On Thu, Nov 29, 2012 at 10:14 AM, Mike Blackwell <mike.blackwell@rrd.com> wrote:
>
>
>
> On Thu, Nov 29, 2012 at 12:09 PM, Jeff Janes <jeff.janes@gmail.com> wrote:
>>
>>
>> But If you do keep the drop index inside the transaction, then you
>> would probably be better off using truncate rather than delete, and
>> rebuild the index non-concurrently and move that inside the
>> transaction as well.
>>
>
>
> Hmm.... From the 9.2 manual it seems that might not work out so well:
>
> TRUNCATE is not MVCC-safe (see Chapter 13 for general information about
> MVCC). After truncation, the table will appear empty to all concurrent
> transactions, even if they are using a snapshot taken before the truncation
> occurred.
>
> It looks like other transactions could find an empty table while it was
> being reloaded under that approach.
They would block during the load, it is just after the load that they
would see the table as empty. I thought that that would only be a
problem for repeatable read or higher, but a test shows that read
committed has that problem as well. But yeah, that could definitely
be a problem with that method.
Cheers,
Jeff