Re: About "Our CLUSTER implementation is pessimal" patch - Mailing list pgsql-hackers

From Leonardo F
Subject Re: About "Our CLUSTER implementation is pessimal" patch
Date
Msg-id 556775.51248.qm@web29015.mail.ird.yahoo.com
Whole thread Raw
In response to Re: About "Our CLUSTER implementation is pessimal" patch  (Leonardo F <m_lists@yahoo.it>)
Responses Re: About "Our CLUSTER implementation is pessimal" patch
List pgsql-hackers
Anyone? I'd like some feedback before moving on to do the seq scan + sort in those
CLUSTER cases where "use_index_scan" returns false...



----- Messaggio originale -----
> Da: Leonardo F <m_lists@yahoo.it>
> A: pgsql-hackers@postgresql.org
> Inviato: Mer 20 gennaio 2010, 18:48:00
> Oggetto: Re: [HACKERS] About "Our CLUSTER implementation is pessimal" patch
>
> > I read the thread "Our CLUSTER implementation is pessimal"
> > http://archives.postgresql.org/pgsql-hackers/2008-08/msg01371.php .
> >
> > I was going to try to add the proper cost_index/cost_sort calls to decide
> which
> > "path" should be executed, as in:
> >
> > http://archives.postgresql.org/pgsql-hackers/2008-09/msg00517.php
>
> I think I got something up and running to check if a table scan + sort is
> supposed
> to be faster than an index scan for a certain CLUSTER operation.
>
> The way I did it is (I guess...) wrong: I created the elements needed by
> get_relation_info, create_seqscan_path, create_index_path, cost_sort.
>
> It has been, obviously, a trial and error approach: I added the member values as
> soon as one function call crashed... and I bet I didn't get all the corner
> cases.
> Is there any better way of doing it?
>
> Leonardo
>
> (this is called in copy_heap_data to decide which path to choose:)
>
> static bool use_index_scan(Oid tableOid, Oid indexOid)
> {
> RelOptInfo *rel;
> PlannerInfo *root;
> Query *query;
> PlannerGlobal *glob;
> Path *seqAndSortPath;
> IndexPath *indexPath;
> RangeTblEntry *rte;
>
> rel = makeNode(RelOptInfo);
> rel->reloptkind = RELOPT_BASEREL;
> rel->relid = 1;
> rel->rtekind = RTE_RELATION;
>
> /* needed by get_relation_info */
> glob = makeNode(PlannerGlobal);
>
> /* needed by get_relation_info: */
> query = makeNode(Query);
> query->resultRelation = 0;
>
> root = makeNode(PlannerInfo);
>
> root->parse = query;
> root->glob = glob;
>
> get_relation_info(root, tableOid, false, rel);
> seqAndSortPath = create_seqscan_path(NULL, rel);
>
> rel->rows = rel->tuples;
>
> rte = makeNode(RangeTblEntry);
> rte->rtekind = RTE_RELATION;
> rte->relid = tableOid;
>
> root->simple_rel_array_size = 2;
> root->simple_rte_array = (RangeTblEntry **)
> palloc0(root->simple_rel_array_size * sizeof(RangeTblEntry *));
> root->simple_rte_array[1] = rte;
>
> root->total_table_pages = rel->pages;
>
> indexPath = create_index_path(root,
> (IndexOptInfo*)(list_head(rel->indexlist)->data.ptr_value), NULL, NULL,
> ForwardScanDirection, NULL);
> cost_sort(seqAndSortPath, root, NULL, seqAndSortPath->total_cost, rel->tuples,
> rel->width, -1);
>
> return indexPath->path.total_cost < seqAndSortPath->total_cost;
> }





pgsql-hackers by date:

Previous
From: Boszormenyi Zoltan
Date:
Subject: Re: lock_timeout GUC patch
Next
From: KaiGai Kohei
Date:
Subject: Re: Largeobject Access Controls (r2460)