On 22 March 2014 05:32, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Thom Brown <thom@linux.com> writes:
>> Is it necessary for a partial index that doesn't include the row to be
>> involved in locking?
>
> Yes. You can't determine whether the index needs to get a new entry
> without examining its metadata, and that's what the lock is mainly about.
I see. Why does this apply to deletes too?
> The only possible alternative would be to take the minimum possible
> lock (AccessShareLock) on each index so its metadata would hold still,
> and then upgrade that to RowExclusiveLock on the one(s) we find need
> insertions. This is not better; it means *more* lock management traffic
> not less, and lock upgrades increase the potential for deadlocks.
Yes, I can see that wouldn't be an improvement.
--
Thom
