2018年12月29日(土) 1:44 Tom Lane <tgl@sss.pgh.pa.us>:
>
> Kohei KaiGai <kaigai@heterodb.com> writes:
> > I've investigated a crash report of PG-Strom for a few days, then I doubt
> > add_partial_path() can unexpectedly release dominated old partial path
> > but still referenced by other Gather node, and it leads unexpected system
> > crash.
>
> Hm. This seems comparable to the special case in plain add_path, where it
> doesn't attempt to free IndexPaths because of the risk that they're still
> referenced. So maybe we should just drop the pfree here.
>
> However, first I'd like to know why this situation is arising in the first
> place. To have the situation you're describing, we'd have to have
> attempted to make some Gather paths before we have all the partial paths
> for the relation they're for. Why is that a good thing to do? It seems
> like such Gathers are necessarily being made with incomplete information,
> and we'd be better off to fix things so that none are made till later.
>
Because of the hook location, Gather-node shall be constructed with built-in
and foreign partial scan node first, then extension gets a chance to add its
custom paths (partial and full).
At the set_rel_pathlist(), set_rel_pathlist_hook() is invoked next to the
generate_gather_paths(). Even if extension adds some partial paths later,
the first generate_gather_paths() has to generate Gather node based on
incomplete information.
If we could ensure Gather node shall be made after all the partial nodes
are added, it may be a solution for the problem.
Of course, relocation of the hook may have a side-effect. Anyone may
expect the pathlist contains all the path-node including Gather node, for
editorialization of the pathlist.
Thanks,
--
HeteroDB, Inc / The PG-Strom Project
KaiGai Kohei <kaigai@heterodb.com>