandra yaml
configuration file contains six of these node entries. In an n-node cluster, n entries need to be manually
maintained and distributed across n nodes. In addition, one topology file needs to be maintained.
Using Cassandra or mongoDB, this type of day-to-day management, including patching and upgrades,
has to be performed without the help of management tools, unless secondary solutions are purchased.
Using a standardized solution such as Oracle Real Application Clusters (RAC), standard management
tools are not only provided free of charge, they are fully cluster-aware and thereby simplify day-to-day
management tasks in clusters of any supported size.
Furthermore, certain management operations are not required in a shared disk-based RDBMS, such as
Oracle RAC, as they are inherent to using sharding as a scalability approach. For example, in case of a
Cassandra database, reorganization of the cluster or the structure of the shards require manual edits as
well as redistribution of configuration files along with a potentially manual (re-)distribution of binaries.
Depending on the use case for the database, those operations may become a regular task as explained
in the subsequent chapter of this paper. Using Oracle RAC, data is stored in a central fashion and will
almost never require physical reorganization or redistribution, lowering the TCO in the long term.
Data Management (access and modification)
From a development point of view, NoSQL databases are attractive, as they allow for a very flexible
and dynamic development style. Unlike relational databases, NoSQL databases do not require a
restrictive relational data model to be established prior to developing the application and populating
data. Key-value pairs are often “good enough” for a lot of specific use cases.
DataStax, Initializing a multiple node cluster (multiple data centers),
Back to the Future with Oracle Database 12c
However, “flexible” does not mean “simple”. When it comes to choosing a standard DBMS specifics
matter, since using a variety of different solutions that do not follow any standards otherwise increases
training requirements and costs. Data modeling and optimization as well as data storage requirements
should therefore be reviewed for each DBMS that is planned to be used in the data center.
As a side effect of using a key-value pair approach to storing data, one may have to de-normalize and
duplicate data for read performance, although one should not de-normalize, if not needed. “It’s all
about finding the right balance” as A Jay Patel in his eBay tech blog titled “Cassandra Data Modeling
Best Practices, Part 1” illustrates14. Finding this balance is a planning intensive task and requires
ongoing optimization, leading to higher operational costs.
As Jay Patel in his blog continues to explain, choosing an appropriate data model optimized for the
(anticipated) access pattern of the application that will be used with the database has a strong influence
on the efficiency and the performance that one can expect from the database. This especially applies to
Cassandra databases, which use a key-value pair approach to store data, on top of which a SQL-like
layer (CQL = Cassandra Query Language) can be used to access the data.
Cassandra supports indexes to look up data, but does not support joins. Ensuring good performance
for frequent queries based on data from related data sets may require keeping the data for each set of
data as well as the joined data. As a side effect, dynamic queries may often still not provide the best
performance, as the data may not have been optimized for them.
Migration (inter-solution data access and migration)
Another challenge in addition to the management costs associated with handling distributed NoSQL
databases will be the migration of projects that outgrow the capabilities of a given NoSQL database
over time. Developers often use NoSQL databases because they very efficiently solve a problem at
hand. Over time, however, these applications may become more complex or more important so that
those databases need to be integrated into an existing infrastructure that ensures certain Service Level
Agreements (SLAs). At this time, a migration to a database tailored to hosting business critical
applications may become required.
Accessing or moving data between different NoSQL databases can be challenging, if the data model
needs to be optimized depending on the specifics of the DBMS. Standardization has traditionally been
an attempt to simplify moving between databases. In the absence of a NoSQL standard to storing data,
JSON seems to be one of the upcoming document-based models that are frequently used to store data
in a format that will satisfy a lot of the “modern development requirements.”
Patel, Jay, Cassandra Data Modeling Best Practices, Part 1,
Back to the Future with Oracle Database 12c
Most of these modern development requirements are based on the idea to change the application data
model on demand, which is contrary to the idea of using schemas. Schema-less data management with
the ability to persist application objects in a document-centric approach becomes more and more
popular; so are the databases supporting this model.
mongoDB supports JSON, as well as CouchDB and Oracle’s NOSQL database. Starting with Oracle
Database 12c Patch Set One (, JSON will also be supported in the Oracle Database, which will
provide access to the JSON data using various APIs, including SQL, as it was announced during
Oracle Open World 2013, session [CON9348] - “Storing and Querying JSON Data in Oracle Database 12c”.
Using JSON as a document-centric standard for storing data has multiple advantages. It is a defined,
yet flexible data model that allows schema-less development. Creating indexes on documents has been
a widely supported feature for a variety of DBMS, although optimized indexes are still required. Last
but not least, a standardized way of storing data will simplify the migration.
Choosing a NoSQL database that supports JSON for storing data will therefore be useful when the
need arises to move data into the Oracle Database for example. There is now a choice to either retain
the JSON documents or convert the data model into a relational model as needed. Either case, one
will immediately benefit from the Oracle Database as an infrastructure.
One of the benefits that Oracle as a company offers in this context is a migration path between
different database solutions. As Andy Mendelsohn, Executive Vice President Oracle Database
Server Technologies, explains in his keynote during the NoSQL Now! 2013 Conference & Expo15,
Oracle offers a variety of standard database management solutions, tailored to various use cases as
illustrated in figure 2 below.
Figure 2: Oracle Databases - Providing the right database technology for the job
Mendelsohn, A., Oracle NoSQL Database,
Back to the Future with Oracle Database 12c
Additional value that Oracle as a DBMS (not just RDBMS) vendor provides is the integration of the
Oracle solutions with Oracle Enterprise Manager (EM) as the standard management tool as well as a
support organization that acts as the primary point of contact for all DBMS related questions. Oracle
Enterprise Manager enables monitoring and automation for any Oracle Database solutions free of
charge, while Oracle Support as the primary point of contact eliminates effortful multi-vendor calls in
case support for any DBMS is required. Consequently, Oracle as a solution provider should always be
considered when it comes to choosing any type of DBMS for your data center.
Scalability (vertically and horizontally)
Horizontal scalability is one of the main selling points for both Cassandra and mongoDB and both
solutions use a sharding approach for scalability. For read-mostly applications, sharding will work;
even for data warehouse applications with ETL-based data loads, sharding works as designed. For
OLTP applications alternative approaches provide a better performance.
When using sharding to scale horizontally, which means using more than one server in the data center
concurrently to run an operation, the question of vertical scalability inevitably arises, so do questions
regarding partially (un)available data.
Vertical and horizontal scalability
In sharded environments, the size of a shard is typically determined by a hash function or the data is
partitioned using a given partition key. The distribution of work across shards thereby depends either
on the system and the hash function or the structure of the data. Either case, the non-trivial question
of how many servers need to serve a shard will need to be answered.
If using more than one NoSQL database as part of the deployment in the data center, using servers as
shared resources and allocating databases from different vendors on the same server is one deployment
model that will soon become a necessity, raising the question on how to isolate the various databases
on the same server. Virtual machines are typically the logical answer. Virtual machines work fine for
this matter, but will increase the management overhead and resource utilization.
Sharding is not the simplest approach when maintaining databases; neither will it necessarily help to
reduce the data stored. Sharding assumes that data is partitioned over various data stores. This means
that the total amount of data that is stored as part of the databases is merely partitioned, not reduced.
If it is further assumed that each shard is replicated to prevent data loss in case of a (partial) failure, the
amount of data stored as part of the database doubles in the same way as if a shared disk approach
with mirroring was used.
Consequently, sharding does not solve all the problems. It certainly avoids some problems, but leaves
others unsolved or to the discretion of the user. The latter approach can lead to a significant amount
of planning as well as maintenance, thereby increasing management costs.
Back to the Future with Oracle Database 12c
Dealing with partially (un)available data
One of the biggest disadvantages of sharding (as mostly used in combination with a BASE consistency
model) is that the DBMS needs to provide a solution on how to deal with partially (un)available data.
Partially (un)available data generally appears in two ways: 1) a shard is not available (despite replication)
at all at the moment a query is run or 2) the data from a shard cannot be returned in a reasonable
amount of time. A subset of the partially available data problem is the “inconsistent data problem”,
which can occur in databases that have chosen to implement eventual consistency.
Eventual consistency (BASE) assumes that a database “converges” at some point in time. In the
meantime, however, it is not assured that all replicas will have the latest update of the modified data.
Querying those replicas during this time can lead to wrong results and therefore requires the
application to be sharding-aware, unless the sharded DBMS prevents reading from those replicas,
which typically leads to a decrease in performance.
Considering various sharding approaches and the configuration choices that need to be made to
optimize their performance (often referred to as “making the application sharding-aware”), it is fair to
say that if similar optimization is performed on a per-application basis in a shared disk environment,
the application would perform equally well, if not better.
Using shared disk, however, the application developer or administrator would typically not have to
think about “making the application sharding-aware”, “partially (un)available data”, “eventual
consistency” and the access pattern that the application will expose in order to optimize the data,
leading to reduced overall planning and managing costs.
High Availability (locally, remotely and long term)
Cassandra and mongoDB provide a defined level of high availability. The fundamental assumption,
however, is that high availability is based on partition tolerance. Both solutions also use clustering as
well as sharding to achieve availability. Clustering is used to coordinate roles in the cluster and to
determine availability of components. Both solutions thereby combine what is typically referred to
as protection against component failures and protection against data loss.
Protection against component failures
Cassandra explicitly mentions its failure tolerance and continuous availability in addition to scalability,
which in case of the Cassandra database is based on a token-ring cluster architecture. Cluster
configuration information is exchanged via special servers in the cluster called “seed nodes”. Cluster
membership information and updates are exchanged via the public network (by default; a dedicated
network is recommended) using the “gossip” protocol.
Back to the Future with Oracle Database 12c
This architecture is suitable for managing server membership in the Cassandra cluster, but limited
when it comes to managing consistency, node fencing or any other application. In addition, using a
database-specific clustering technology has various disadvantages and Oracle is the first to admit that
having database-specific clustering solutions are hard to justify, as this was the approach taken when
Oracle Real Application Clusters (RAC) 9i was first released more than ten years ago.
For a production environment the need to use various clusters on the same infrastructure will increase
the chattiness on the (public) network and increase the management and training costs as well as the
complexity of the system, since multiple specific cluster solutions need to be maintained individually.
Using different cluster models for availability and scalability also makes the migration between
different DBMS difficult, as one would have to consider the dependencies on the cluster for each
database management system. This complicates the integration into an existing infrastructure, as
applications using a DBMS might require a cluster that (partially) overlaps with the DBMS cluster
for example. Using a general purpose cluster for both layers simplifies this integration and thereby
allows for reducing the number of items that need to be managed to in the data center.
Oracle stopped using Oracle RAC-specific clustering solutions on Linux and Windows when Oracle
Clusterware 10g was release. Oracle Clusterware is a generic clustering solution that is available on all
Operating Systems supported for Oracle RAC. Oracle Clusterware serves as the integrated foundation
for running Oracle RAC16 databases, while it provides a cluster-based resource management solution
for any application. As part of the Oracle Grid Infrastructure, Oracle Clusterware can also be used to
manage MySQL databases in the cluster17.
Protection against data loss
Unlike Cassandra and mongoDB, traditional RDBMS typically use a separation between clustering and
protection against storage failure, following the distinction between local and remote high availability.
Local high availability is typically provided by a cluster in this case, while remote high availability is
achieved by a replication mechanism that typically does not require tight coupling between nodes either
on server- or on the storage-level. The replication mechanism that Cassandra uses follows a similar
paradigm by allowing hosting a replica in a different data center, but in order to make efficient use of
this replica, it needs to be in the same database cluster.
In case of an Oracle Database, the local database is typically protected by Oracle RAC and a remote
protection is provided by Oracle Active Data Guard. The assumption is that the shared disk-based
For more information on Oracle Real Application Clusters (RAC), see
For more information see the “Oracle Grid Infrastructure (Bundled) Agents” Reference Guide:
Back to the Future with Oracle Database 12c
Oracle RAC cluster uses a first level storage failure protection such as a local RAID implementation,
while the site synchronized via Oracle Active Data Guard would be used as a secondary storage failure
protection site, should the complete storage in the first site fail. In addition, Oracle Active Data Guard
can be used to outsource read-only queries and acts as a secondary protection, should all the servers in
the primary cluster fail at the same time.
The downside of separating the protection provided by the cluster and by the replication as in the
Oracle RAC / Oracle Active Data Guard example is that the data potentially needs to be maintained
three times (active set of shared disks, their mirror, and the Data Guard copy). Using parity-based
RAID implementations can help to reduce the amount of data that has to be stored.
Sharding, however, does not necessarily help to reduce the amount of data that needs to be stored on
disk. If the assumption is that high availability of the shard is maintained while re-organization takes
place, likeliness is that multiple copies of the sharded data need to be maintained. If full redundancy
is maintained for every shard that constitutes the database, no data reduction takes place at all, while
there is still additional overhead in maintaining this setup.
Security (for data at rest and in transit)
Most NoSQL databases are used for applications that work on “big data”, which means that there is
a “huge amount of” security needed to protect this data from un-authorized access. Unauthorized data
access can occur when the data is at rest (when being stored on disk) or while in transition (being read,
updated, written or while generally being transferred via the network).
Given that in sharded environments data is not only read on a frequent basis, but potentially also
transferred via the network due to re-organization or relocation needs, the DBMS should ideally
provide independent security features such as encryption for data at rest as well as for data in transit
(network encryption). For the latter, a standard solution can be used.
If local disks are used for storing data in shards, removing the data from those disks should also
consider providing an “erase” functionality to prevent un-authorized access to data after a subsequent
de-commission of those disks, especially if encryption functionality is not offered or used. While it
could be argued that this type of functionality is not necessarily provided by other, shared disk-based
solutions, either, it needs to be considered that these solutions by definition either assume enterpriselevel storage (as opposed to local storage), for which respective operational flows have already been
established in most data centers, or they provide a default encryption solution.
Similar considerations apply to other areas, which are often overlooked as they are taken for granted
using traditional multi-purpose RDBMS. Such areas cover, but are not limited to, backup and recovery
for example. Using the Oracle RDBMS as an example, backup and recovery has been a strong design
point right from the start. RMAN (the Oracle Recovery Manager) is the free-of-charge standard tool to
back up an Oracle Database in various states (online or offline). It is well integrated with third party
backup solutions, supporting useful features such as incremental backups as well as snapshot based
backups, of which all can be combined with security features inherent to the Oracle Database.
Back to the Future with Oracle Database 12c
The IT business evolves in circles with a tendency to form a spiral. This discussion shows that the
recent increase in the number of NoSQL databases, mostly considered Niche Players in Gartner’s
recent Magic Quadrant (MQ) for Operational Database Management Systems, formerly OLTP
[Gartner 2013], is no surprise. Most of these solutions are purpose-built DBMS that were developed
to address very specific use cases.
The popularity of some of those solutions and the support that they have received is a result of the
cost advantages they provide at first glance. For some, rather specific use cases, using those purposebuilt databases makes perfect sense.
However, NoSQL databases have not solved any problem that cannot be solved using traditional
RDBMS. This means, it is more likely that the currently leading RDBMS vendors will close the
perceived gaps in functionality leading to a broad but rather insignificant adjustment of the market.
The purpose-built nature of most NoSQL DBMS allows them to benefit from advances in other
technology areas, such as networking, faster. Traditional relational database management systems
have a longer history and need to maintain backward compatibility, hence, may need some time to
adopt new technologies, while they eventually will adapt accordingly.
Most of the leading RDBMS vendors listed in Gartner’s recent Magic Quadrant (MQ) have already
solved the problems that nearly all NoSQL databases will have to solve when hitting certain thresholds
as a result of growth. Growth and requirements of business critical applications will force almost all
NoSQL DBMS vendors to provide comprehensive solutions for data consistency, security and overall
management, cutting into the cost benefits they seem to provide at first glance.
Concluding, the leading relational DBMS providers will not rest to provide the same benefits that
NoSQL databases provide. They will add the additional benefits they already contain and offer the
integrated solution at a competitive price point. First in line: Oracle with its rich DBMS solution
portfolio, including its flagship solution, the Oracle Database 12c. Welcome to the future.
Back to the Future with Oracle Database 12c
Appendix A – References
[Brewer 2000] Brewer, Eric (2000), Towards Robust Distributed Systems:
[Brewer 2012] Brewer, Eric (2012), CAP Twelve Years Later: How the "Rules" Have Changed:
[Chang et. al. 2006] Fay Chang, Jeffrey Dean, Sanjay Ghemawat,Wilson C.Hsieh,
Deborah A.Wallach, Mike Burrows, Tushar Chandra, Andrew Fikes, Robert E.Gruber (2006),
Bigtable: A Distributed Storage System for Structured Data:
[Codd 1970] Codd, E.F. (1970), A Relational Model of Data for Large Shared Data Banks,
[DeCandia et. al. 2007] Giuseppe DeCandia, Deniz Hastorun, Madan Jampani, Gunavardhan Kakulapati,
Avinash Lakshman, Alex Pilchin, Swami Sivasubramanian, Peter Vosshall and Werner Vogels, “Dynamo:
Amazon's Highly Available Key-Value Store”, in the Proceedings of the 21st ACM Symposium on Operating Systems
Principles, Stevenson, WA, (2007):
[Gartner 2013] Gartner (October 2013), Magic Quadrant for Operational Database Management Systems:
[Michalewicz 2003] Michalewicz, M. (2003): Eigenschaften von Hochleistungsdatenbanken (im Cluster) vor
dem Hintergrund der Verfügbarkeit am Beispiel der Oracle9i Datenbank mit Oracle Real Application Clusters,
Datenbank-Spektrum 7: pages 28-37
[Petersen et. al. 1997] Petersen, K.; Spreitzer, M. J.; Terry, D. B.; Theimer, M. M.; Demers, A. J.
(1997), "Flexible update propagation for weakly consistent replication". ACM SIGOPS Operating Systems Review
31 (5): 288. doi:10.1145/269005.266711
[Stonebraker 1986] Stonebraker, M. (1986), The Case for Shared Nothing:
Back to the Future with Oracle Database 12c
Copyright © 2014, Oracle and/or its affiliates. All rights reserved.
April 2014
Author: Markus Michalewicz
This document is provided for information purposes only, and the contents hereof are subject to change without notice. This
Contributing Authors: Bob Thome
document is not warranted to be error-free, nor subject to any other warranties or conditions, whether expressed orally or implied in
Oracle Corporation
World Headquarters
500 Oracle Parkway
Redwood Shores, CA 94065
law, including implied warranties and conditions of merchantability or fitness for a particular purpose. We specifically disclaim any
liability with respect to this document, and no contractual obligations are formed either directly or indirectly by this document. This
document may not be reproduced or transmitted in any form or by any means, electronic or mechanical, for any purpose, without our
prior written permission.
Oracle and Java are registered trademarks of Oracle and/or its affiliates. Other names may be trademarks of their respective owners.
Worldwide Inquiries:
Phone: +1.650.506.7000
Intel and Intel Xeon are trademarks or registered trademarks of Intel Corporation. All SPARC trademarks are used under license and
Fax: +1.650.506.7200
are trademarks or registered trademarks of SPARC International, Inc. AMD, Opteron, the AMD logo, and the AMD Opteron logo are
trademarks or registered trademarks of Advanced Micro Devices. UNIX is a registered trademark of The Open Group. 0113
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF