Re: ALTER TABLE lock strength reduction patch is unsafe

On 03/01/2014 12:06 PM, Simon Riggs wrote:
> On 27 February 2014 08:48, Simon Riggs <simon@2ndquadrant.com> wrote:
>> On 26 February 2014 15:25, Andres Freund <andres@2ndquadrant.com> wrote:
>>> On 2014-02-26 15:15:00 +0000, Simon Riggs wrote:
>>>> On 26 February 2014 13:38, Andres Freund <andres@2ndquadrant.com> wrote:
>>>>> Hi,
>>>>>
>>>>> On 2014-02-26 07:32:45 +0000, Simon Riggs wrote:
>>>>>>> * This definitely should include isolationtester tests actually
>>>>>>>   performing concurrent ALTER TABLEs. All that's currently there is
>>>>>>>   tests that the locklevel isn't too high, but not that it actually works.
>>>>>> There is no concurrent behaviour here, hence no code that would be
>>>>>> exercised by concurrent tests.
>>>>> Huh? There's most definitely new concurrent behaviour. Previously no
>>>>> other backends could have a relation open (and locked) while it got
>>>>> altered (which then sends out relcache invalidations). That's something
>>>>> that should be tested.
>>>> It has been. High volume concurrent testing has been performed, per
>>>> Tom's original discussion upthread, but that's not part of the test
>>>> suite.
>>> Yea, that's not what I am looking for.
>>>
>>>> For other tests I have no guide as to how to write a set of automated
>>>> regression tests. Anything could cause a failure, so I'd need to write
>>>> an infinite set of tests to prove there is no bug *somewhere*. How
>>>> many tests are required? 0, 1, 3, 30?
>>> I think some isolationtester tests for the most important changes in
>>> lock levels are appropriate. Say, create a PRIMARY KEY, DROP INHERIT,
>>> ...  while a query is in progress in a nother session.
>> OK, I'll work on some tests.
>>
>> v18 attached, with v19 coming soon
> v19 complete apart from requested comment additions

I've started to look at this patch and re-read the thread.  The first
thing I noticed is what seems like an automated replace error.  The docs
say "This form requires only an SHARE x EXCLUSIVE lock." where the "an"
was not changed to "a".

Attached is a patch-on-patch to fix this.  A more complete review will
come later.

--
Vik