Re: Write visibility map during CLUSTER/VACUUM FULL

On Thu, Sep 12, 2019 at 4:55 PM Alexander Korotkov
<a(dot)korotkov(at)postgrespro(dot)ru> wrote:
> On Wed, Sep 11, 2019 at 3:30 PM Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
> > On Sun, Sep 1, 2019 at 1:37 PM Alexander Korotkov
> > <a(dot)korotkov(at)postgrespro(dot)ru> wrote:
> > > I found it weird that CLUSTER/VACUUM FULL don't write visibility map.
> > > Attached patch implements writing visibility map in
> > > heapam_relation_copy_for_cluster().
> > >
> > > I've studied previous attempt to implement this [1]. The main problem
> > > of that attempt was usage of existing heap_page_is_all_visible() and
> > > visibilitymap_set() functions. These functions works through buffer
> > > manager, while heap rewriting is made bypass buffer manager.
> > >
> > > In my patch visibility map pages are handled in the same way as heap
> > > pages are.
> > >
> >
> > I haven't studied this patch in detail, but while glancing I observed
> > that this doesn't try to sync the vm pages as we do for heap pages in
> > the end (during end_heap_rewrite). Am I missing something?
>
> You're not missed anything. Yes, VM need sync. Will fix this. And I
> just noticed I need a closer look to what is going on with TOAST.

Attached patch syncs VM during end_heap_rewrite().

However, VM for TOAST still isn't read. It appear to be much more
difficult to write VM for TOAST, because it's written by insertion
tuples one-by-one. Despite it seems to fill TOAST heap pages
sequentially (assuming no FSM exists yet), it's quite hard to handle
page-switching event with reasonable level of abstraction.
Nevertheless, I find this patch useful in current shape. Even if we
don't write VM for TOAST, it's still useful to do for regular heap.
Additionally, one of key advantages of having VM is index-only scan,
which don't work for TOAST anyway.

------
Alexander Korotkov
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company