On Tue, May 14, 2019 at 9:11 AM Michael Paquier <michael@paquier.xyz> wrote:
On Tue, May 14, 2019 at 08:08:05AM +0200, Chris Travers wrote: > Having thought about this a bit, I think the best solution would be to have > grant take out an access share lock to the tables granted. This would > prevent concurrent alter table operations from altering the schema > underneath the grant as well, and thus possibly cause other race conditions. > > Any thoughts?
"tuple concurrently updated" is an error message which should never be user-facing, and unfortunately there are many scenarios where it can be triggered by playing with concurrent DDLs: https://postgr.es/m/20171228063004.GB6181@paquier.xyz
If you have an idea of patch, could you write it? Having an isolation test case would be nice as well.
I will give Nick a chance to do the patch if he wants it (I have reached out). Otherwise sure.
I did notice one more particularly exotic corner case that is not resolved by this proposed fix.
If you have two transactions with try to grant onto the same pg entity (table etc) *both* will typically fail on the same error.
I am not sure that is a bad thing because I am not sure how concurrent grants are supposed to work with MVCC but I think that would require a fundamentally different approach.