Re: WIP: Detecting SSI conflicts before reporting constraint violations

On Sat, Mar 26, 2016 at 5:04 AM, David Steele <david(at)pgmasters(dot)net> wrote:
> Hi Thomas,
>
> On 3/13/16 8:20 PM, Thomas Munro wrote:
>
>> <...> I will have another look at this in
>> a few days but for now I need to do some other things, so I'm posting
>> these observations in case they are in some way helpful...
>
>
> It's not clear to me what state this patch should be in but the thread has
> been idle for over a week.
>
> Have you had the chance to take a look as you indicated above?

Here's a summary of where we are today:

1. It looks like we have general consensus that this is a problem and
that we should fix it, but not about whether a change should be
backpatched, if/when we arrive at an acceptable patch.

2. We have a 1 line patch (+ comments and isolation tests) which
covers the cases that I've heard complaints about. These cases all
involve either checking if a value already exists with SELECT or
computing a new value based on existing values with something like
SELECT MAX(id) + 1 and then inserting it in concurrent transactions.

3. In the process of designing isolation tests, I found a case that
doesn't cover: where one session simply inserts, while the another
checks explicitly whether it's OK to insert, finds that it is, and
then tries to do so. I haven't figured out how to detect an SSI
conflict before the UCV in this case.

Realistically I'm not going to have a solution to the third problem
before the 31st.

--
Thomas Munro
http://www.enterprisedb.com