1. Statistic for global temporary tables (including number of tuples, pages and all visible flag). My position is the following: while in most cases it should not be a problem, because users rarely create indexes or do analyze for temporary tables, there can be situations when differences in data sets of global temporary tables in different backends can really be a problem. Unfortunately I can not propose good solution for this problem. It is certainly possible to create some private (per-backend) cache for this metadata. But it seems to requires changes in many places.
Yeah. I don't really like just sharing them but it's not that bad either.
2. Your concerns about performance penalty of global temp tables accessed through shared buffers comparing with local temp tables access through local buffers. I think that this concern is not actual any more because there is implementation of global temp tables using local buffers. But my experiments doesn't show significant difference in access speed of shared and local buffers. As far as shared buffers are used to be much larger than local buffers, there are more chances to hold all temp relation in memory without spilling it to the disk. In this case access to global temp table will be much faster comparing with access to local temp tables.
You ignore the costs of evicting non-temporary data from shared_buffers, i.e. contention for space. Also increased chance of backends being forced to do direct write-out due to lack of s_b space for dirty buffers.
> In case of pulling all content of temp table in memory (pg_prewarm) global temp table with shared buffers becomes faster.
Who would ever do that?
I forget or do not notice some of your questions, would you be so kind as to repeat them?