Availability is the fraction of time a TP system is up and running and able to do useful work — that is, it isn't...
down due to hardware or software failures, operator errors, preventative maintenance, power failures, or the like. Availability is an important measure of the capability of a TP system because the TP application usually is offering a service that's "mission critical," one that's essential to the operation of the enterprise, such as airline reservations, managing checking accounts in a bank, processing stock transactions in a stock exchange, or offering a retail storefront on the Internet. Obviously, if this type of system is unavailable, the business stops operating. Therefore, the system must operate nearly all the time.
Just how highly available does a system have to be? We see from the table in Figure 1.11 that if the system is available 96% of the time, that means it's down nearly an hour a day. That's too much time for many types of businesses, which would consider 96% availability to be unacceptable.
An availability of 99% means that the system is down about 100 minutes per week (i.e., 7 days/week X 24 hours/day X 60 minutes/hour X 1/100). Many TP applications would find this unacceptable if it came in one 100-minute period of unavailability. It might be tolerable, provided that it comes in short outages of just a few minutes at a time. But in many cases, even this may not be tolerable, for example in the operation of a stock exchange where short periods of downtime can produce big financial losses.
For more information on this title
This is an excerpt from Principles of Transaction Processing by Philip Bernstein and Eric Newcomer. Printed with permission from Morgan Kaufmann, a division of Elsevier. Copyright 2009.
Print Book ISBN : 9781558606234
eBook ISBN : 9780080948416
An availability of 99.9% means that the system is down for about an hour per month, or under two minutes per day. Further, 99.999% availability means that the system is down five minutes a year. That number may seem incredibly ambitious, but it is attainable; telephone systems typically have that level of availability. People sometimes talk about availability in terms of the number of 9s that are attained; for example, "five 9s" means 99.999% available.
Figure 1.11 Downtime at Different Availability Level. The number of nines after the decimal point is of practical significance.
|1 hour/20 years||99.99942|
Some systems need to operate for only part of the day, such as 9 AM to 5 PM on weekdays. In that case, availability usually is measured relative to the hours when the system is expected to be operational. Thus, 99.9% availability means that it is down at most 2.4 minutes per week (i.e., 40 hours/week X 60 minutes/hour X 1/1000).
Today 's TP system customers typically expect availability levels of at least 99%, although it certainly depends on how much money they're willing to spend. Generally, attaining high availability requires attention to four factors:
- The environment — making the physical environment more robust to avoid failures of power, communications, air conditioning, and the like
- System management — avoiding failures due to operational errors by system managers and vendors ' field service
- Hardware — having redundant hardware, so that if some component fails, the system can immediately and automatically replace it with another component that's ready to take over
- Software — improving the reliability of software and ensuring it can automatically and quickly recover after a failure
This book is about software, and regrettably, of the four factors, software is the major contributor to availability problems. Software failures can be divided into three categories: application failures, database system failures, and operating system failures.
Because we're using transactions, when an application fails, any uncommitted transaction it was executing aborts automatically. Its updates are backed out, because of the atomicity property. There's really nothing that the system has to do other than re-execute the transaction after the application is running again.
When the database system fails, all the uncommitted transactions that were accessing the database system at that time have to abort, because their updates may be lost during the database system failure. A system management component of the operating system, database system, or transactional middleware has to detect the failure of the database system and tell the database system to reinitialize itself. During the reinitialization process, the database system backs out the updates of all the transactions that were active at the time of the failure, thereby getting the database into a clean state, where it contains the results only of committed transactions.
A failure of the operating system requires it to reboot. All programs, applications, and database systems executing at the time of failure are now dead. Everything has to be reinitialized after the operating system reboots. On an ordinary computer system all this normally takes between several minutes and an hour, depending on how big the system is, how many transactions were active at the time of failure, how long it takes to back out the uncommitted transactions, how efficient the initialization program is, and so on. Very high availability systems, such as those intended to be available in excess of 99%, typically are designed for very fast recovery. Even when they fail, they are down only for a very short time. They usually use some form of replicated processing to get this fast recovery. When one component fails, they quickly delegate processing work to a copy of the component that is ready and waiting to pick up the load.
The transaction abstraction helps the programmer quite a bit in attaining high availability, because the system is able to recover into a clean state by aborting transactions. And it can continue from where it left off by rerunning transactions that aborted as a result of the failure. Without the transaction abstraction, the recovery program would have to be application-specific. It would have to analyze the state of the database at the time of the failure to figure out what work to undo and what to rerun. We discuss high availability issues and techniques in more detail in Chapter 7, and replication technology in Chapter 9.
In addition to application, database system, and operating system failures, operator errors are a major contributor to unplanned downtime. Many of these errors can be attributed to system management software that is hard to understand and use. If the software is difficult to tune, upgrade, or operate, then operators make mistakes. The ideal system management software is fully automated and requires no human intervention for such routine activities.
1.7 Styles of systems
We've been talking about TP as a style of application, one that runs short transaction programs that access a shared database. TP is also a style of system, a way of configuring software components to do the type of work required by a TP application. It's useful to compare this style of system with other styles that you may be familiar with, to see where the differences are and why TP systems are constructed differently from the others. There are several other kinds of systems that we can look at here:
- Batch processing systems, where you submit a job and later receive output in the form of a file
- Real-time systems, where you submit requests to do a small amount of work that has to be done before some very early deadline
- Data warehouse systems, where reporting programs and ad hoc queries access data that is integrated from multiple data sources
Designing a system to perform one of these types of processing is called system engineering. Rather than engineering a specific component, such as an operating system or a database system, you engineer an integrated system by combining different kinds of components to perform a certain type of work. Often, systems are engineered to handle multiple styles, but for the purposes of comparing and contrasting the different styles, we'll discuss them as if each type of system were running in a separately engineered environment. Let's look at requirements for each of these styles of computing and see how they compare to a TP system.
Batch processing systems
A batch is a set of requests that are processed together, often long after the requests were submitted. Data processing systems of the 1960s and early 1970s were primarily batch processing systems. Today, batch workloads are still with us. But instead of running them on systems dedicated for batch processing, they often execute on systems that also run a TP workload. TP systems can execute the batches during nonpeak periods, since the batch workload has flexible response-time requirements. To make the comparison between TP and batch clear, we will compare a TP system running a pure TP workload against a classical batch system running a pure batch workload, even though mixtures of the two are now commonplace.
A batch processing system executes each batch as a sequence of transactions, one transaction at a time. Since transactions execute serially there's no problem with serializability. By contrast, in a TP system many transactions can execute at the same time, and so the system has extra work to ensure serializability.
For example, computing the value of a stock market portfolio could be done as a batch application, running once a day after the close of financial markets. Computing a monthly bill for telephone customers could be a batch application, running daily for a different subset of the customer base each day. Generating tax reporting documents could be a batch application executed once per quarter or once per year.
The main performance measure of batch processing is throughput, that is, the amount of work done per unit of time. Response time is less important. A batch could take minutes, hours, or even days to execute. By contrast, TP systems have important response time requirements, because generally there's a user waiting at a display for the transaction's output.
A classical batch processing application takes its input as a record-oriented file whose records represent a sequence of request messages. Its output is also normally stored in a file. By contrast, TP systems typically have large networks of display devices for capturing requests and displaying results.
Batch processing can be optimized by ordering the input requests consistently with the order of the data in the database. For example, if the requests correspond to giving airline mileage credit for recent flights to mileage award customers, the records of customer flights can be ordered by mileage award account number. That way, it's easy and efficient to process the records by a merge procedure that reads the mileage award account database in account number order. By contrast, TP requests come in a random order. Because of the fast response time requirement, the system can't spend time sorting the input in an order consistent with the database. It has to be able to access the data randomly, in the order in which the data is requested.
Classical batch processing takes the request message file and existing database file(s) as input and produces a new master output database as a result of running transactions for the requests. If the batch processing program should fail, there's no harm done because the input fi le and input database are unmodified — simply throw out the output file and run the batch program again. By contrast, a TP system updates its database on-line as requests arrive. So a failure may leave the database in an inconsistent state, because it contains the results of uncompleted transactions. This atomicity problem for transactions in a TP environment doesn't exist in a batch environment.
Finally, in batch the load on the system is fixed and predictable, so the system can be engineered for that load. For example, you can schedule the system to run the batch at a given time and set aside sufficient capacity to do it, because you know exactly what the load is going to be. By contrast, a TP load generally varies during the day. There are peak periods when there's a lot of activity and slow periods when there's very little. The system has to be sized to handle the peak load and also designed to make use of extra capacity during slack periods.
TP systems are similar to real-time systems, such as a system collecting input from a satellite or controlling a factory's shop floor equipment. TP essentially is a kind of real-time system, with a real-time response time demand of 1 to 2 seconds. It responds to a real-world process consisting of end-users interacting with display devices, which communicate with application programs accessing a shared database. So not surprisingly, there are many similarities between the two kinds of systems.
Real -time systems and TP systems both have predictable loads with periodic peaks. Real-time systems usually emphasize gathering input rather than processing it, whereas TP systems generally do both.
Due to the variety of real-world processes they control, real-time systems generally have to deal with more specialized devices than TP, such as laboratory equipment, factory shop floor equipment, or sensors and control systems in an automobile or airplane.
Real -time systems generally don't need or use special mechanisms for atomicity and durability. They simply process the input as quickly as they can. If they lose some of that input, they ignore the loss and keep on running. To see why, consider the example of a system that collects input from a monitoring satellite. It's not good if the system misses some of the data coming in. But the system certainly can't stop operating to go back to fix things up like a TP system would do — the data keeps coming in and the system must do its best to continue processing it. By contrast, a TP environment can generally stop accepting input for a short time or can buffer the input for awhile. If there is a failure, it can stop collecting input, run a recovery procedure, and then resume processing input. Thus, the fault-tolerance requirements between the two types of systems are rather different.
Real -time systems are generally not concerned with serializability. In most real-time applications, processing of input messages involves no access to shared data. Since the processing of two different inputs does not affect each other, even if they're processed concurrently, they'll behave like a serial execution. No special mechanisms, such as locking, are needed. When processing real-time inputs to shared data, the notion of serializability is as relevant as it is to TP. However, in this case, real-time applications generally make direct use of low-level synchronization primitives for mutual exclusion, rather than relying on a general-purpose synchronization mechanism that is hidden behind the transaction abstraction.
Data warehouse systems
TP systems process the data in its raw state as it arrives. Data warehouse systems integrate data from multiple sources into a database suitable for querying.
For example, a distribution company decides each year how to allocate its marketing and advertising budget. It uses a TP system to process sales orders that includes the type and value of each order. The customer database tells each customer's location, annual revenue, and growth rate. The finance database includes cost and income information, and tells which product lines are most profitable. The company pulls data from these three data sources into a data warehouse. Business analysts can query the data warehouse to determine how best to allocate promotional resources.
Data warehouse systems execute two kinds of workloads: a batch workload to extract data from the sources, cleaning the data to reconcile discrepancies between them, transforming the data into a common shape that's convenient for querying, and loading it into the warehouse; and queries against the warehouse, which can range from short interactive requests to complex analyses that generate large reports. Both of these workloads are quite different than TP, which consists of short updates and queries. Also unlike TP, a data warehouse's content can be somewhat out-of-date, since users are looking for trends that are not much affected by the very latest updates. In fact, sometimes it's important to run on a static database copy, so that the results of successive queries are comparable. Running queries on a data warehouse rather than a TP database is also helpful for performance reasons, since data warehouse queries would slow down update transactions, a topic we'll discuss in some detail in Chapter 6. Our comparison of system styles so far is summarized in Figure 1.12 .
Other system types
Two other system types that are related to TP are timesharing and client-server.
In a timesharing system, a display device is connected to an operating system process, and within that process the user can invoke programs that interact frequently with the display. Before the widespread use of PCs, when timesharing systems were popular, TP systems often were confused with timesharing, because they both involve managing lots of display devices connected to a common server. But they're really quite different in terms of load, performance requirements, and availability requirements:
- A timesharing system has a highly unpredictable load, since users continually make different demands on the system. By comparison, a TP load is very regular, running similar load patterns every day.
- Timesharing systems have less stringent availability and atomicity requirements than TP systems. The TP concept of ACID execution doesn't apply.
- Timesharing applications are not mission-critical to the same degree as TP applications and therefore have weaker availability requirements.
- Timesharing system performance is measured in terms of system capacity, such as instructions per second and number of on-line users. Unlike TP, there are no generally accepted benchmarks that accurately represent the behavior of a wide range of timesharing applications.
Figure 1.12 Comparison of System Types. Transaction processing has different characteristics than the other styles, and therefore requires systems that are specially engineered to the purpose.
|Transaction Processing||Batch||Real-time||Data Warehouse|
|Isolation||serializable, multiprogrammed execution||serial, uniprogrammed execution||no transaction concept||no transaction concept|
|Workload||high variance||predictable||predictability depends on the application||predictable loading, high variance queries|
|Performance metric||response time and throughput||throughput||response time, throughput, missed deadlines||throughput for loading, response time for queries|
|Input||network of display devices submitting requests||record-oriented file||network of devices submitting data and operations||network of display devices submitting queries|
|Data Access||random access||accesses sorted to be consistent with database order||unconstrained||possibly sorted for loading, unconstrained for queries|
|Recovery||after failure, ensure database has committed updates and no others||after failure, rerun the batch to produce a new master file||application's responsibility||application's responsibility|
In a client-server system, a large number of personal computers communicate with shared servers on a local area network. This kind of system is very similar to a TP environment, where a large number of display devices connect to shared servers that run transactions. In some sense, TP systems were the original client-server systems with very simple desktop devices, namely, dumb terminals. As desktop devices have become more powerful, TP systems and personal computer systems have been converging into a single type of computing environment with different kinds of servers, such as file servers, communication servers, and TP servers.
There are many more system types than we have space to include here. Some examples are embedded systems, computer-aided design systems, data streaming systems, electronic switching systems, and traffic control systems.
Why Engineer a TP System?
Each system type that we looked at is designed for certain usage patterns. Although it is engineered for that usage pattern, it actually can be used in other ways. For example, people have used timesharing systems to run TP applications. These applications typically do not scale very well or use operating system resources very efficiently, but it can be done. For example, people have built special-purpose TP systems using real-time systems, and batch systems to run on a timesharing system.
TP has enough special requirements that it's worth engineering the system for that purpose. The amount of money businesses spend on TP systems justifies the additional engineering work vendors do to tailor their system products for TP — for better performance, reliability, and ease-of-use.
1.8 TP System Configurations
When learning the principles of transaction processing, it is helpful to have a feel for the range of systems where these principles are applied. We already saw some examples in Section 1.5 on TP benchmarks. Although those benchmark applications have limited functionality, they nevertheless are meant to be representative of the kind of functionality that is implemented for complete practical applications.
In any given price range, including the very high end, the capabilities of TP applications and systems continually grow, in large part due to the steadily declining cost of computing and communication. These growing capabilities enable businesses to increase the functionality of classical TP applications, such as travel reservations and banking. In addition, every few years, these capabilities enable entirely new categories of businesses. In the past decade, examples include large-scale Internet retailers and social networking web sites.
There is no such thing as an average TP application or system. Rather, systems that implement TP applications come in a wide range of sizes, from single servers to data centers with thousands of machines. And the applications themselves exhibit a wide range of complexity, from a single database with few dozen transaction types to thousands of databases running hundreds of millions of lines of code. Therefore, whatever one might say about typical TP installations will apply only to a small fraction of them and will likely be outdated within a few years.
A low-end system could be a departmental application supporting a small number of users who perform a common function. Such an application might run comfortably on a single server machine. For example, the sales and marketing team of a small company might use a TP application to capture sales orders, record customer responses to sales campaigns, alert sales people when product support agreements need to be renewed, and track the steps in resolving customer complaints. Even though the load on the system is rather light, the application might require hundreds of transaction types to support many different business functions.
By contrast, the workload of a large Internet service might require thousands of server machines. This is typical for large-scale on-line shopping, financial services, travel services, multimedia services (e.g., sharing of music, photos, and videos), and social networking. To ensure the service is available 24 hours a day, 7 days a week (a.k.a. 24 X 7), it often is supported by multiple geographically distributed data centers. Thus if one data center fails, others can pick up its load.
Like hardware configuration, software configurations cover a wide range. The system software used to operate a TP system may be proprietary or open source. It may use the latest system software products or ones that were introduced decades ago. It may only include a SQL database system and web server, or it may include several layers of transactional middleware and specialized database software.
The range of technical issues that need to be addressed is largely independent of the hardware or software configuration that is chosen. These issues include selecting a programming model; ensuring the ACID properties; and maximizing availability, scalability, manageability, and performance. These issues are the main subject of this book.
A transaction is the execution of a program that performs an administrative function by accessing a shared database. Transactions can execute on-line, while a user is waiting, or off-line (in batch mode) if the execution takes longer than a user can wait for results. The end-user requests the execution of a transaction program by sending a request message.
A transaction processing application is a collection of transaction programs designed to automate a given business activity. A TP application consists of a relatively small number of predefined types of transaction programs. TP applications can run on a wide range of computer sizes and may be centralized or distributed, running on local area or wide area networks. TP applications are mapped to a specially engineered hardware and software environment called a TP system.
The three parts of a TP application correspond to the three major functions of a TP system:
- Obtain input from a display or special device and construct a request.
- Accept a request message and call the correct transaction program.
- Execute the transaction program to complete the work required by the request.
Database management plays a significant role in a TP system. Transactional middleware components supply functions to help get the best price/performance out of a TP system and provide a structure in which TP applications execute.
There are four critical properties of a transaction: atomicity, consistency, isolation, and durability. Consistency is the responsibility of the program. The remaining three properties are the responsibility of the TP system.
- Atomicity: Each transaction performs all its operations or none of them. Successful transactions commit; failed transactions abort. Commit makes database changes permanent; abort undoes or erases database changes.
- Consistency: Each transaction is programmed to preserve database consistency.
- Isolation: Each transaction executes as if it were running alone. That is, the effect of running a set of transactions is the same as running them one at a time. This behavior is called serializability and usually is implemented by locking.
- Durability: The result of a committed transaction is guaranteed to be on stable storage, that is, one that survives power failures and operating system failures, such as a magnetic or solid-state disk.
If a transaction updates multiple databases or resource managers, then the two - phase commit protocol is required. In phase one, it ensures all resource managers have saved the transaction's updates to stable storage. If phase one succeeds, then phase two tells all resource managers to commit. This ensures atomicity, that is, that the transaction commits at all resource managers or aborts at all of them. Two-phase commit usually is implemented by a transaction manager, which tracks which resource managers are accessed by each transaction and runs the two-phase commit protocol.
Performance is a critical aspect of TP. A TP system must scale up to run many transactions per time unit, while giving one- or two-second response time. The standard measures of performance are the TPC benchmarks, which compare TP systems based on their maximum transaction rate and price per transaction for a standardized application workload.
A TP system is often critical to proper functioning of the enterprise that uses it. Therefore, another important property of TP systems is availability; that is, the fraction of time the system is running and able to do work. Availability is determined by how frequently a TP system fails and how quickly it can recover from failures.
TP systems have rather different characteristics than batch, real-time, and data warehouse systems. They therefore require specialized implementations that are tuned to the purpose. These techniques are the main subject of this book.
Making business transaction processing and applications work
Transaction processing system examples and SOA approaches
Key benchmarks for measuring transaction processing performance
Get caught up on the buzz about Amazon Redshift
Download a free PDF of this chapter: Principles of Transaction Processing, Second Edition
To purchase the book or similar titles, visit the Elsevier website
Read other excerpts from data management books in the chapter download library
Learn how in-database analytics can reduce customer turnover
Read about the analytical advantages of data warehouse systems