Thanks guys for the valuable info. The key take away for me is clear: keep transactions short under all circumstances.
Cheers,
Eduard
Am Fr., 1. Dez. 2023 um 01:31 Uhr schrieb Andres Freund <andres@anarazel.de>:
Hi,
On 2023-11-30 18:51:35 -0500, Tom Lane wrote: > On what grounds do you assert that? Operations on shared catalogs > are visible across databases. Admittedly they can't be written by > ordinary DML, and I'm not sure that we make any promises about DDL > writes honoring serializability. But I'm unwilling to add > "optimizations" that assume that that will never happen.
I'd say the issue is more that it's quite expensive to collect the information. I tried in the past to make the xmin computation in GetSnapshotData() be database specific, but it quickly shows in profiles, and GetSnapshotData() unfortunately is really performance / scalability critical.
If that weren't the case, we could check a shared horizon for shared tables, and a non-shared horizon otherwise.
In some cases we can compute a "narrower" horizon when it's worth the cost, but quite often we lack the necessary data, because various backends have stored the "global" xmin in the procarray.