Hi Ronan,
On 9/17/21 10:09 AM, Ronan Dunklau wrote:
> Hello Bertrand,
>
> Le mardi 4 mai 2021, 11:55:43 CEST Drouvot, Bertrand a écrit :
>> Implementation overview:
>>
>> * A new catalog snapshot is added: DirtyCatalogSnapshot.
>> * This catalog snapshot is a dirty one to be able to look for
>> in-flight dependencies.
>> * Its usage is controlled by a new UseDirtyCatalogSnapshot variable.
>> * Any time this variable is being set to true, then the next call to
>> GetNonHistoricCatalogSnapshot() is returning the dirty snapshot.
>> * This snapshot is being used to check for in-flight dependencies and
>> also to get the objects description to generate the error messages.
>>
> I quickly tested the patch, it behaves as advertised, and passes tests.
Thanks for looking at it!
>
> Isolation tests should be added to demonstrate the issues it is solving.
Good point. I'll have a look.
>
> However, I am bit wary of this behaviour of setting the DirtyCatalogSnapshot
> global variable which is then reset after each snapshot acquisition: I'm
> having trouble understanding all the implications of that, if it would be
> possible to introduce an unforeseen snapshot before the one we actually want
> to be dirty.
I don't think that could be possible as long as:
- this is a per backend variable
- we pay attention where we set it to true
But i might be missing something.
Do you have any corner cases in mind?
> I don't want to derail this thread, but couldn't predicate locks on the
> pg_depend index pages corresponding to the dropped object / referenced objects
> work as a different approach ?
I'm fine to have a look at another approach if needed, but does it mean
we are not happy with the current approach proposal?
Thanks
Bertrand