and Newcomer (1997) distinguish between:
A transaction-processing (TP)
system is the hardware and software that implements the transaction programs. A
TP monitor is a portion of a TP system that acts as a kind of funnel or
concentrator for transaction programs, connecting multiple clients to multiple
server programs (potentially accessing multiple data sources). In a distributed
system, a TP monitor will also optimize the use of the network and hardware
resources. Examples of TP monitors include
IBM’s Customer Information
Control System (CICS),
IBM’s Information Management System (IMS),
BEA’s Tuxedo, and
Microsoft Transaction Server
The focus of this article is on the fundamentals of online
transactions (e.g. the technical side of things). The critical concepts are:
Business transactions. A business
transaction is an interaction in the real world, usually between an enterprise
and a person, where something is exchanged.
Online transaction. An online transaction is the execution of a program that performs an
administrative or real-time function, often by accessing shared data sources,
usually on behalf of an online user (although some transactions are run offline
in batch). This transaction program
contains the steps involved in the business transaction.
This definition of an online transaction is important because it makes it
clear that there is far more to this topic than database transactions.
2. The ACID Properties
An important fundamental of transactions are the four
properties that they must exhibit:
- Atomicity. The whole transaction occurs or nothing in the
transaction occurs; there is no in between. In SQL, the changes become
permanent when a COMMIT statement is issued, and they are aborted when a
ROLLBACK statement is issued. For example, the transfer of funds between two
accounts is a transaction. If we transfer $20 from account A to account B,
then at the end of the transaction A’s balance will be $20 lower and B’s
balance will be $20 higher (if the transaction is completed) or neither
balance will have changed (if the transaction is aborted).
When the transaction starts the entities are in a consistent state, and when
the transaction ends the entities are once again in a consistent, albeit
different, state. The implication is that the referential integrity rules
and applicable business rules still apply after the transaction is
- Isolation. All transactions work as if they alone were
operating on the entities. For example, assume that a bank account contains
$200 and each of us is trying to withdraw $50. Regardless of the order of
the two transactions, at the end of them the account balance will be $100,
assuming that both transactions work. This is true even if both transactions
occur simultaneously. Without the isolation property two simultaneous
withdrawals of $50 could result in a balance of $150 (both transactions saw
a balance of $200 at the same time, so both wrote a new balance of $150).
Isolation is often referred to as serializability.
The entities are stored in a persistent media, such as a relational database
or file, so that if the system crashes the transactions are still permanent.
So far I have discussed flat transactions, transactions
whose steps are individual activities. A nested transaction is a transaction
where some of its steps are other transactions, referred to as subtransactions.
Nested transactions have several important features:
|As the name suggests there are two phases to the 2PC protocol: the
attempt phase where each system tries its part of the transaction and the commit
phase where the systems are told to persist the transaction.
The 2PC protocol requires the existence of a transaction manager to
coordinate the transaction. The transaction manager will assign a unique transaction ID
to the transaction to identify it. The
transaction manager then sends the various transaction steps to each system of
record so they may attempt them, each system responding back to the transaction
manager with the result of the attempt. If
an attempted step succeeds then at this point the system of record must lock the
appropriate entities and persist the potential changes in some manner (to ensure
durability) until the commit phase. Once the transaction manager hears back from all systems of
record that the steps succeeded, or once it hears back that a step failed, then
it either sends out a commit or abort request to every system involved.
When a program starts a new transaction, if it already
inside of an existing transaction then a subtransaction is started otherwise
a new top level transaction is started.
There does not need to be a limit on the depth of
When a subtransaction aborts then all of its steps are
undone, including any of its subtransactions.
However, this does not cause the abort of the parent transaction,
instead the parent transaction is simply notified of the abort.
When a subtransaction is executing the entities that it
is updating are not visible to other transactions or subtransactions (as per
the isolation property).
When a subtransaction commits then the updated entities
are made visible to other transactions and subtransactions.
Although transactions are often thought of as a database issue the reality
could be further from the truth. From
the introduction of TP monitors such as CICS and Tuxedo in the 1970s and 1980s,
to the CORBA-based object request brokers (ORBs) of the early 1990s to the EJB
application servers of the early 2000s transaction have clearly been far more
than a database issue. This section
explores three approaches to implementing transactions that involve both object
and relational technology. This
material is aimed at application developers as well as Agile DBAs that need to
explore strategies that they may not have run across in traditional
data-oriented literature. These implementation options are:
- Database transactions
- Object transactions
- Distributed object
- Including non-transactional steps
The simplest way for an application to implement transactions is
to use the features supplied by the database.
Transactions can be started, attempted, then committed or aborted via SQL
code. Better yet, database APIs
such as Java Database Connectivity (JDBC) and Open Database Connectivity (ODBC)
provide classes that support basic transactional functionality.
At the time of this writing support for transaction control is one of the
most pressing issues in the web services community and full support for nested
transactions is underway within the EJB community.
As you see in Figure 1, databases
aren’t the only things that can be involved in transactions. The fact is that objects, services, components, legacy
applications, and non-relational data sources can all be included in
Figure 1. Transactions can involve more
than just databases.
The advantage of adding behaviors implemented by objects (and
similarly services, components, and so on) to transactions are that they become
far more robust. Can you imagine
using a code editor, word processor, or drawing program without an undo
function? If not, then I believe it
becomes reasonable to expect both behavior invocation as well as data
transformations as steps of a transaction.
Unfortunately this strategy comes with a significant disadvantage –
increased complexity. For this to work your business objects need to be
transactionally aware. Any behavior
that can be invoked as a step in a transaction requires supporting attempt,
commit, and abort/rollback operations. Adding
support for object-based transactions is a non-trivial endeavor.
Just like it is possible to have distributed data transactions it
is possible to have distributed object transactions as well.
To be more accurate, as you see in Figure
1 it’s just distributed transactions period – it’s not just about
databases any more, but it’s databases plus objects plus services plus
components plus… and so on.
Sometimes you find that you need to include a
non-transactional source within a transaction.
A perfect example is an update to information contained in an LDAP
directory or the invocation of a web service, neither of which at the time of
this writing support transactions. The
problem is as soon as a step within a transaction is non-transactional the
transaction really isn’t a transaction any more. You have four basic strategies available to you for dealing
with this situation:
Which approach should you take? I prefer strategies #1 and #4 – when it comes to
transactions I want to do it right or not do it at all.
The problem with implementing full transactional logic is that it can be
a lot of work. I’ll consider the
attempt and abort strategy when it is possible to live with the results of a
collision, and strategy #2 as a last resort. A major issue is that
strategy #4 is the only one to pass the ACID test.
Remove the non-transactional step from your
transaction. In practice this is
rarely an option, but if it's a viable strategy then consider doing so.
This strategy, which could be thought of as the “hope the parent transaction
doesn’t abort” strategy, enables you to include a non-transactional step
within your transaction. You will need to simulate the attempt, commit, and
abort protocol used by the transaction manager. The attempt and abort
behaviors are simply stubs that do nothing other than implement the
requisite protocol logic. The one behavior that you do implement, the
commit, will invoke the non-transactional functionality that you want. A
different flavor of this approach, which I’ve never seen used in practice,
would put the logic in the attempt phase instead of the commit phase.
Implement attempt and abort.
This is an extension to the previous technique whereby you basically
implement the “do” and “undo” logic but not the commit. In this case, the
work is done in the attempt phase; the assumption is that the rest of the
transaction will work, but if it doesn’t, you still support the ability to
roll back the work. This is an “almost transaction” because it doesn’t avoid
the problems with collisions described earlier.
Make it transactional. With this approach, you fully implement the requisite
attempt, commit, and abort behaviors. The implication is that you will need
to implement all the logic to lock the affected resources and to recover
from any collisions. An example of this approach is supported by the
J2EE Connector Architecture
(JCA), in particular by the LocalTransaction interface.