So far, we've discussed database systems in general terms. It's time now to consider Berkeley DB in particular and see how it fits into the framework we have introduced. The key question is, what kinds of applications should use Berkeley DB?
Berkeley DB is an open source embedded database library that provides scalable, high-performance, transaction-protected data management services to applications. Berkeley DB provides a simple function-call API for data access and management.
By "open source," we mean that Berkeley DB is distributed under a license that conforms to the Open Source Definition. This license guarantees that Berkeley DB is freely available for use and redistribution in other open source products. Sleepycat Software sells commercial licenses for redistribution in proprietary applications, but in all cases the complete source code for Berkeley DB is freely available for download and use.
Berkeley DB is embedded because it links directly into the application. It runs in the same address space as the application. As a result, no inter-process communication, either over the network or between processes on the same machine, is required for database operations. Berkeley DB provides a simple function-call API for a number of programming languages, including C, C++, Java, Perl, Tcl, Python, and PHP. All database operations happen inside the library. Multiple processes, or multiple threads in a single process, can all use the database at the same time as each uses the Berkeley DB library. Low-level services like locking, transaction logging, shared buffer management, memory management, and so on are all handled transparently by the library.
The library is extremely portable. It runs under almost all UNIX and Linux variants, Windows, and a number of embedded real-time operating systems. It runs on both 32-bit and 64-bit systems. It has been deployed on high-end Internet servers, desktop machines, and on palmtop computers, set-top boxes, in network switches, and elsewhere. Once Berkeley DB is linked into the application, the end user generally does not know that there's a database present at all.
Berkeley DB is scalable in a number of respects. The database library itself is quite compact (under 300 kilobytes of text space on common architectures), but it can manage databases up to 256 terabytes in size. It also supports high concurrency, with thousands of users operating on the same database at the same time. Berkeley DB is small enough to run in tightly constrained embedded systems, but can take advantage of gigabytes of memory and terabytes of disk on high-end server machines.
Berkeley DB generally outperforms relational and object-oriented database systems in embedded applications for a couple of reasons. First, because the library runs in the same address space, no inter-process communication is required for database operations. The cost of communicating between processes on a single machine, or among machines on a network, is much higher than the cost of making a function call. Second, because Berkeley DB uses a simple function-call interface for all operations, there is no query language to parse, and no execution plan to produce.
Berkeley DB applications can choose the storage structure that best suits the application. Berkeley DB supports hash tables, Btrees, simple record-number-based storage, and persistent queues. Programmers can create tables using any of these storage structures, and can mix operations on different kinds of tables in a single application.
Hash tables are generally good for very large databases that need predictable search and update times for random-access records. Hash tables allow users to ask, "Does this key exist?" or to fetch a record with a known key. Hash tables do not allow users to ask for records with keys that are close to a known key.
Btrees are better for range-based searches, as when the application needs to find all records with keys between some starting and ending value. Btrees also do a better job of exploiting locality of reference. If the application is likely to touch keys near each other at the same time, the Btrees work well. The tree structure keeps keys that are close together near one another in storage, so fetching nearby values usually doesn't require a disk access.
Record-number-based storage is natural for applications that need to store and fetch records, but that do not have a simple way to generate keys of their own. In a record number table, the record number is the key for the record. Berkeley DB will generate these record numbers automatically.
Queues are well-suited for applications that create records, and then must deal with those records in creation order. A good example is on-line purchasing systems. Orders can enter the system at any time, but should generally be filled in the order in which they were placed.
Berkeley DB offers important data management services, including concurrency, transactions, and recovery. All of these services work on all of the storage structures.
Many users can work on the same database concurrently. Berkeley DB handles locking transparently, ensuring that two users working on the same record do not interfere with one another.
The library provides strict ACID transaction semantics, by default. However, applications are allowed to relax the isolation guarantees the database system makes.
Multiple operations can be grouped into a single transaction, and can be committed or rolled back atomically. Berkeley DB uses a technique called two-phase locking to be sure that concurrent transactions are isolated from one another, and a technique called write-ahead logging to guarantee that committed changes survive application, system, or hardware failures.
When an application starts up, it can ask Berkeley DB to run recovery. Recovery restores the database to a clean state, with all committed changes present, even after a crash. The database is guaranteed to be consistent and all committed changes are guaranteed to be present when recovery completes.
An application can specify, when it starts up, which data management services it will use. Some applications need fast, single-user, non-transactional Btree data storage. In that case, the application can disable the locking and transaction systems, and will not incur the overhead of locking or logging. If an application needs to support multiple concurrent users, but doesn't need transactions, it can turn on locking without transactions. Applications that need concurrent, transaction-protected database access can enable all of the subsystems.
In all these cases, the application uses the same function-call API to fetch and update records.
Berkeley DB was designed to provide industrial-strength database services to application developers, without requiring them to become database experts. It is a classic C-library style toolkit, providing a broad base of functionality to application writers. Berkeley DB was designed by programmers, for programmers: its modular design surfaces simple, orthogonal interfaces to core services, and it provides mechanism (for example, good thread support) without imposing policy (for example, the use of threads is not required). Just as importantly, Berkeley DB allows developers to balance performance against the need for crash recovery and concurrent use. An application can use the storage structure that provides the fastest access to its data and can request only the degree of logging and locking that it needs.
Because of the tool-based approach and separate interfaces for each Berkeley DB subsystem, you can support a complete transaction environment for other system operations. Berkeley DB even allows you to wrap transactions around the standard UNIX file read and write operations! Further, Berkeley DB was designed to interact correctly with the native system's toolset, a feature no other database package offers. For example, Berkeley DB supports hot backups (database backups while the database is in use), using standard UNIX system utilities, for example, dump, tar, cpio, pax or even cp.
Finally, because scripting language interfaces are available for Berkeley DB (notably Tcl and Perl), application writers can build incredibly powerful database engines with little effort. You can build transaction-protected database applications using your favorite scripting languages, an increasingly important feature in a world using CGI scripts to deliver HTML.
Copyright Sleepycat Software