Re: Usage of epoch in txid_current

On Thu, Mar 28, 2019 at 8:30 AM Thomas Munro <thomas.munro@gmail.com> wrote:
> Thanks for the reviews!  Pushed.

Any ideas we should move towards 64-bit xids in more places?  That has
been discussed several times already.  I think last time it was
discussed in person during FOSDEM PGDay 2018 Developer Meeting [1].
There we've discussed that it probably doesn't worth it to change
32-bit on-disk xids in heap.  It's better to leave existing heap "as
is", but allow other pluggable table access methods to support 64-bit
xids.  Given now we have pluggable table access methods, we may build
a plan on full support of 64-bit xids in core.

In my vision sketchy plan may look like this.

1. Change all non-heap types of xids from TransactionId to
FullTransactionId.  But in shared memory change TransactionId to
pg_atomic_uint64 in order to guarantee atomicity of access, which we
require in multiple places.
2. Also introduce FullMultixactId, and apply to MultixactId the
similar change as #1.
3. Change SLRU on-disk format to handle 64-bit xids and multixacts.
In particular make SLRU page numbers 64-bit too.  Add SLRU upgrade
procedure to pg_upgrade.
4. Change relfrozenxid and relminmxid to something like rellimitxid
and rellimitmxid.  So, we don't imply there is restriction of 32-bit
xids.  Instead, we let table AM place (or don't place) a limit to
advancing nextXid, nextMultixact.
5. Table AM API would be switched to 64-bit xids/multixacts, while
heap will remain using 32-bit.  So, heap should convert 32-bit xid
value to 64-bit basing on nextXid/nextMultixact.  Assuming we set
rellimitxid and rellimitmxid for relation as oldestxid + 2^32 and
oldestmxid + 2^32, we may safely assume the 32-bit values of
xids/multixacts are in 2^32 range before nextXid/nextMultixact.
Thanks to this even in heap we would be able to operate 2^32
xids/multixacts simultaneously instead of 2^31 we have now.

So, this is how I see the possible plan.  I would be glad to see a feedback.

Unfortunately, I don't have enough time to implement all of this.  But
I think there are many interested parties in finally having 64-bit
xids in core.  Especially assuming we now have pluggable table AMs,
and it would be ridiculous to spear limitation of 32-bit xids to new
table AMs.

Links

1. https://wiki.postgresql.org/wiki/FOSDEM/PGDay_2018_Developer_Meeting

------
Alexander Korotkov
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company