On 20.08.2019 20:01, Pavel Stehule wrote:
Another solution is wait on ZHeap storage and replica can to have own UNDO log.
I thought about implementation of special table access method for temporary tables.
+1
Unfortunately implementing special table access method for temporary tables doesn't solve all problems.
XID generation is not part of table access methods.
So we still need to assign some XID to write transaction at replica which will not conflict with XIDs received from master.
Actually only global temp tables can be updated at replica and so assigned XIDs can be stored only in tuples of such relations.
But still I am not sure that we can use arbitrary XID for such transactions at replica.
Also I upset by amount of functionality which has to be reimplemented for global temp tables if we really want to provide access method for them:
1. CLOG
2. vacuum
3. MVCC visibility
And still it is not possible to encapsulate all changes need to support writes to temp tables at replica inside table access method.
XID assignment, transaction commit and abort, subtransactions - all this places need to be patched.
--
Konstantin Knizhnik
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company
