On Wed, Feb 22, 2017 at 12:01 AM, Amit Kapila <amit.kapil...@gmail.com> wrote:
> On Tue, Feb 21, 2017 at 1:09 AM, Simon Riggs <si...@2ndquadrant.com> wrote:
>> On 20 February 2017 at 10:27, Amit Kapila <amit.kapil...@gmail.com> wrote:
>>> On Mon, Feb 20, 2017 at 3:01 PM, Simon Riggs <si...@2ndquadrant.com> wrote:
>>>> On 20 February 2017 at 09:15, Amit Kapila <amit.kapil...@gmail.com> wrote:
>>>>> On Mon, Feb 20, 2017 at 7:26 AM, Masahiko Sawada <sawada.m...@gmail.com> 
>>>>> wrote:
>>>>>> On Fri, Feb 17, 2017 at 3:41 AM, Robert Haas <robertmh...@gmail.com> 
>>>>>> wrote:
>>>>>>> On Thu, Feb 16, 2017 at 6:17 AM, Simon Riggs <si...@2ndquadrant.com> 
>>>>>>> wrote:
>>>>>>>> On 15 February 2017 at 08:07, Masahiko Sawada <sawada.m...@gmail.com> 
>>>>>>>> wrote:
>>>>>>>>> It's a bug. Attached latest version patch, which passed make check.
>>>>>>>> 2. The current btree vacuum code requires 2 vacuums to fully reuse
>>>>>>>> half-dead pages. So skipping an index vacuum might mean that second
>>>>>>>> index scan never happens at all, which would be bad.
>>>>>>> Maybe.  If there are a tiny number of those half-dead pages in a huge
>>>>>>> index, it probably doesn't matter.  Also, I don't think it would never
>>>>>>> happen, unless the table just never gets any more updates or deletes -
>>>>>>> but that case could also happen today.  It's just a matter of
>>>>>>> happening less frequently.
>>>>> Yeah thats right and I am not sure if it is worth to perform a
>>>>> complete pass to reclaim dead/deleted pages unless we know someway
>>>>> that there are many such pages.
>>>> Agreed.... which is why
>>>> On 16 February 2017 at 11:17, Simon Riggs <si...@2ndquadrant.com> wrote:
>>>>> I suggest that we store the number of half-dead pages in the metapage
>>>>> after each VACUUM, so we can decide whether to skip the scan or not.
>>>>> Also, I think we do reclaim the
>>>>> complete page while allocating a new page in btree.
>>>> That's not how it works according to the README at least.
>>> I am referring to code (_bt_getbuf()->if (_bt_page_recyclable(page))),
>>> won't that help us in reclaiming the space?
>> Not unless the README is incorrect, no.
> Just to ensure that we both have the same understanding, let me try to
> write what I understand about this reclaim algorithm.  AFAIU, in the
> first pass vacuum will mark the half dead pages as Deleted and in the
> second pass, it will record such pages as free in FSM so that they can
> be reused as new pages when the indexam asked for a new block instead
> of extending the index relation.  Now, if we introduce this new GUC,
> then there are chances that sometimes we skip the second pass where it
> would not have been skipped.
> Note that we do perform the second pass in the same vacuum cycle when
> index has not been scanned for deleting the tuples as per below code:

The first pass uses cycle id given by _bt_start_vacuum, but the second
pass always uses cycle id 0.

> btvacuumcleanup()
> {
> ..
> if (stats == NULL)
> {
> stats = (IndexBulkDeleteResult *) palloc0(sizeof(IndexBulkDeleteResult));
> btvacuumscan(info, stats, NULL, NULL, 0);
> ..
> }
> In above code stats won't be NULL, if the vacuum has scanned index for
> deleting tuples (btbulkdelete).  So, based on this I think it will
> skip scanning the index (or recycling pages marked as deleted) in the
> second vacuum only when there are no dead tuple removals in that
> vacuum.  Do we agree till here?


> I understand that there could be some delay in reclaiming dead pages
> but do you think it is such a big deal that we completely scan the
> index for such cases or even try to change the metapage format?

IIUC, I think that we need to have the number of half-dead pages in meta page.
Isn't it a problem that the freespace map of btree index is not
vacuumed if all vacuums skip the second pass?


Masahiko Sawada
NTT Open Source Software Center

Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:

Reply via email to