> On 13 Feb 2019, at 19:27, Andres Freund <andres@anarazel.de> wrote:
>
> Hi,
>
> Turns out in portions of the regression tests a good chunk of the
> runtime is inside AddNewAttributeTuples() and
> recordMultipleDependencies()'s heap insertions. Looking at a few
> profiles I had lying around I found that in some production cases
> too. ISTM we should use heap_multi_insert() for both, as the source
> tuples ought to be around reasonably comfortably.
>
> For recordMultipleDependencies() it'd obviously better if we collected
> all dependencies for new objects, rather than doing so separately. Right
> now e.g. the code for a new table looks like:
>
> recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
>
> recordDependencyOnOwner(RelationRelationId, relid, ownerid);
>
> recordDependencyOnNewAcl(RelationRelationId, relid, 0, ownerid, relacl);
>
> recordDependencyOnCurrentExtension(&myself, false);
>
> if (reloftypeid)
> {
> referenced.classId = TypeRelationId;
> referenced.objectId = reloftypeid;
> referenced.objectSubId = 0;
> recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
> }
>
> and it'd obviously be more efficient to do that once if we went to using
> heap_multi_insert() in the dependency code. But I suspect even if we
> just used an extended API in AddNewAttributeTuples() (for the type /
> collation dependencies), it'd be a win.
When a colleague was looking at heap_multi_insert in the COPY codepath I
remembered this and took a stab at a WIP patch inspired by this email, while
not following it to the letter. It’s not going the full route of collecting
all the dependencies for creating a table, but adding ways to perform
multi_heap_insert in the existing codepaths as it seemed like a good place to
start.
It introduces a new function CatalogMultiInsertWithInfo which takes a set of
slots for use in heap_multi_insert, used from recordMultipleDependencies and
InsertPgAttributeTuples (which replace calling InsertPgAttributeTuple
repeatedly). The code is still a WIP with some kludges, following the show-
early philosophy.
It passes make check and some light profiling around regress suites indicates
that it does improve a bit by reducing the somewhat costly calls. Is this
along the lines of what you were thinking or way off?
cheers ./daniel