Hi,
> To achieve this we can allocate Form_pg_class structures (for shared
> relations… a small number) on shared memory.
But why would this be necessary / a good idea? Even if we decided it
were, it seems like it'd end up being quite invasive. But I doubt it's
a good plan, because relcache entries want / need to be updated
differently in the transaction that does the changes (as it needs to see
the effect of catalog changes before commit) than other sessions (which
only should see them after commit).
It will be a good idea as, we can avoid maintaining file
(creation/deletion/updation) for period of engine running and we do not need to
invalidate other backend cache.
For transaction(which change catalog), we can allocate a new private
Form_pg_class structure and do operations on it and update shared structure
post commit.
Routine roughly will be-
1) A backend having a relcache entry for a shared rel where rd_rel part points
to shared memory structure.
Rel->rd_rel is storing a shared memory pointer.
2) Wants to update the entry...allocate a new private structure and memcpy the
shared content to this new memory and point rd_rel to private memory
Rel->rd_rel is storing a pointer to process specific structure.
3) Transaction committed and we copy the private memory content to shared
memory area and point rd_rel again to shared memory.
4) free private memory.
5) Other backends do not do any invalidation but still get the latest updated
values.
Why is this good idea?
Lets take an example (assuming we have 1000 postgres backends running).
With shared memory scheme-
Operation wise, for a transaction, we allocate/free once (private
memory allocation) and memcpy data to and fro (from shared to private and back
to shared)...
Overall memory footprint 1 shared copy and 1 private only when updating.
No file creation/deletion/updation.
With current file scheme-
Operation wise, for a transaction, we use private cache but we need to
invalidate 1000 other caches( which will be atleast 1000 memcpy and
allocate/free) and may involve reading back in page of pg_class.
Overall memory footprint 1000 private copies.
We have to create/delete/update init file and synchronize around it.
Having said that we may not worry about transaction for updating all
values...(I think relfrozenxid can be updated by a CAS operation ...still
thinking on it).
-Nishant
