A Critique of ANSI SQL Isolation Levels (1995) [pdf] | Hacker News Reader

A Critique of ANSI SQL Isolation Levels (1995) [pdf]

microsoft.com

76 pointsluu12 comments

This is a classic paper, and well worth reading if you use relational databases for anything serious. Using 'phenomena' is a nice framing, and one that I found particularly effective for understanding how to write correct database transactions at various isolation levels. Perhaps the most surprising result here is "REPEATABLE READ »« Snapshot Isolation", meaning that it's not possible to order one of the two as "stronger" isolation.

The author list here is a round-up of real database legends, including the late Jim Gray and Pat O'Neil, and Phil Bernstein and Elizabeth O'Neil who are still doing work in databases and systems.

During my days at Microsoft I had "worked with"^ with 4 out of 6 co-authors of this paper. It's a weird feeling like I suddenly got back there.

^ not a whole a lot, admittedly

Also, to your point, I believe it is possible to strengthen Snapshot Isolation to the point where it is strictly stronger than Repeatable Read. There was a paper at VLDB circa 2007, I belie it was done by a researcher from Oracle. Can't recall any details, other than the guy was pretty young and well-spoken.

I believe that you're thinking of Michael J. Cahill:

https://courses.cs.washington.edu/courses/cse444/08au/544M/R...

This became the basis of PostgreSQL's serializable snapshot isolation feature:

https://drkp.net/papers/ssi-vldb12.pdf

One funny thing about that VLDB'12 paper is Kevin Grittner's affiliation. Not often you see a state supreme court publishing cutting-edge database research. I'd love to hear the story around that.

I used to work with Kevin, back when I was with VMware. I can confirm that it's an interesting story!

Quite possible. In the years since my brain has completely unloaded transaction processing to make room for supply chain management, so I can no longer relate to the subject to see if it's the same research. :(

I recall I approached him in the break and congratulated him on the fantastic presentation. And it was fantastic - like a murder mystery story, it kept you on the edge of the seat, answering questions it has created earlier and culminating into the answer. An older guy, probably his science advisor, protested that it's not only the presentation that was good, and the research itself was not too shabby. To that I retorted that for all we know there was a lot of good research in the room but we will never find out because delivery was lacking. Research must beget research, it's what makes science from hobby. I think we reached an agreement. I often think back to that moment - what we build should lend itself to further building, or it becomes a dead end. Neither science nor technology are made from dead-ends, the proverbial shoulders of giants must stack up for the civilization to advance.

Here is the thesis itself:

https://ses.library.usyd.edu.au/bitstream/handle/2123/5353/m...

I do remember thinking that it was unusually well argued and presented. It's much more accessible to the practitioner than a lot of the other research in this area. In my experience this is a positive signal about the quality of the research.

That's gotta be him. I just read the abstract, it's mountain-creek-clear.

If anyone would like to learn more about this topic, Jepsen has some great articles on isolation levels and consistency models.

https://jepsen.io/consistency

(1995)

Added. Thanks!

Reply on news.ycombinator.com