On 27/09/2021 14:59, Aleksander Alekseev wrote:
Hi hackers,
As a matter of fact, I think the patch I suggested is the right approach:
let me elaborate on why.
[...]
It is straightforward to replace it by implementing the Table AM methods
above, but we are missing a callback on dropping the table. If we have that,
we can record the table-to-be-dropped in a similar manner to how the heap AM
does it and register a transaction callback using RegisterXactCallback.
Since no one objected in 5 months, I assume Mats made a good point. At least,
personally, I can't argue.
I agree that having a table AM callback at relation drop would make it
more consistent with creating and truncating a relation. Then again, the
indexam API doesn't have a drop-callback either.
But what can you actually do in the callback? WAL replay of dropping the
storage needs to work without running any AM-specific code. It happens
as part of replaying a commit record. So whatever action you do in the
callback will not be executed at WAL replay. Also, because the callback
merely *schedules* things to happen at commit, it cannot generate
separate WAL records about dropping resources either.
Mats's in-memory table is an interesting example. I guess you don't even
try WAL-logging that, so it's OK that nothing happens at WAL replay. As
you said, the callback to schedule deletion of the shared memory block
and use an end-of-xact callback to perform the deletion. You're
basically re-inventing a pending-deletes mechanism similar to smgr's.
I think you could actually piggyback on smgr's pending-deletions
mechanism instead of re-inventing it. In the callback, you can call
smgrGetPendingDeletes(), and drop the shared memory segment for any
relation in that list.
- Heikki
