Older blog entries for nconway (starting at number 50)

I noticed that Joe Hellerstein, Mike Stonebraker, and James Hamilton (DBMS luminaries all) have published a nice, reasonably high-level paper describing the architecture and design principles of a typical database management system: "Architecture of a Database System".

MarkMail is now indexing all 630,000+ messages from the PostgreSQL mailing list archives. If, like me, you've been frustrated when trying to use the search engine and archives at archives.postgresql.org, I suggest checking out MarkMail. It's been working very well for me so far.

Ian Lance Taylor's blog has an interesting post on signed overflow behavior in C. According to the C standard, integer overflow results in undefined behavior, and modern versions of GCC take advantage of this to generate more efficient code. This topic was raised on -hackers by Tom a few years ago — at the time, only the -fwrapv flag was implemented by GCC. Now that GCC 4.2 provides -Wstrict-overflow, this might be worth investigating further.

The broader point here is that while this optimization is completely legal according to the C standard, it is inconsistent with the traditional C semantics, and runs the risk of breaking code that depends on integer overflow having the expected behavior. At least GCC now provides a flag to emit warnings for potentially broken code, which IMHO is a prerequisite for doing aggressive optimizations of this type. There's another interesting post on Ian Lance Taylor's blog that discusses this situation in general (e.g. alias optimizations are another instance where the C standard contradicts the traditional expectations of C programmers).

A few years ago, I did a summer internship with a group at Microsoft that was building a multimaster filesystem replication product. This was a very rewarding experience for several reasons. Now that the replication product has been shipped (in Windows 2003-R2, Vista, and Windows Live Messenger), I was happy to see that my mentor for that summer, Nikolaj Bjørner, has published a paper containing "lessons learned" from the project: "Models and Software Model Checking of a Distributed File Replication System". The paper is worth reading, for a few reasons:

1. Why is filesystem replication such a hard problem, particularly in the asynchronous, multi-master case?
   The paper talks about the basic problem and the approach the group took to solving it.
2. Perhaps more interestingly, how do you go about constructing a high-quality implementation of such a product?
   I was impressed by the group's emphasis on correctness. Nikolaj and Dan (the technical lead for the group) both had a CS theory background, so this is perhaps not surprising -- but it's interesting to see some of the practical techniques that they used to ensure they built a correct replication system:
   • A detailed specification (on the order of a few hundred pages)
   • A prototype of the system in OCaml, written concurrently with the specification but before the real implementation work began
   • A high-level, executable specification of the replication protocol in AsmL. This served as both a readable description of the protocol, as well as a way to automatically generate useful test cases.
   • Using model checking to verify the correctness of certain particularly complex aspects of the protocol (distributed garbage collection, conflict resolution).
   • A "simulator" that walked a random tree of filesystem operations, pausing after each node to verify that the system had correctly replicated the resulting filesystem state. Once a leaf node in the tree was reached, the simulator then backtracked, exploring another branch of the tree. The simulator was also clearly inspired by model checking techniques. By replacing certain components of the real system with virtualized ones (e.g. using a toy in-memory database), this tool could be used to test large numbers of scenarios very quickly.
   • Exhaustive testing. Using the simulator and a cluster of test machines, more than 500 billion test cases were examined.

On May 31, 2008, a tribute to honor the life and work of Jim Gray will be held at UC Berkeley. There's a technical session, for which registration is required, preceded by a general session that is open to the public. As the invitation email I received (thanks Elein!) states:

This is not a memorial, because Jim is still listed as missing, and will be so listed until about Jan 28, 2011. It is important that it is not referred to as a memorial, because it can't be a memorial until then. We believe that it is good to go ahead and recognize Jim's contributions, to honor him in a Tribute, before such a long time has passed.

There's a new post by Stonebraker up at The Database Column. I don't have much to add to the post itself, although it's interesting to hear some information about the old Sequoia 2000 project. I notice that Google and Yahoo were invited to the workshop — at first glance, it seems to me that the data management problems faced by the big web companies are quite dissimilar to the challenges facing "big science", but perhaps that's not the case.

This weekend's conference in Portland was a great experience. Much thanks to Selena Deckelmann, Josh Drake, and all the other volunteers for organizing and running the conference. Everything ran amazingly smoothly!

I've posted the slides to my talk on "Query Execution Techniques in PostgreSQL". I thought the talk went fairly well, although unfortunately I didn't have enough time to get to everything I wanted to discuss.

In the talk, one of the algorithms I discussed was the "hybrid hash join", which is the common hash join algorithm used by most modern DBMSs, including PostgreSQL. The night before, Jeff Davis tipped me off to the fact that the inventor of the hybrid hash join algorithm, Dr. Len Shapiro from PSU, was going to be in the audience! Thankfully I didn't get the details of the hybrid hash join wrong :) It was a pleasure to meet Dr. Shapiro, whose students are doing some interesting work improving hash index bulk build performance.

When running Postgres in EXEC_BACKEND mode on Linux, make sure to do:

echo 0 > /proc/sys/kernel/randomize_va_space

Preferably, before spending a while wondering about the non-deterministic process startup errors that will otherwise occur.

I gave a revised version of the "Introduction to Hacking PostgreSQL" tutorial at PgCon earlier today. I've posted the slides, handouts, and example patch here; this version uses a completely new example patch, and much of the introductory material has been revised. You can also find the slides at the PgCon page for the talk, which also includes a link to give me feedback.

I've been busy at school recently, finishing my undergraduate thesis and taking my final exams. I've got one more exam on Thursday, but once that's finished, I should be pretty much finished my undergraduate degree (fingers crossed).

In a complexity theory class I'm taking, I recently gave a talk on Kolmogorov complexity, introducing the basic ideas and discussing some notable applications. I'm no expert at AIT, so take the contents with a grain of salt.

Work

I've decided to take a full-time position at Amalgamated Insight, working on their PostgreSQL-based data stream management system (DSMS). I interned at AI last summer and I was very impressed, so I'm excited at the chance to work for them again. This also means I'll be moving to the Bay Area in early June.

Speaking of stream processing, I'll be giving a talk about data stream query processing at PgCon 2007. Gavin and I are also doing an introduction to modifying the PostgreSQL source code, similar in spirit to the "Introduction to Hacking PostgreSQL" tutorial we gave at the Anniversary Summit. The details are still a little hazy, but the goal is to make the session rewarding to both newcomers and to those who attended the previous tutorial.