On Sun, Nov 14, 2010 at 1:12 PM, Heikki Linnakangas
<heikki.linnakangas@enterprisedb.com> wrote:
> On 14.11.2010 00:29, Robert Haas wrote:
>>
>> On Sat, Nov 13, 2010 at 12:13 PM, Tom Lane<tgl@sss.pgh.pa.us> wrote:
>>>
>>> Hannu Krosing<hannu@2ndQuadrant.com> writes:
>>>>
>>>> On Sat, 2010-11-13 at 10:51 -0500, Tom Lane wrote:
>>>>>
>>>>> If a table has no indexes, we will always decide that any same-page
>>>>> update operation is a HOT update, since obviously it isn't modifying
>>>>> any indexed columns. But is there any benefit to doing so?
>>>
>>>> If we do the in-page "mini vacuum" even without HOT, then there should
>>>> be no benefit from index-less HOT updates.
>>>
>>> AFAICS we do: heap_update marks the page as prunable whether it's a HOT
>>> update or not. The only difference between treating the update as HOT vs
>>> not-HOT is that if there was more than one HOT update, the intermediate
>>> tuples could be completely reclaimed by page pruning (ie, their line
>>> pointers go away too). With not-HOT updates, the intermediate line
>>> pointers would have to remain in DEAD state until vacuum, since page
>>> pruning wouldn't know if there were index entries pointing at them.
>>> But that seems like a pretty tiny penalty.
>>
>> I'm not at all convinced that's a tiny penalty.
>
> Me neither. It's a tiny penalty when you consider one update, but if you
> repeatedly update the same tuple, you accumulate dead line pointers until
> the next real vacuum runs. With HOT updates, you reach a steady state where
> page pruning is all you need. Then again, if you're repeatedly updating a
> row in a table with no indexes, presumably it's a very small table or you
> would create an index on it. And frequently autovacuuming a small index is
> quite cheap too.
The case here is when you have say a control table that is managing a
gapless sequence, or a materialization table with a very small number
of records. These type of tables get updated very frequently, perhaps
in every transaction. People without detailed implementation
knowledge of postgresql might assume that leaving an index off the
table is faster in these situations.
The danger here is that if autovacuum is stalled for whatever reason,
you get exponentially bad behavior as the table gets stuffed with bad
records. index-less hot was put in intentionally. As autovacuum gets
smarter and smarter, the reasoning to do this get weaker.
merlin