International Journal in Foundations of Computer Science & Technology (IJ FCST), Vol.4, No.

4, July 2014

DOI:10.5121/ijfcst.2014.4405 63

REAL-TIME EVENTUAL CONSISTENCY

Jan Lindstrm
1
and Mawahib Musa Elbushra
2

1
SkySQL The MariaDB Compary, Espoo, Finland
2
Sudan University of Science & Technology, College of Graduate Studies, Khartoum,
Sudan

ABSTRACT

Many real-time systems are naturally distributed and these distributed systems require not only high-
availability but also timely execution of transactions. Consequently, eventual consistency, a weaker type of
strong consistency is an attractive choice for a consistency level. Unfortunately, standard eventual
consistency, does not contain any real-time considerations. In this paper we have extended eventual
consistency with real-time constraints and this we call real-time eventual consistency. Followed by this new
definition we have proposed a method that follows this new definition. We present a new algorithm using
revision diagrams and fork-join data in a real-time distributed environment and we show that the proposed
method solves the problem.

KEYWORDS

Real-time databases, eventual consistency, transaction processing.

1. INTRODUCTION

A real-time database system (RTDBS) is a database systemintended to handle workloads whose
state is constantly changing [9, 27]. Traditional databases contain persistent data, mostly
unaffected by time in contrast. As an example consider application keeping track of all of the
markets in a stock market that changes very rapidly and is dynamic. A real-time database is a
database where transactions have deadlines or timing constraints [30]. Real-time
databases are usually used in real-time applications which require timely access to data. Different
applications have different timeliness requirements; frommicroseconds to minutes [33].

Many real-world applications contain tasks that require time-constrained access to data as well as
access to data item that has temporal validity [26]. As an example: a telephone switching system,
network management, navigation systems, stock trading, and command and control systems. As
an example following task require time-constrained access: looking up the 800 directory,
obstacle detection and avoidance, radar tracking and recognition of objects. Tasks in these real-
time applications collect data from the dynamic environment, process data gathered earlier, and
produce timely and correct response. Additionally, these real-time tasks process both temporal
data, which loses its validity after a specific time intervals, as well as historical data.

Traditional databases, from now on referred to as databases, manipulate persistent data.
Transactions process this data and preserve its consistency. The goal of transaction and query
processing methods chosen in databases is to optimizethroughput or response time. Real-time
database systems in contrary can also process temporal data, i.e., data that becomes outdated after
a certain time. Task in real-time systems have timing constraint because application environment
contains data that is outdated in certain time periods. Therefore, real-time tasks use periods or
deadlines. These requirement leads to a new goal where real-time systems has to meet the time
constraints of the tasks.
International Journal in Foundations of Computer Science & Technology (IJ FCST), Vol.4, No.4, July 2014


64
Termreal-time does not just mean fast [33] and real-time does not mean timing constraints that
are in nanoseconds or microseconds. Real-time imply the need to manage explicit time
constraints in a predictable fashion using time-cognizant methods to deal with time or periodicity
constraints associated with tasks. Databases are necessary in real-time applications because they
include many useful features that allow (1) the description of data, (2) the maintenance of
correctness and integrity of the data, (3) efficient access to the data, and (4) the correct executions
of query and transaction execution in spite of concurrency and failures [29].

Many complex real-time systems require distribution and sharing of extensive amounts of data,
with full or partial replication of database. Consider for example integrated vehicle control system
were autonomous nodes control individual sub-systems and handle lot of data locally under hard
timing constraints (e.g., fuel injection, ignition control). These systems also require parameters
from other subsystems on a less critical time scale (e.g. transmission, environment sensors).
Similarly, in automated manufacturing, stock trading and, massively multiplayer online role-
playing games (MMORPG). In these application scenarios, temporary inconsistencies of the
distributed database can be allowed. Therefore, updates are made locally and propagated to all
notes in a way that guarantees eventual consistency [37, 38]. To achieve high scalability at low
cost, distributed databases are typically highly distributed systems running on commodity
hardware. Scale out requires adding a new off the shelf server.

However, Brewer [3,7] introduced CAP theorem proposing that it is impossible for a distributed
computer systemto concurrently provide all three of the following guarantees:

Consistency : all nodes see the same data at the same time.
Availability: a guarantee that every request receives a response about whether it was
successful or failed.
Partition tolerance: the system continues to operate despite lost messages or partial
failure of the system.
This, theoremis also known as Brewer's theorem[7, 3] and it was later formally shown by Gilbert
and Lynch [16].

This arouses the concern on which property is the least important one, since all three properties
are both important and anticipated in distributed systems [3, 8]. Many applications are often used
together with databases where the so called ACID (Atomicity, Consistency, Isolation, and
Durability) properties apply and thus strong consistency is required. Databases use transactions to
provide strong consistency.

High availability is expected because systems are connected to a network and if the network is
available, the system is also supposed to be available [18]. Furthermore, tolerance to network
partitions is also expected because it is a basic fault-tolerance technique. Distributed system
should continue service even if a network link goes down or a server crashes [16].

Network partitioning can be totally avoided or make very unlikely using a several servers on the
same rack. This solution is not useful for cloud systembecause it does not scale and decreases the
tolerance against other failure types. Use of more reliable links between the networks does not
eliminate the chance of partitioning, in contrary increases the cost significantly. Thus, network
partitioning will happen and systemdesigner has to select either consistency or availability.

Consistency requirement is similar to the atomicity property in ACID where every single
transaction will be atomic in strictly consistent database. Different server can have different views
of the same data itemif the database is not strongly consistent.

Distributed system should be designed to allow arbitrarily loss of messages sent from one server
to another to avoid partition tolerance [16]. Storing all data at one server is not an option because
International Journal in Foundations of Computer Science & Technology (IJ FCST), Vol.4, No.4, July 2014


65
of strict availability requirement. This is because when the only source fails the entire system
becomes unavailable. Thus, partition tolerance enables for system states to be kept in different
locations. Distributed database that is partition tolerant allows to performread/write operations
while partitioned. If distributed databaseis not partition tolerant, when partitioned, the database
may become completely unusable or only available even for read operations.

Therefore, maintaining a single-system service in a distributed systemhas a cost as CAP theorem
shows [3]. Updates cannot be synchronously propagated to all servers without blocking when
processes within a distributed systemare partitioned. This blocking presents unavailability to
some or all of its users. High availability system can offer low latency even when network is
portioned, because it can respond to a users request without contacting other servers [3].
Allowing partitioning is not an option as noted on [18, thus choice is either consistency or
availability.

Brewer [8] noted that two of the three requirements can be guaranteed simultaneously if one of
the requirements can be traded-off. It could be possible to have every kind of combination, CP,
AP and CA. However, every system wants to be tolerant to partitions. Therefore, only open
question in CAP is what do we do when some servers are not available?! Choosing CA would be
a system that does not tolerate network partitioning and making system unavailable. CP or AP are
only choices in the distributed databases. Therefore, most of the NOSQL-databases provide only
eventual consistency, a weaker type of strong consistency.

Eventual consistency states that in an updatable replicated database, eventually all copies of each
data itemconverge to the same value. The origin of eventual consistency can be traced back to
Thomas majority consensus algorithm [36]. Eventual consistency was also used by Terry et al.
[35] and later made more popular by Amazon in their Dynamo system, which supported only
eventual consistency [13].

Eventual consistency is a weaker form a consistency compared to strong consistency; the storage
systemguarantees that if no new updates are made to the object, eventually all accesses will
return the last updated value [5]. When failures does not happen, the maximum size of the
inconsistency window can be calculated based on communication delays, the load on the system,
and the number of servers participating in the replication. The most popular system that uses
eventual consistency is DNS (Domain Name System). Update to a name is distributed based on
order and time intervals configured; eventually, all client applications will see the latest name
[37]. When an applications fetches a data item there is no guarantee that the data accessed is
consistent, therefore the conflicts have to be resolved in application [37].

In this paper our major motivation is to use eventual consistency as a consistency level in a real-
time distributed relational database system. We make following contributions in this paper:

To authors knowledge this is the first paper were real-time eventual consistency is defined
and used on real-time database systems.
While lacy replication is used on DeeDS and this lacy replication is sometimes called as
eventual consistent it has not been shown that DeeDS can maintain eventual consistency on
all configurations and all failure cases. Furthermore, traditional eventual consistency does
not provide any real-time constraints.
Proposed method allows multiple-masters and requires that transactions is successfully
executed on all servers before transaction deadline.
We will rigorously show that proposed real-time eventual consistency method produces
correct results.
International Journal in Foundations of Computer Science & Technology (IJ FCST), Vol.4, No.4, July 2014


66
The rest of this paper is organized as follows. We first review the related research in Section 2.
This is followed by definition of real-time eventual consistency in Section 3. A new real-time
eventually consistent method is presented in Section 4. Finally, conclusions are presented in
Section 5.

2. RELATED WORK

Most of the previous work on real-time databases has been based on simulation. However, several
prototypes of real-time databases have been introduced. Real-time disk-based transaction
processing systemRT-CARAT [20], was one of the firsts [20]. Similarly, REACH (Real-time
Active and Heterogeneous mediator systemproject) [10] and the STRIP (Stanford Real-time
Information Processor) project [1].

Kimand Son [22] have presented StarBase real-time database architecture. StarBase architecture
has been developed over a real-time microkernel operating systemand it is based on relational
model. Wolfe \& al. [39] have implemented a prototype of object-oriented real-time database
architecture RTSORAC. RTSORAC architecture is based on open object-oriented database
architecture with real-time extensions. Database is implemented over a thread-based POSIX-
compliant operating system. Similarly, DeeDS project at the University of Skvde [2, 15, 17]
and the BeeHive project at the University of Virginia [34] are examples of more recent RTDBS
prototype projects.

M2RTSS main-memory object oriented database systemarchitecture is presented by Cha \& al.
[12]. M2RTSS is object-oriented storage manager implementing real-time transaction
scheduling, and recovery. Real-Time Object-Oriented Database Architecture for Intelligent
Networks (RODAIN) [23], is a real-time, object-oriented, and fault-tolerant database
management system. RODAIN is a main-memory real-time database using priority and criticality
based scheduling and optimistic concurrency control.

Concurrently, commercial real-time'' database systemproducts have started to appear in the
marked such as Eaglespeed-RTDB [14], Clustra [21], Timesten [24], and SolidDB [4, 28]. While
these products are not real-time databases based on strict definition of real-time since most of
them only have very limited real-time features, they represent a very important step forward to
the success of real-time database systems. In these commercial database management systems
many use in-memory database techniques to achieve a better real-time performance.

Distributed real-time database systems have been also studied on literature e.g. [19, 20, 25, 27,
32, 34]. To authors knowledge DeeDS is the only distributed real-time database system
mentioning eventual consistency. Therefore, we review DeeDS more detailed way.

The DeeDS [2, 15, 17] database prototype stores the fully replicated database in main memory to
make transaction timeliness independent of disk accesses. Because, DeeDS uses full replication
transactions can be executed on the local node entirely, independent of network delays. Updates
are send to other servers using detached replication independently fromthe execution of the
transaction. Clients using the databasesees the distributed database as a single local database and
does need not to handle data location or how to synchronize concurrent updates of different
replicas of the same data. Full replication uses a lot of resources, since all updates needs to be
replicated to all the servers. Additionally, this causes scalability problemfor resource usage, such
as bandwidth for replication of updates, storage for data item replicas, and processing for
replicating updates and resolving conflicts for concurrent updates at different servers for the same
data item. Currently, real-time databases do not scale well but in same time the need for larger
distributed real-time databases is increasing.
International Journal in Foundations of Computer Science & Technology (IJ FCST), Vol.4, No.4, July 2014


67
The DeeDS architecture supports multi-master replication. The replication scheme adopted in
DeeDS (referred to as lazy replication) is used in order to make real-time database access
predictable [2]. That is, by replicating all data onto every node and accepting updates on every
node, an application will never have to access data over the network. Instead it can always access
and change data locally. Further, since eventual consistency is employed, updates can be made on
different nodes simultaneously without any need to lock the entire database, only the local
replica. When an update has committed locally, it is propagated to the other nodes. Thus, an
application can always apply any changes necessary locally, without waiting for the remote
replicas. This implies that the database may become temporarily globally inconsistent, and that
conflicts may occur. Any conflict that occurs must be detected and resolved. Local consistency is
enforced by pessimistic locking of the local replica of the database.

Replication in DeeDS is handled by the replication module. The replication module interacts with
the storage manager TDBM, and with the DeeDS Operating system Interface (DOI). TDBM is a
store manager with support for nested transactions [6]. The DOI is a layer added between the
DeeDS database and the underlying hardware which makes it possible to make the DeeDS
platformindependent [11]. DeeDS replication module and architecture is shown in the Figure 1.



Figure 1. Replication in DeeDS [3].

The replication module logs all updates on data made during a local transaction. When the local
transaction has been committed, the log is propagated to all other DeeDS nodes. The changes are
then integrated into the remote replica(s). This is currently implemented in DeeDS by using a
number of modules: the logger, the propagator, the integrator, the version vector handler (VV
handler) and the log filter (depicted in above figure). The logger is responsible for logging any
changes made during a local transaction. When a transaction has been committed, the log is
forwarded to the propagator module. The propagator distributes the update message containing
the log(s) to all other replicas of the database. Currently, the log is distributed over TCP via point
to point connections.

A more efficient solution would be to use broadcast or multicast. On the remote nodes, the update
message is received by the integrator module. The integrator module consists of several sub
modules, these are: the receiver, the conflict detector, the conflict resolver and the updater. The
receiver sub module is responsible for receiving update messages sent by propagators on other
DeeDS nodes. The conflict detector then checks if the update made is in conflict with any earlier
updates on this local node. This is done by using version vectors2 and the log filter. Three
different types of conflicts may occur: write-write conflicts, read-write conflicts, and read-write
cycles [2].
International Journal in Foundations of Computer Science & Technology (IJ FCST), Vol.4, No.4, July 2014


68
If a global conflict is detected, the conflict resolver has to resolve the conflict. To achieve
predictability, DeeDS detects and resolves conflicts locally, without communication between
replicas over the network. If a conflict is detected, it is vital that it is resolved within a predictable
amount of time in order to achieve predictable real-time database access, and in a deterministic
way for global consistency. That is, if two or more replicas are in conflict, all replicas must
resolve the conflict in the same way. DeeDS support a set of simple generic replication policies.

The chosen value may be:

based on the new value resulting froman operation, e.g., mean min or max value of two
conflicting updates
chosen fromthe highest prioritized replica
based on timestamps (e.g., use the oldest or newest value)

If no conflict has been detected or when any detected conflict has been resolved, the updater
writes the changes to the local copy of the database, and to the log filter. This must be done as
one atomic action. During the integration process, the involved database objects and the log filter
must be locked so that no other process manipulates them when the conflict detector is working.
If the local database and the log filter are not kept mutually consistent, conflicts may be
undetected with potential system failure as the result. Integration is done by using a special kind
of transactions, integration transactions which run on a lower priority than regular database
transactions. This makes sure that local access to the database does not need to wait an
unpredictable amount of time for any integration processing. Thus, local real-time guarantees are
kept.

These design choices makes real-time database accesses in DeeDS predictable since all accesses
are made in local main memory. Further, the database is guaranteed to be globally consistent
eventually since all changes are propagated to other nodes, where conflicts are detected and
resolved in a predictable and deterministic way.

3. REAL-TIME EVENTUAL CONSISTENCY

In this section we first define necessary terminology.

DEFINITION 1: Real-time database data itemcan be denoted by d:(value, read timestamp,
write timestamp, absolute validity interval), where d{value} denotes the current state of the d,
read timestamp denotes when the last committed transaction has read the state of the d, write
timestamp denotes when the last committed transaction has written the d, i.e., when the
observation relating to a d was made, and absolute validity interval, denotes the length of the time
interval following read timestamp during which d is considered to have absolute validity.

A set of data items used to derive a new data itemform a relative consistency set R. Each such set
R is associated with a relative validity interval. Assume that d R. Data itemd has a correct state
if and only [30]:

d{value} is logically consistent, i.e., satisfies all integrity constraints.
d is temporally consistent:
Data item d R is absolutely consistent iff {current time} - d{observation time} <
d{absolute validity interval}.
Data items are relatively consistent iff {d' R | d{timestamp} d{timestamp} | <
R{relative validity interval}}.


International Journal in Foundations of Computer Science & Technology (IJ FCST), Vol.4, No.4, July 2014


69
DEFINITION 2: Real-Time Eventual consistency.

Real-time ordering: If update conflicts, a higher priority transaction wins and lower
priority transaction is aborted.
Real-Time Eventual delivery: An update executed at node before its deadline evenly
executes at all nodes.
Real-Time Termination: All update executions terminate before their deadlines or they
are aborted.
Convergence: Nodes that have executed the same updates eventually reach equivalent
state (and stay).

EXAMPLE 1: Consider a case where data itemR=0 on all three nodes. Assume that we have
following sequence of writes and commits: W(R=3) C W(R=5) C W(R=7) C in node 0. Now read
on node 1 could return R=3 and read fromnode 2 could return R=5. This is eventually consistent
as long as eventually read from all nodes return the same value. Note that that this value could be
R=3. Eventual consistency does not restrict the order which the writes must be executed.

4. PROPOSED METHOD

Proposed real-time eventually consistent distributed commit protocol is now presented.
Presentation of the proposed real-time eventually consistent method start from presenting how
real-time transactions are executed on master (or primary) server and then how changes are
replicated to other servers in the system.

Read only real-time transactions can be executed on any of the servers reachable by the client.
Because database state does not change on read only real-time transaction, there is no effect on
database consistency. Thus, global transaction identifier is not assigned to the real-only
transactions. Instead, only a local real-time transaction identifier is used. Because, full replication
is used in this work, single database server can execute the transaction.

Write transactions change database state and they are executed on master first and when master
has committed the real-time transaction, we replicate it to other servers for execution. Any node
that can receive messages from client can be used as master but normally the server replying
fastest is used as master.

Locally, data consistency is ensured by using a multiversion model (Multiversion Concurrency
Control, MVCC). For each transaction a snapshot of committed data is created when accessed.
This snapshot of data is consistent and transaction cant view inconsistent data produced by other
concurrent transaction updates on the same data rows, providing transaction isolation for each
database session. MVCC provides less lock contention in order to allow for reasonable
performance in multiuser environments. Furthermore, standard write-a-head logging is used to
guarantee durability.

1 GTID :=Source_id: +next_transaction_id;
2 If GTID is stored in masters log {
3 End; /* already executed */
4 }
5 While every statement Sk in the transaction T {
6 If Sk is contains at least one write statement {
7 If #servers reachable frommaster <(#total servers / 2) +1 { /* Majority?*/
8 Abort transaction T;
9 }
International Journal in Foundations of Computer Science & Technology (IJ FCST), Vol.4, No.4, July 2014


70
10 }
11 fork {
12 revision r ={ tuple }
13 If Sk is read {
14 Execute operation;
15 } Else If Sk is Write {
16 call conflict_detection(T, T);
17 If revision =current version {
18 Execute transaction;
19 } Else {
20 request current version
21 }
22 If database inconsistent {
23 Abort;
24 }
25 }; /* Sk is write */
26 }; /* while executes statements on master*/
27 J oin r;}
28 If (deadline <current_time) abort;
29 Write transaction operations and GTID to the log;
30 Send transaction redo log and GTID to other server for execution;

Figure 1. Master server algorithm.

Real-time transaction is executed firstly on the master server, the master server assigns real-time
transaction a global transaction identifier GTID. Since global transaction identifier contains the
servers UUID, it remains constant for transactions executed on that server. Global transaction
identifier is represented as a pair where GTID =source_id:transaction_id, and transaction_id is
globally unique. Possible method to implement globally unique transaction identifiers is to use
Lamport's clocks.

Global transaction identifiers are persistently stored to the ARIES style redo log using a new log
event called Gtid_log. Gtid_log record is written before a group of operations for a given
transaction is to be written to the log. Storing global transaction identifiers in log of the master
server makes sure that no real-time transaction is re-executed more than once and two different
real-time transactions cannot have the same global transaction identifier.

When master receives a transaction for execution we first check that global transaction identifier
is unique in a log file, which means global transaction identifier is not found from the log. If
global transaction identifier is found, we have already executed the real-time transaction and we
can skip this real-time transaction.

Transactions containing INSERT, UPDATE or DELETE statements are write transactions and
master server must be able to send messages and receive replies frommajority of the servers.
When real-time transaction is a read operation the transaction will find a snapshot of the data
using a fork operation that allows creating a copy of the data itemand read it. For real-time
transactions containing INSERT, UDPATE or DELETE statements, the proposed method finds a
snapshot of data itemusing fork. Proposed method validates that updated data itemis not old
snapshot or version of data item by equation: current version equal to new version minus one.
Then we check integrity constraints to guarantee that the new update not violate integrity
constraints.
International Journal in Foundations of Computer Science & Technology (IJ FCST), Vol.4, No.4, July 2014


71
After all changes has been done, we check that real-time constraint is not violated using
transaction deadline and current time.

Finally we use join all operation to merge these updates with main version and then global
transaction identifier is written to the log (immediately preceding the real-time transaction itself
in the log). Send transaction log and global transaction identifier to other servers, this means
other server cant see this new data itemversion until committed in the master server.

Now we consider how real-time transactions are executed on a server that is not a master. Either,
real-time transaction is read-only transaction or transaction is already committed on a master
server and now replicated to this server for execution. For read only real-time transactions we
only need to provide current consistent version of the data item. There is no need to use any
special method for this. Additional processing is required for write real-time transactions received
from master and this processing is presented below.

We can execute read-only real-time transactions on any of the servers, but transactions containing
write statements (update, insert or delete SQL-statements) must to be first executed on a master
server. Thus, servers are not autonomous i.e. master makes the decision whether the transaction
commits or aborts. If transaction is aborted on a non-master server normal REDO-log records of
the transaction are requested from the master server and that server executes ARIES style
recovery for that transaction.

Every server removes transaction from queue after it checks global transaction identifier is not
already executed on this server. If it is, we can skip the transaction. This is followed by validation
that this transaction based on its global transaction identifier is next transaction based on already
executed transactions in the log file. If not, this server sends a REDO log request to master server
and master server sends REDO logs for all missing transactions. Finally server writes transaction
operations and global transaction identifier to the log.

When executing write real-time transactions we need to check that data item versions are
consistent. They must represent consistent view. Data itemversions must represent consistent
view and integrity constraints should not be violated after transaction execution, if they do we
send REDO log request to the master server. Master server will provide REDO logs for all data
items in a read and write set of the transaction. These REDO logs are then executed on this server
using traditional ARIES style recovery. After transaction commit database is consistent
(eventuall) database state and the transaction log records and global transaction identifier can be
written to the log file followed by log record for transaction commit.

1 Remove transaction T fromQueue;
2 If gtid received can be found fromlog file{
3 End; /* Skip */
4 }
5 If transaction_id >transaction_id +1 on log {
6 Put transaction T back to Queue; /* Ordered execution */
7 };
8 For every statement Sk in transaction T {
9 Check data items versions touched on Sk are consistent;
10 Execute the statement;
11 If database inconsistent {
12 Abort transaction;
13 Send REDO log request to Master;
14 End;
International Journal in Foundations of Computer Science & Technology (IJ FCST), Vol.4, No.4, July 2014


72
15 }
16 } /* End For;
17 Write transaction operations and GTID to the log;

Figure 2. Other servers.

Finally conflict detection is done as follows:

1. If t, t TH and t conflicts with t then
2. If f t
priority
>=t
priority
then
3. t
v
t
4. Else t
v
t.
5. EndIf
6. Endif

Figure 3. Conflict detection.

We have to first define few precise definitions in order to show that proposed method produces
correct consistency level and we do this similarly as in [5]. Firstly, all operations must be part of a
transaction and transactions can be described using following three types of operations (query-
update interface):

1 Updates u U issued by the transactions
2 Pairs (q, p) forma query q Q that is executed by the transaction together with a result v
V fromthe database system.
3 Abort or Commit issued by the transaction.
Formally, we can represent the activity as a streamof operations forming a history.

DEFINITION 3: A history H for a set T transactions and query-update interface (Q, V, U) is a
map H which maps each transaction t T to a finite or infinite sequence H (t) operation from
alphabet =U (Q x V){end}.

Next we define the order where operations are executed on a transaction i.e. programorder.

DEFINITION 4: Program order. For a given history H, there is a partial order p over
operations in the H such that e p e iff e appear before e in some sequence of the H (t).

Then we define an equivalence relation.

DEFINITION 5: Factoring: We define an equivalence relation _t over events such that e t e
iff trans (e) =trans (e'). For any partial order over operations, we say that factors over t iff
for any operations x and y fromdifferent transactions x y implies x' y' for any x, y such that x
t x' and y t y. Thus, factoring forms a partial order on the transactions.

With following formalization we can specify the information about relationships between events
declaratively, without referring to implementation-level concepts, such as replicas or messages.
Namely, F takes as a parameter not a sequence, but an operation context, which encapsulates all
we need to know about a systemexecution to determine the return value of a given operation.
Eventual consistency weakens strict consistency models by allowing queries in a transaction t to
see only a subset of all transactions that are globally ordered before t. This is done by separating a
visibility order (a partial order that defines what updates are visible to a query), and an arbitration
order (a partial order that determines the relative order of updates).
International Journal in Foundations of Computer Science & Technology (IJ FCST), Vol.4, No.4, July 2014


73
DEFINITION 6: A history H is real-time eventually consistent if there exist two partial orders
v (the visibility order) and a (the arbitration order) over events in H, such that the following
conditions are satisfied for all events e1, e2, e EH:

1. Arbitration extends visibility: if e1v e2 then e1a e2.
2. Total order on past events: if e1v e and e2v e, then either e1a e2 or e2a e1.
3. Compatible with program order: if e1p e2 then e1v e2.
4. Consistent query results: for all (q, v) EH, v=q^#(apply ({e H)e v q},a,s0)).
Thus query returns the state as it results from applying all preceding visible updates (as
determined by the visibility order) to the initial state, in the order given by the arbitration
order.
5. Atomicity: Both v and a factor over t.
6. Isolation: If e1committed (EH) and e1v e2, then e1p e2. That is, events in
uncommitted transactions are visible only to later events by the same client.
7. Real-Time Eventual delivery: For all committed transactions t, there exists only finitely
many transactions t' TH such that t v t' and tdeadline <=commit_time.
8. Real-Time Ordering: If t, t TH and t conflicts with t then t v t if tpriority >=
tpriority otherwise t v t. Thus, real-time ordering forms a partial ordering of
transactions.

The reason why eventual consistency can tolerate temporary network partitions is that the
arbitration order can be constructed incrementally, i.e. may remain only partially determined for
some time after a transaction commits. Thus, conflicting updates can be committed even in the
presence of network partitions.

THEOREM 1: A history H produced by the proposed method is real-time eventually consistent.
PROOF:

Compatible with program order: All transaction operations are executed on all servers in
the order they appear on program in step 9. Thus, proposed real-time eventually
consistent method is compatible with programorder.
Arbitration extents visibility: In step 2 we make sure that slaves execute transactions in
the same order as they are executed on master. Local concurrency control uses MVCC
that produces locally serializable histories, thus we know that revision diagramis acyclic
and a partial order. Because data items versions use revision diagramfork and join
operations (see [6]) v is acyclic, transitive and partial order. Thus we know that if e1 e
and e2 e, then either e1 e2 or e2 e1. Now define v =a =, then if e1v e2 then
e1a e2.
Consistent query results: In step 2 query can continue if and only if all preceding
transactions are executed. Result of the query contains state as it results from applying all
preceding visible updates.
Atomicity: Step 2 of the proposed real-time method produce total order i.e. if e1 e and
e2 e, thus either e1 e2 or e2 e1. Now define v =a =, then we can easily see
that both v and a factor over t
Isolation: MVCC that is used as local concurrency control method (steps 20--24)
produces partial ordering (local serializability) and MVCC also maintains program order,
thus if e1 committed (EH) and e1v e2, then e1p e2. Thus, updates produced by
uncommitted transactions are visible only to later events by the same client.
Real-Time Eventual Delivery: For all executed transactions t at step 28 we check that
transaction is executed before its deadline, thus tdeadline <=commit_time. When server
crashes or there is network failure between servers, the server is unavailable. After server
is restored or server can again send and receive messages fromrest of servers, we
International Journal in Foundations of Computer Science & Technology (IJ FCST), Vol.4, No.4, July 2014


74
requests log records of the missing transactions. ARIES method is then used to process
received log records and thus eventually all operations are delivered.
When server is up it can serve only read-only transactions if the server does not belong to
majority. If there is no network partitioning, write transactions are executed only after the
missing transactions are executed (step 2). Thus, there can exist only finitely many
transactions t' TH such that t v' for every transaction t.
Real-Time Ordering: In conflict detection step 16 and Figure 3 we make sure that t, t
TH and t conflicts with t then t v t if tpriority >=tpriority otherwise t v t.

5. CONCLUSIONS

Many real-time systems are naturally distributed and these distributed systems require not only
high-availability but also timely execution of transactions. CAP summarizes trade-offs from
decades of distributed-system designs and shows that maintaining a single-system image in a
distributed system has a cost. Therefore, distributed databases can either be strongly consistent or
available. Consequently, eventual consistency, a weaker type of strong consistency is an attractive
choice for a consistency level. Unfortunately, standard eventual consistency, does not contain any
real-time considerations. In this paper we have extended eventual consistency with real-time
constraints and this we call real-time eventual consistency. Followed by this new definition we
have proposed a method that follows this new definition.

As a future research we will build a prototype real-time database system using real-time eventual
consistency. Another research question is how to allow more flexibility and additional rules on
both conflict resolution and recovery of transactions. Furthermore, the question what is the
strongest consistency level that can be provided to transactions and still avoid problems with CAP
theorem on bounded error scenario in network.

REFERENCES

[1] ADELBERG, B., Kao, B., Garcia-Molina, H.: Overview of the Stanford Real-Time Information
Processor STRIP. SIGMOD Record, vol 25, no 1, pp 34-37, 1996.
[2] ANDLER, S., Hansson, J., Mellin, J., Eriksson, J., and Eftring, B.: An overview of the DeeDS real-
time database architecture. In Proc. 6th IntlWorkshop on Parallel and Distributed Real-Time
Systems, 1998.
[3] BAILIS, P., and Ghodsi, A.: Eventual consistency today: limitations, extensions, and beyond , In
communications of the ACM vol. 56, no. 5, PP. 55-63, May 2013.
[4] BALLARD, C., Bechman, D., Huumonen, A., Laiho, K., Lindstrm, J., Milek, M., Roche, M., Seery,
J., Vakkila, K., Watters, J., Wolski, A.: IBM SolidDB: Delivering Data with Extreme Speed. An IBM
Redbooks publications, 2011.
[5] BURSKHARDT, S., Leijen, D., Fhndrich, M., and Sagiv, M.: Eventually Consistent Transactions.
ESOP , pp.67-86, 2012
[6] BRACHMAN, B., Neufeld, G.: TDBM: A DBM Library with Atomic Transactions, In Proc.
USENIX, San Antonio, 1992
[7] BREWER, E.: Towards Robust Distributed Systems, (invited Talk) Principles of Distributed
Computing, Portland, Oregon, SIGOPS, And SIGACT News, July 2000.
[8] BREWER, E.: CAP twelve years later: How the rules have changed. IEEE Computer, vol. 45, no.
2, pp. 23-29, February 2012.
[9] BUCHMANN, A.: Real Time Database Systems. Encyclopedia of Database Technologies and
Applications. Idea Group, 2002.
[10] BUCHMANN, A., Branding, H., Kudrass, T., Zimmermann, J.: REACH: A Real-Time, Active and
Heterogenous mediator system. IEEE Data Engineering Bulleting, vol 15, no 1, pp 44-47, 1992.
[11] CASAIS, E., Ranft, M., Schiefer, B., Theobald, D., Zimmer, W.: STONE An Overview,
ForschungszentrumInformatik (FZI), Germany, 1992.
International Journal in Foundations of Computer Science & Technology (IJ FCST), Vol.4, No.4, July 2014


75
[12] CHA, S., Park, B., Lee, S.,Song, S, Park, J., Lee, J., Park, S., Hur, D., Kim, G.: Object-Oriented
Design of Main-Memory DBMS for Real-Time Applications. In Proceedings of the 2nd International
Workshop on Real-Time Computing Systems and Applications, pp 109-115, 1995.
[13] DECANDIA, G., Hastorun, D., Jampani, M., Kakulapati, G., Lakshman, A., Pilchin, A.,
Sivasubramanian, S., Vosshall, P., and Vogels, W.: Dynamo: Amazon's highly available key-value
store. In Proceeding 21st ACM Symposiumon Operating Systems Principles (SOSP), pp. 205-220,
2007.
[14] EAGLESPEED: Lockheed Martin Corporation.
[15] ERIKSSON, D.: How to implement Bounded-Delay replication in DeeDS, B.Sc. dissertation,
Department of Computer Science, Hgskolan i Skvde, 1998.
[16] GILBERT, S., LYNCH, N.: Brewer's Conjecture and the Feasibility of Consistent, Available,
Partition-Tolerant Web Service; SIGACT News, vol 33, no 2, pp 51-59, June 2002.
[17] GUSTAVSSON, S., Andler, S.: Continuous Consistency Management in Distributed Real-Time
Databases with Multiple Writers of Replicated Data. Workshop on Parallel and Distributed Real-Time
Systems (WPDRTS'05, 2005.
[18] HALE, C.: You cant sacrifice partition tolerance; Available from http://codahale.com/you-cant-
sacrificepartition-tolerance.
[19] HARITSA, J., Ramamritham, K., Gupta, R.: Real-Time Commit Prosessing. Real-Time Database
Systems Architecture and Techniques, pp 227-244, Kluwer, 2001.
[20] HUANG. J., Stankovic, J., Towsley, D., Ramamritham, K.: Experimental Evaluation of Real-Time
Transaction Prosessing. In Proceedings of the 10th IEEE Real-Time Systems Symposium, pp 144-
153, 1989.
[21] HVASSHOVD, S.-V., Torbjrsen, ., Bratsberg, S. Holager, P.: The ClustRa Telecom Database:
High Availability, High Throughput, and Real-Time Response. In Proceedings of the 21th VLDB
Conference, pp 469-477, 1995.
[22] KIM, Y.-K., Sang, H.: Developing a Real-Time Dabase: The StarBase Experience. Real-Time
Database Systems: Issues and Applications, pp 305-324, Kluwer, 1997.
[23] KIVINIEMI, J., Niklander, T., Porkka, P, Raatikainen, K.: Transaction Prosessing in the RODAIN
Real-Time Database System. Real-Time Database and Information Systems, pp 355-375, Kluwer,
1997.
[24] LAHIRI, T., Neimat, M.-A., Folkman, S.: Oracle TimesTen: An In-Memory Database for Enterpise
Applications. IEEE Data Engineering Bulletin, vol 36, no 2, pp 6-14, 2013.
[25] LIN, K., Lin, M.-J.: Enhancing Availability in Distributed Real-Time Databases. ACM SIGMOD
Record, vol 17, no 1, pp 34-43, 1988.
[26] LOCKE, D.: Applications and System Characteristics. Real-Time Database Systems Architecture
and Techiques. Kluwer Academic Publishers, pp 17-26, 2001.
[27] LINDSTRM, J.: Real Time Database Systems. Encyclopedia of Computer Science and Engineering.
Vol 4, pp 2257-2369, John Wiley & Sons, January 2009.
[28] LINDSTRM, J., Raatikka, V., Ruuth, J., Soini, P., Vakkila, K.: IBM SolidDB: In-Memory Database
Optimized for Extreme Speed and Availability. IEEE Data Engineering Bulletin, vol 36, no 2, pp 14-
20, 2013.
[29] PURIMETLA, B., Sivasankaran, R., Ramamritham, K., Stankovic, J.: Real-Time Databases: Issues
and Applications. Advances in Real-Time Systems, pp 487-507, Prentice Hall, 1996.
[30] RAMAMRITHAM, K.: Real-Time Databases. Distributed and Parallel Databases, vol 1, pp. 199-226,
April, 1993.
[31] SIVASUBRAMANIAN, S.: Amazon dynamoDB: a seamlessly scalable non-relational database
service. In proceeding SIGMOD International Conference on Management of Data, pp. 729-730,
2011.
[32] SHANKER, U., Misra, M., Sarje, A.: Distributed real time database systems: background and
literature review. Distributed and Parallel Databases, vol. 23, no. 2, pp 127-149, 2008.
[33] STANKOVIC, J., Son. S., Hansson, J.: Misconceptions about real-time databases. IEEE Computer,
vol 32, no. 6, pp 29-36, June 1999.
[34] STANKOVIC. J ., Son, S., Liebeherr, J.: BeeHive: Global Multimedia Database Support for
Dependable Real-Time Applications. In proceedings of the Second International Workshop on Real-
Time Databases, pp 409-422, 1997.
[35] TERRY, D. B., Demers, A. J., Petersen, K., Spreitzer, M.J., Theimer, M.M., Welch, B. B.: Session
guarantees for Weakly Consistent Replicated Data. In proceeding of parallel and distributed
information system(PDIS). pp. 140-149, 1994.
International Journal in Foundations of Computer Science & Technology (IJ FCST), Vol.4, No.4, July 2014


76
[36] THOMAS, R.: A Majority Concensus Approach to Concurrency Control for Multiple Copy
Databases. ACM Transactions on Database Systems, vol 4, no 2,pp 180-209, 1979.
[37] VOGELS, W.: Scalable Web services: Eventually Consistent, ACM Queue, vol. 6, no. 6, pp 14-16,
October 2009.
[38] VOGELS, W.: Eventually consistent, Communications of the ACM, vol. 52, no. 1, pp. 4044,
January 2009.
[39] WOLFE, V., DiPippo, L, Prichard, J, Peckham, J., Fortier, P.: The Design of Real-Time Extensions to
the Open Object-Oriented Database System. Technical Report TR-94-236. University of Thode
Island, Department of Computer Science and Statistics, February 1994.

Authors

Dr. Jan Lindstrmis the principal engineer at SkySQL working on InnoDB storage
engine and Galera cluster. Before joining SkySQL he was software developer for
IBM DB2 and development manager for IBM solidDB core development. He joined
IBM with the acquisition of Solid Information Technology in 2008. Before joining
Solid in 2006, Jan worked on Innobase and spent almost 10 years working in the
database field as a researcher, developer, author, and educator. He has developed
experimental database systems, and has authored, or co-authored, a number of
research papers. His research interests include real-time databases, in-memory
databases, distributed databases, transaction processing and concurrency control. Jan has a MSc. and Ph.D.
in Computer Science fromthe University of Helsinki, Finland.

Mawahib Elbushra received her MSc on Computer Science fromthe College of Graduate Studies, Sudan
University of Science &Technology. She is currently aiming PhD on Computer Science. Her research
interests include cloud databases and distributed databases.