Book Review: PostgreSQL 9.0 High Performance 75
eggyknap writes "Thanks in large part to the oft-hyped 'NoSQL' movement, database performance has received a great deal of press in the past few years. Organizations large and small have replaced their traditional relational database applications with new technologies like key-value stores, document databases, and other systems, with great fanfare and often great success. But replacing a database system with something radically different is a difficult undertaking, and these new database systems achieve their impressive results principally because they abandon some of the guarantees traditional database systems have always provided." Keep reading for the rest of eggyknap's review.
For those of us who need improved performance but don't have the luxury of redesigning our systems, and even more for those of us who still need traditional transactions, data integrity, and SQL, there is an option. Greg Smith's book, PostgreSQL 9.0 High Performance takes the reader step-by-step through the process of building an efficient and responsive database using "the world's most advanced open source database".| PostgreSQL 9.0 High Performance | |
| author | Gregory Smith |
| pages | Packt Publishing |
| publisher | 468 |
| rating | Packt Publishing |
| reviewer | 184951030X |
| ISBN | takes the reader step-by-step through the process of building an efficient and responsive database using "the world's most advanced open source database" |
| summary | 8/10 |
Greg Smith has been a major contributor to PostgreSQL for many years, with work focusing particularly on performance. In PostgreSQL 9.0 High Performance, Smith starts at the lowest level and works through a complete system, sharing his experience with systematic benchmarking and detailed performance improvement at each step. Despite the title, the material applies not only to PostgreSQL's still fairly new 9.0 release, but to previous releases as well. After introducing PostgreSQL, briefly discussing its history, strengths and weaknesses, and basic management, the book dives into a detailed discussion of hardware and benchmarking, and doesn't come out for 400 pages.
Databases vary, of course, but in general they depend on three main hardware factors: CPU, memory, and disks. Smith discusses each in turn, and in substantial detail, as demonstrated in a sample chapter available from the publisher, Packt Publishing. After describing the various features and important considerations of each aspect of a database server's hardware, the book introduces and demonstrates powerful and widely available tools for testing and benchmarking. This section in particular should apply easily not only to administrators of PostgreSQL databases, but users of other databases, or indeed other applications as well, where CPU, memory, or disk performance is a critical factor. Did you know, for instance, the difference between "write-through" and "write-back" caching in disk, and why it matters to a database? Or did you know that disks perform better depending on which part of the physical platter they're reading? How does memory performance compare between various common CPUs through the evolution of their different architectures?
At every step, Smith encourages small changes and strict testing, to ensure optimum results from your performance efforts. His discussion includes methods for reducing and correcting variability, and sticks to easily obtained and interpreted tools, whose output is widely understood and for which support is readily available. The underlying philosophy has correctly been described as "measure, don't guess," a welcome relief in a world where system administrators often make changes based on a hunch or institutional mythology.
Database administrators often limit their tools to little more than building new indexes and rewriting queries, so it's surprising to note that those topics don't make their appearance until chapters 9 and 10 respectively, halfway through the book. That said, they receive the same detailed attention given earlier to database hardware, and later on to monitoring tools and replication. Smith thoroughly explains each of the operations that may appear in PostgreSQL's often overwhelming query plans, describes each index type and its variations, and goes deeply into how the query planner decides on the best way to execute a query.
Other chapters cover such topics as file systems, configuration options suitable for various scenarios, partitioning, and common pitfalls, each in depth. In fact, the whole book is extremely detailed. Although the tools introduced for benchmarking, monitoring, and the like are well described and their use nicely demonstrated, this is not a book a PostgreSQL beginner would use to get started. Smith's writing style is clear and blessedly free of errors and confusion, as is easily seen by his many posts on PostgreSQL mailing lists throughout the years, but it is deeply detailed, and the uninitiated could quickly get lost.
This is also a very long book, and although not built strictly as a reference manual, it's probably best treated as one, after an initial thorough reading. It covers each topic in such detail that each must be absorbed before further reading can be beneficial. Figures and other non-textual interruptions are, unfortunately, almost nowhere to be found, so despite the author's clear and easy style, it can be a tiring read.
It is, however, one of the clearest, most thorough, and best presented descriptions of the full depth of PostgreSQL currently available, and doubtless has something to teach any frequent user of a PostgreSQL database. Those planning a new database will welcome the straightforward and comprehensive presentation of hardware-level details that are difficult or impossible to change after a system goes into production; administrators will benefit from its discussion of configuration options and applicable tools; and users and developers will embrace its comprehensive description of query planning and optimization. PostgreSQL 9.0 High Performance will be a valuable tool for all PostgreSQL users interested in getting the most from their database.
You can purchase PostgreSQL 9.0 High Performance from amazon.com. Slashdot welcomes readers' book reviews -- to see your own review here, read the book review guidelines, then visit the submission page.
Re:Cut out the middleman (Score:5, Insightful)
Point 1:
Databases already use B-trees for their indexes.
Point 2:
Coding business logic for performance instead of verifiability, testability or stability is beyond retarded.
Re:Cut out the middleman (Score:5, Insightful)
Um...
I considered modding, but "wrong" is not a mod.
B-Trees are certainly not the fastest way to store and retrieve data. You may want to investigate those strange things called "hashes". Fortunately for you, most database systems know how.
As well, "assembler" isn't the fastest coding system. It is so very slow and tedious to use that algorithm tends to get overlooked. Back to your "B-Tree" again -- have you tried coding a B-Tree in assembler?
To give a hard example -- back in the days of OS/2, the filesystem was coded in assembler. I imagine that it was considered the right way. After being recoded in C, the thing ran faster -- mostly because the developers could concentrate on better ways of doing things, rather than the drudge work of getting the assembler correct.
Another example -- replacing a "C" coded inner loop in a console game with a scheme interpreted version, that ended up being faster because it could fit into cache.
Sometimes upstream is simply better, even for performance.
Database systems are in the same category. It is rather difficult to optimize specifically, and, by the time you did this, typically the data store would be "out of touch" with the actual business requirements. So, most of us use a good database, with a query optimizer, and let it take care of that part of the problem. Yes, it may need some tweaking, but those tweaks will tend to be independent of the business logic. This can give rise to specialists who can be called on if needed.
Using your model, the "assembler tuned B-Tree" custom coded, there would be no paladins available to assist in the tuning. Indeed, moving to another architecture would be difficult (or impossible). It would be ridiculously expensive, and would likely never be upgraded. Hey, CICS and IMS are still in use, aren't they? And THEY are more portable than your suggestion. (CICS and IMS just turned 40).
PS. There are CICS and IMS paladin around. There wouldn't be any for your "solution". Also, IMS can use direct storage links, and, because data in IMS is hierarchical, even IMS can outperform "B-Tree".