Thread: parallel append vs. simple UNION ALL

As I mentioned in the commit message for the Parallel Append commit
(ab72716778128fb63d54ac256adf7fe6820a1185), it's kind of sad that this
doesn't work with UNION ALL queries, which are an obvious candidate
for such parallelization.  It turns out that it actually does work to
a limited degree: assuming that the UNION ALL query can be converted
to a simple appendrel, it can consider a parallel append of
non-partial paths only.  The attached patch lets it consider a
parallel append of partial paths by doing the following things:

1. Teaching set_subquery_pathlist to create *partial* SubqueryScan
paths as well as non-partial ones.
2. Teaching grouping_planner to create partial paths for the final rel
if not at the outermost query level.
3. Modifying finalize_plan to allow the gather_param to be passed
across subquery boundaries.

#3 is the only part I'm really unsure about; the other stuff looks
pretty cut and dried.

I have a draft patch that handles the case where the union can't be
converted to a simple appendrel, too, but that's not quite baked
enough to post yet.

For those for whom the above may be too technical to follow, here's an example:

pgbench -i 40
explain (costs off) select a.bid from pgbench_accounts a,
pgbench_branches b where a.bid = b.bid and aid % 1000 = 0 union all
select a.bid from pgbench_accounts a where aid % 1000 = 0;

Unpatched:

 Append
   ->  Gather
         Workers Planned: 2
         ->  Hash Join
               Hash Cond: (a.bid = b.bid)
               ->  Parallel Seq Scan on pgbench_accounts a
                     Filter: ((aid % 1000) = 0)
               ->  Hash
                     ->  Seq Scan on pgbench_branches b
   ->  Gather
         Workers Planned: 2
         ->  Parallel Seq Scan on pgbench_accounts a_1
               Filter: ((aid % 1000) = 0)

Patched:

 Gather
   Workers Planned: 2
   ->  Parallel Append
         ->  Hash Join
               Hash Cond: (a.bid = b.bid)
               ->  Parallel Seq Scan on pgbench_accounts a
                     Filter: ((aid % 1000) = 0)
               ->  Hash
                     ->  Seq Scan on pgbench_branches b
         ->  Parallel Seq Scan on pgbench_accounts a_1
               Filter: ((aid % 1000) = 0)

In this particular case the change doesn't buy very much, but the
second plan is better because avoid shutting down one set of workers
and starting a new set.  That's more efficient, plus it allows the two
branches to be worked in parallel rather than serially.  On a small
enough scale factor, even without the patch, you get this...

 Gather
   Workers Planned: 2
   ->  Parallel Append
         ->  Nested Loop
               Join Filter: (a.bid = b.bid)
               ->  Seq Scan on pgbench_branches b
               ->  Seq Scan on pgbench_accounts a
                     Filter: ((aid % 1000) = 0)
         ->  Seq Scan on pgbench_accounts a_1
               Filter: ((aid % 1000) = 0)

...but that's not good because now we have regular sequential scans
instead of partial sequential scans.

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

On Sat, Dec 23, 2017 at 4:53 PM, Robert Haas <robertmhaas@gmail.com> wrote:
> As I mentioned in the commit message for the Parallel Append commit
> (ab72716778128fb63d54ac256adf7fe6820a1185), it's kind of sad that this
> doesn't work with UNION ALL queries, which are an obvious candidate
> for such parallelization.  It turns out that it actually does work to
> a limited degree: assuming that the UNION ALL query can be converted
> to a simple appendrel, it can consider a parallel append of
> non-partial paths only.  The attached patch lets it consider a
> parallel append of partial paths ...

Here's an extended series of patches that now handles both the simple
UNION ALL case (where we flatten it) and the unflattened case:

0001 is pretty much the same as the subquery-smarts.patch file I
attached to the previous email.  I don't see much reason not to go
ahead and commit this, although it could use a test case.  It makes
the simple/flattened case work.  After some study I think that the
gather-parameter handling is correct, although if somebody felt like
reviewing that portion especially I wouldn't say no.

0002 rewrites recurse_union_children to work iteratively rather than
recursively and renames it to plan_union_children.  This probably
isn't 100% necessary, but it seems to me that the resulting code is
easier to understand, and it reduces the risk of blowing out the
stack.  There should be no user-visible behavior change.

0003 rewrites the setop planner to create a separate upper rel for
each stage of setop planning and uses them to return paths instead of
returning paths directly.  This is necessary preparatory work for
anything that wants to consider multiple possible paths for queries
that go through the full setop planner, but it shouldn't have any
visible impact all by itself.

0004 causes generate_union_path() to consider both the traditional
method and also Gather -> Parallel Append -> [partial path for each
subquery].  This is still a bit rough around the edges and there's a
lot more that could be done here, but I'm posting what I have for now
in the (perhaps vain) hope of getting some feedback.  With this, you
can use Parallel Append for the UNION ALL step of a query like SELECT
.. UNION ALL .. SELECT ... EXCEPT SELECT ...

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

> On Sat, Feb 24, 2018 at 2:55 AM, Robert Haas <robertmhaas@gmail.com> wrote:
>>
>> 0001 is pretty much the same as the subquery-smarts.patch file I
>> attached to the previous email.  I don't see much reason not to go
>> ahead and commit this, although it could use a test case.  It makes
>> the simple/flattened case work.  After some study I think that the
>> gather-parameter handling is correct, although if somebody felt like
>> reviewing that portion especially I wouldn't say no.

I had a look at 0001 patch. Other than the issue raised by Rajkumar,
it looks good functionally.

Regarding the finalize_plan() changes, I see that in the patch, the
Gather rescan param is now included in the valid_params while calling
finalize_plan() for the SubqueryScan, which looks correct. But I was
thinking that instead of doing that just before the recursive
finalize_plan(), it looks better if we do that at the initial section
of finalize_plan(). We already add initSetParam to valid_params. There
itself we can also add gather_params. Something like this :

@@ -2314,6 +2314,10 @@ finalize_plan(PlannerInfo *root, Plan *plan,
        if (initSetParam)
                valid_params = bms_union(valid_params, initSetParam);

+       /* Same applies for Gather rescan param */
+       if (gather_param >= 0)
+               valid_params = bms_add_member(valid_params, gather_param);


On 27 February 2018 at 16:51, Rajkumar Raghuwanshi
<rajkumar.raghuwanshi@enterprisedb.com> wrote:
>
>
> I have applied 0001 on pg-head, and while playing with regression tests, it
> crashed with below test case.
>
> postgres=# SET min_parallel_table_scan_size=0;
> SET
> postgres=# SELECT * FROM information_schema.foreign_data_wrapper_options
> ORDER BY 1, 2, 3;
> server closed the connection unexpectedly
>     This probably means the server terminated abnormally
>     before or while processing the request.
> The connection to the server was lost. Attempting reset: Failed.

This is happening because there is a ProjectionPath as one of the
subpaths, and that is not parallel safe.

 Sort
   Sort Key: ((current_database())::information_schema.sql_identifier),
((w.fdwname)::information_schema.sql_identifier),
((((pg_options_to_table(w.fdwoptions))).option_name)::information_schema.sql_identifier)
   ->  Result
         ->  ProjectSet
               ->  Hash Join
                     Hash Cond: (w.fdwowner = u.oid)
                     ->  Seq Scan on pg_foreign_data_wrapper w
                           Filter: (pg_has_role(fdwowner,
'USAGE'::text) OR has_foreign_data_wrapper_privilege(oid,
'USAGE'::text))
                     ->  Hash
                           ->  Seq Scan on pg_authid u

In grouping_planner() where partial paths are generated for final_rel,
we can skip non-parallel-safe paths.

-- 
Thanks,
-Amit Khandekar
EnterpriseDB Corporation
The Postgres Database Company

On Sat, Feb 24, 2018 at 2:55 AM, Robert Haas <robertmhaas@gmail.com> wrote:

>
> Here's an extended series of patches that now handles both the simple
> UNION ALL case (where we flatten it) and the unflattened case:
>

The patches look clean. I particularly looked at 0003.

patch 0001
+    /*
+     * Generate partial paths for final_rel, too, if outer query levels might
+     * be able to make use of them.
+     */
I am not able to understand the construct esp. the if clause. Did you want to
say "... if there are outer query levels. Those might ..." or something like
that?

0002
                (op->all == top_union->all || op->all) &&
This isn't really your change.  Checking
op->all is cheaper than checking equality, so may be we should check that first
and take advantage of short-circuit condition evaluation. If we do that above
condition reduces to (op->all || !top_union->all) which is two boolean
conditions, even cheaper. But may be the second optimization is not worth the
loss of readability.

"identically-propertied UNIONs" may be "UNIONs with identical properties".

0003
Probably we want to rename generate_union_path() as generate_union_rel() or
generate_union_paths() since the function doesn't return a path anymore.
Similarly for generate_nonunion_path().

In recurse_set_operations()
-        return NULL;            /* keep compiler quiet */
This line is deleted and instead rel is initialized to NULL. That way we loose
any chance to detect a future bug because of a block leaving rel uninitialized
through compiler warning. May be we should replace "return NULL" with "rel =
NULL", which will not be executed because of the error.

+    /* Build path list and relid set. */
+    foreach(lc, rellist)
+    {
With the changes in this patch, we could actually use add_paths_to_append_rel()
to create an append path. That function builds paths with different pathkeys,
parameterization (doesn't matter here) and also handles parallel append. So we
can avoid code duplication and also leverage more optimizations like using
MergeAppend instead of overall sort etc. But that function doesn't have ability
to add a final node like make_union_unique(). A similar requirement has arisen
in partition-wise join where we need to add a final node for finalising
aggregate on top of paths created by add_paths_to_append_rel().  May be we can
change that function to return a list of paths, which are then finalized by the
caller and added to "append" rel. But I don't think doing all that is in the
scope of this patch set.

0004
+        if (!op->all)
+            ppath = make_union_unique(op, ppath, tlist, root);
We could probably push the grouping/sorting down to the parallel workers. But
again not part of this patchset.

-- 
Best Wishes,
Ashutosh Bapat
EnterpriseDB Corporation
The Postgres Database Company

On Tue, Feb 27, 2018 at 6:21 AM, Rajkumar Raghuwanshi
<rajkumar.raghuwanshi@enterprisedb.com> wrote:
> I have applied 0001 on pg-head, and while playing with regression tests, it
> crashed with below test case.
>
> postgres=# SET min_parallel_table_scan_size=0;
> SET
> postgres=# SELECT * FROM information_schema.foreign_data_wrapper_options
> ORDER BY 1, 2, 3;
> server closed the connection unexpectedly
>     This probably means the server terminated abnormally
>     before or while processing the request.
> The connection to the server was lost. Attempting reset: Failed.

Hmm, nice catch.  I think part of the problem here is that commit
69f4b9c85f168ae006929eec44fc44d569e846b9, wherein Andres introduced
the ProjectSet node, didn't really do the right thing in terms of
testing parallel-safety.  Before that commit, is_parallel_safe(root,
(Node *) scanjoin_target->exprs)) was really testing whether the tlist
produced during the scan/join phase was parallel-safe.  However, after
that commit, scanjoin_target->exprs wasn't really the final target for
the scan/join phase any more; instead, it was the first of possibly
several targets computed by split_pathtarget_at_srfs().  Really, the
right thing to do is to test the *last* element in that list for
parallel-safety, but as the code stands we end up testing the *first*
element.  So, if there's a parallel-restricted item in the target list
(this query ends up putting a CoerceToDomain in the target list, which
we currently consider parallel-restricted), it looks we can
nevertheless end up trying to project it in what is supposed to be a
partial path.

There are a large number of cases where this doesn't end up mattering
because the next upper_rel created will not get marked
consider_parallel because its target list will also contain the same
parallel-restricted construct, and therefore the partial paths
generated for the scan/join rel will never get used -- except to
generate Gather/Gather Merge paths, which has already happened; but
that step didn't know about the rest of the scan/join targets either,
so it won't have used them.  However, both create_distinct_paths() and
the code in grouping_planner that populates final_rel assume that they
don't need to retest the target for parallel-safety because no
projection is done at those levels; they just inherit the
parallel-safety marking of the input rel, so in those cases if the
input rel's marking is wrong the result is populated upward.

There's another way final_rel->consider_parallel can be wrong, too: if
the FROM-list is empty, then we create a join rel and set its
consider_parallel flag without regard to the parallel-safety of the
target list.  There are comments in query_planner() says that this
will be dealt with "later", but this seems not to be true. (Before
Tom's commit da1c91631e3577ea5818f855ebb5bd206d559006, the comments
simply ignored the question of whether a check was needed here, but
Tom seems to have inserted an incorrect justification for the
already-wrong code.)

I'm not sure to what degree, if at all, any of these problems are
visible given that we don't use final_rel->consider_parallel for much
of anything.  Certainly, it gets much easier to trigger a problem with
0001 applied, as the test case shows.  But I'm not entirely convinced
that there's no problem even without that.  It seems like every upper
rel that is setting its consider_parallel flag based on the first
element of some list of targets rather than the last is potentially
vulnerable to ending up with the wrong answer, and I'm afraid that
might have some adverse consequence that I haven't quite pinned down
yet.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

On Wed, Mar 7, 2018 at 5:36 AM, Rajkumar Raghuwanshi
<rajkumar.raghuwanshi@enterprisedb.com> wrote:
> With 0001 applied on PG-head, I got reference leak warning and later a
> server crash.
> this crash is reproducible with enable_parallel_append=off also.
> below is the test case to reproduce this.

New patches attached, fixing all 3 of the issues you reported:

0001 is a new patch to fix the incorrect parallel safety marks on
upper relations.  I don't know of a visible effect of this patch by
itself, but there might be one.

0002 is the same as the old 0001, but I made a fix in
SS_charge_for_initplans() which fixed your most recent crash report.
Either this or the previous change also fixed the crash you saw when
using tab-completion.  Also, I added some test cases based on your
failing examples.

0003-0005 are the same as the old 0002-0004.

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

On Thu, Mar 8, 2018 at 2:46 AM, Rajkumar Raghuwanshi
<rajkumar.raghuwanshi@enterprisedb.com> wrote:
> On Thu, Mar 8, 2018 at 12:27 AM, Robert Haas <robertmhaas@gmail.com> wrote:
>>
>> New patches attached, fixing all 3 of the issues you reported:
>
> Thanks. new patches applied cleanly on head and fixing all reported issue.

Great.  Committed 0001.  Are you planning any further testing of this
patch series?

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

On Thu, Mar 1, 2018 at 8:30 AM, Ashutosh Bapat
<ashutosh.bapat@enterprisedb.com> wrote:
> The patches look clean. I particularly looked at 0003.
>
> patch 0001
> +    /*
> +     * Generate partial paths for final_rel, too, if outer query levels might
> +     * be able to make use of them.
> +     */
> I am not able to understand the construct esp. the if clause. Did you want to
> say "... if there are outer query levels. Those might ..." or something like
> that?

Well, that's what I meant, but I didn't think it was necessary to
spell it out in quite that much detail.

> 0002
>                 (op->all == top_union->all || op->all) &&
> This isn't really your change.  Checking
> op->all is cheaper than checking equality, so may be we should check that first
> and take advantage of short-circuit condition evaluation. If we do that above
> condition reduces to (op->all || !top_union->all) which is two boolean
> conditions, even cheaper. But may be the second optimization is not worth the
> loss of readability.

I doubt this makes any difference.  The compiler should be smart
enough to do this the same way regardless of exactly how we write it,
and if it's not, it can't make more than a few instructions worth of
difference.  This code is not nearly performance-critical enough for
that to matter.  Also, it's not the job of this patch to whack this
around.

> "identically-propertied UNIONs" may be "UNIONs with identical properties".

Likewise, I didn't write those words, so I don't plan to change them
just because I might have written them differently.

> 0003
> Probably we want to rename generate_union_path() as generate_union_rel() or
> generate_union_paths() since the function doesn't return a path anymore.
> Similarly for generate_nonunion_path().

Good point.  Changed.

> In recurse_set_operations()
> -        return NULL;            /* keep compiler quiet */
> This line is deleted and instead rel is initialized to NULL. That way we loose
> any chance to detect a future bug because of a block leaving rel uninitialized
> through compiler warning. May be we should replace "return NULL" with "rel =
> NULL", which will not be executed because of the error.

I don't think this is really a problem.  If some code path fails to
initialize rel, it's going to crash when postprocess_setop_rel() calls
set_cheapest().  Any warning the compiler gives is more likely to be a
false-positive than an actual problem.

> +    /* Build path list and relid set. */
> +    foreach(lc, rellist)
> +    {
> With the changes in this patch, we could actually use add_paths_to_append_rel()
> to create an append path. That function builds paths with different pathkeys,
> parameterization (doesn't matter here) and also handles parallel append. So we
> can avoid code duplication and also leverage more optimizations like using
> MergeAppend instead of overall sort etc. But that function doesn't have ability
> to add a final node like make_union_unique(). A similar requirement has arisen
> in partition-wise join where we need to add a final node for finalising
> aggregate on top of paths created by add_paths_to_append_rel().  May be we can
> change that function to return a list of paths, which are then finalized by the
> caller and added to "append" rel. But I don't think doing all that is in the
> scope of this patch set.

Yeah, I thought about all of that and came to similar conclusions.

> 0004
> +        if (!op->all)
> +            ppath = make_union_unique(op, ppath, tlist, root);
> We could probably push the grouping/sorting down to the parallel workers. But
> again not part of this patchset.

Yep.  There's a lot more work that could be done to improve setop
planning, but I think getting even this much done for v11 would be a
significant step forward.

Updated patches attached.

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company<div class="gmail_extra"><br><div
class="gmail_quote">On Thu, Mar 1, 2018 at 8:30 AM, Ashutosh Bapat
<span dir="ltr"><<a href="mailto:ashutosh.bapat@enterprisedb.com"
target="_blank">ashutosh.bapat@enterprisedb.com</a>></span>
wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0
.8ex;border-left:1px #ccc solid;padding-left:1ex"><span class="">On
Sat, Feb 24, 2018 at 2:55 AM, Robert Haas <<a
href="mailto:robertmhaas@gmail.com">robertmhaas@gmail.com</a>>
wrote:<br>
<br>
><br>
</span><span class="">> Here's an extended series of patches that
now handles both the simple<br>
> UNION ALL case (where we flatten it) and the unflattened case:<br>
><br>
<br>
</span>The patches look clean. I particularly looked at 0003.<br>
<br>
patch 0001<br>
+    /*<br>
+     * Generate partial paths for final_rel, too, if
outer query levels might<br>
+     * be able to make use of them.<br>
+     */<br>
I am not able to understand the construct esp. the if clause. Did you
want to<br>
say "... if there are outer query levels. Those might ..." or something like<br>
that?<br>
<br>
0002<br>
                (op->all ==
top_union->all || op->all) &&<br>
This isn't really your change.  Checking<br>
op->all is cheaper than checking equality, so may be we should
check that first<br>
and take advantage of short-circuit condition evaluation. If we do
that above<br>
condition reduces to (op->all || !top_union->all) which is two boolean<br>
conditions, even cheaper. But may be the second optimization is not
worth the<br>
loss of readability.<br>
<br>
"identically-propertied UNIONs" may be "UNIONs with identical properties".<br>
<br>
0003<br>
Probably we want to rename generate_union_path() as generate_union_rel() or<br>
generate_union_paths() since the function doesn't return a path anymore.<br>
Similarly for generate_nonunion_path().<br>
<br>
In recurse_set_operations()<br>
-        return NULL;       
    /* keep compiler quiet */<br>
This line is deleted and instead rel is initialized to NULL. That way
we loose<br>
any chance to detect a future bug because of a block leaving rel
uninitialized<br>
through compiler warning. May be we should replace "return NULL" with "rel =<br>
NULL", which will not be executed because of the error.<br>
<br>
+    /* Build path list and relid set. */<br>
+    foreach(lc, rellist)<br>
+    {<br>
With the changes in this patch, we could actually use
add_paths_to_append_rel()<br>
to create an append path. That function builds paths with different
pathkeys,<br>
parameterization (doesn't matter here) and also handles parallel
append. So we<br>
can avoid code duplication and also leverage more optimizations like using<br>
MergeAppend instead of overall sort etc. But that function doesn't
have ability<br>
to add a final node like make_union_unique(). A similar requirement
has arisen<br>
in partition-wise join where we need to add a final node for finalising<br>
aggregate on top of paths created by add_paths_to_append_rel(). 
May be we can<br>
change that function to return a list of paths, which are then
finalized by the<br>
caller and added to "append" rel. But I don't think doing all that is in the<br>
scope of this patch set.<br>
<br>
0004<br>
+        if (!op->all)<br>
+            ppath =
make_union_unique(op, ppath, tlist, root);<br>
We could probably push the grouping/sorting down to the parallel
workers. But<br>
again not part of this patchset.<br>
<span class="HOEnZb"><font color="#888888"><br>
--<br>
Best Wishes,<br>
Ashutosh Bapat<br>
</font></span><div class="HOEnZb"><div class="h5">EnterpriseDB Corporation<br>
The Postgres Database Company<br>
</div></div></blockquote></div><br><br clear="all"><div><br></div>--
<br><div class="gmail_signature"
data-smartmail="gmail_signature">Robert Haas<br>EnterpriseDB: <a
href="http://www.enterprisedb.com"
target="_blank">http://www.enterprisedb.com</a><br>The Enterprise
PostgreSQL Company</div>
</div>

On Fri, Mar 9, 2018 at 1:28 AM, Robert Haas <robertmhaas@gmail.com> wrote:
>
>> 0003
>> Probably we want to rename generate_union_path() as generate_union_rel() or
>> generate_union_paths() since the function doesn't return a path anymore.
>> Similarly for generate_nonunion_path().
>
> Good point.  Changed.

It looks like it was not changed in all the places. make falied. I
have fixed all the instances of these two functions in the attached
patchset (only 0003 changes). Please check.

-- 
Best Wishes,
Ashutosh Bapat
EnterpriseDB Corporation
The Postgres Database Company

On Tue, Mar 13, 2018 at 12:35 AM, Ashutosh Bapat
<ashutosh.bapat@enterprisedb.com> wrote:
> It looks like it was not changed in all the places. make falied. I
> have fixed all the instances of these two functions in the attached
> patchset (only 0003 changes). Please check.

Oops.  Thanks.

I'm going to go ahead and commit 0001 here.  Any more thoughts on the rest?

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

On Wed, Mar 14, 2018 at 2:09 AM, Robert Haas <robertmhaas@gmail.com> wrote:
> On Tue, Mar 13, 2018 at 12:35 AM, Ashutosh Bapat
> <ashutosh.bapat@enterprisedb.com> wrote:
>> It looks like it was not changed in all the places. make falied. I
>> have fixed all the instances of these two functions in the attached
>> patchset (only 0003 changes). Please check.
>
> Oops.  Thanks.
>
> I'm going to go ahead and commit 0001 here.  Any more thoughts on the rest?

Nope. I am good with the patchset.

-- 
Best Wishes,
Ashutosh Bapat
EnterpriseDB Corporation
The Postgres Database Company

On Fri, Mar 16, 2018 at 7:35 AM, Rajkumar Raghuwanshi
<rajkumar.raghuwanshi@enterprisedb.com> wrote:
> On Fri, Mar 9, 2018 at 1:04 AM, Robert Haas <robertmhaas@gmail.com> wrote:
>> Great.  Committed 0001.  Are you planning any further testing of this
>> patch series?
>
> Sorry I missed the mail.
> Yes, I have further tested patches and find no more issues.

OK, thanks to both you and Ashutosh Bapat.  Committed 0002 and 0003.

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

On Mon, Mar 19, 2018 at 11:57 AM, Robert Haas <robertmhaas@gmail.com> wrote:
> On Fri, Mar 16, 2018 at 7:35 AM, Rajkumar Raghuwanshi
> <rajkumar.raghuwanshi@enterprisedb.com> wrote:
>> On Fri, Mar 9, 2018 at 1:04 AM, Robert Haas <robertmhaas@gmail.com> wrote:
>>> Great.  Committed 0001.  Are you planning any further testing of this
>>> patch series?
>>
>> Sorry I missed the mail.
>> Yes, I have further tested patches and find no more issues.
>
> OK, thanks to both you and Ashutosh Bapat.  Committed 0002 and 0003.

And now committed 0004.  This area could use significantly more work,
but I think it's better to have this much than not.

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company