On Mon, Oct 10, 2016 at 7:59 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:
>> On 9/9/16 12:39 PM, Tom Lane wrote:
>>> I wouldn't really imagine that it's sensible to set READ ONLY
>>> mid-transaction at all, but if it means anything to do that, surely
>>> it ought to mean that no updates can happen *after* you set it.
>
>> I think there is a bit of code missing in check_transaction_read_only().
>> We prevent changing from read-only to read-write after the first query
>> but not vice versa. That seems like an oversight.
>
> The comments around the code make it absolutely clear that it's
> intentional, not an "oversight". Whether it's a good idea is open
> for discussion, certainly, but I don't see how you can imagine that
> it wasn't considered.
I seem to remember that the current state of affairs evolved near
the end of 9.1 development, when it surfaced that this GUC could be
changed at will during a transaction and that made some nice SSI
optimizations impossible. If memory serves, Tom preferred that we
not lose the ability to change from READ WRITE to READ ONLY within
a transaction, and that wasn't hard to accommodate (we capture the
state of the flag at the start of a serializable transaction and
use that for determining serializable optimizations), so I didn't
really care. I don't remember anyone arguing against this way at
the time.
Current behavior seems harmless and possibly useful to me, but it
seems marginal enough I wouldn't care if the change from READ WRITE
to READ ONLY was also prohibited.
--
Kevin Grittner
EDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company