Horton

Re: Potential G2-item cycles under serializable isolation

Kyle Kingsbury <aphyr@jepsen.io>

View thread View on postgr.es
From: Kyle Kingsbury <aphyr@jepsen.io>
To: Thomas Munro <thomas.munro@gmail.com>
Cc: Peter Geoghegan <pg@bowt.ie>, Andres Freund <andres@anarazel.de>, Kevin Grittner <kgrittn@gmail.com>, Jeff Davis <pgsql@j-davis.com>, PostgreSQL mailing lists <pgsql-bugs@lists.postgresql.org>
Date: 2020-06-11T12:39:13Z
Lists: pgsql-bugs 
On 6/11/20 1:40 AM, Thomas Munro wrote:
> What do you think about this documentation update?

I think this is really helpful! You can go a little further if you like. Right 
now, SSI vs serializable and SI vs RR are both described as "differences in 
behavior", which kinda leaves it unclear as to how those levels are related. If 
you want to follow Berenson, Adya, et al.'s broad interpretation, I'd say 
something like

"Snapshot isolation prevents some anomalies, like phantoms, which repeatable 
read allows. It also allows some anomalies, like G2-item (write skew), which 
repeatable read prevents."

... and if you like, you could follow that with a discussion of how SI, thanks 
to ambiguity in the ANSI specification, could also be interpreted as stronger 
than RR. If you stick with the strict interpretation, I'd add a note here about 
how readers familiar with the SI literature might be confused, and that you're 
diverging from Berenson et al., but your position is defensible under the spec.

"Serializable Snapshot Isolation is stronger than serializability: it allows 
only serializable histories--and prohibits some additional histories in addition."

If you wanna talk about the performance rationale, that's valid too. :)

--Kyle

Commits

  1. Doc: Add references for SI and SSI.

  2. Improve comments for [Heap]CheckForSerializableConflictOut().

  3. Avoid update conflict out serialization anomalies.

  4. Implement genuine serializable isolation level.