Re: BUG #14808: V10-beta4, backend abort

On Fri, Sep 15, 2017 at 9:45 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Attached is a draft patch for this.

Some initial feedback:

Compiles cleanly and make check-world passes here.

with wcte as (insert into table1 values (42))
   insert into table2 values ('hello world');
 NOTICE:  trigger = table2_trig, new table = ("hello world")
 NOTICE:  trigger = table1_trig, new table = (42)
+with wcte as (insert into table1 values (43))
+  insert into table1 values (44);
+NOTICE:  trigger = table1_trig, new table = (43), (44)

The effects of multiple ModifyTable nodes on the same table are
merged.  Good.  I doubt anyone will notice but it might warrant a note
somewhere that there is a user-visible change here: previously if you
did this (without TTs) your trigger would have fired twice.

+create trigger my_table_col_update_trig
+  after update of b on my_table referencing new table as new_table
+  for each statement execute procedure dump_insert();
+ERROR:  transition tables cannot be specified for triggers with column lists

Potential SQL standard non-compliance avoided.  Good.

 delete from refd_table where length(b) = 3;
-NOTICE:  trigger = trig_table_delete_trig, old table = (2,"two a"), (2,"two b")
-NOTICE:  trigger = trig_table_delete_trig, old table = (11,"one a"),
(11,"one b")
+NOTICE:  trigger = trig_table_delete_trig, old table = (2,"two a"),
(2,"two b"), (11,"one a"), (11,"one b")

The effects of fk cascade machinery are merged as discussed.  Good, I
think, but I'm a bit confused about how this works when the cascading
operation also fires triggers.

All the other tests show no change in behaviour.  Good.

What is going on here?

=== setup ===
create or replace function dump_delete() returns trigger language plpgsql as
$$
  begin
    raise notice 'trigger = %, old table = %, depth = %',
                 TG_NAME,
                 (select string_agg(old_table::text, ', ' order by a)
from old_table),
                 pg_trigger_depth();
    return null;
  end;
$$;
create table foo (a int primary key, b int references foo(a) on delete cascade);
create trigger foo_s_trig after delete on foo
  referencing old table as old_table
  for each statement
  execute procedure dump_delete();
create trigger foo_r_trig after delete on foo
  referencing old table as old_table
  for each row
  execute procedure dump_delete();
insert into foo values (1, null), (2, 1);
===8<===

postgres=# delete from foo where a = 1;
NOTICE:  trigger = foo_r_trig, old table = (1,), depth = 1
NOTICE:  trigger = foo_s_trig, old table = (1,), depth = 1
NOTICE:  trigger = foo_r_trig, old table = (2,1), depth = 1
NOTICE:  trigger = foo_s_trig, old table = (2,1), depth = 1
NOTICE:  trigger = foo_s_trig, old table = <NULL>, depth = 1
DELETE 1

> Thomas Munro <thomas.munro@enterprisedb.com> writes:
>> 1.  Merging the transition tables when there are multiple wCTEs
>> referencing the same table.  Here's one idea:  Rename
>> MakeTransitionCaptureState() to GetTransitionCaptureState() and use a
>> hash table keyed by table OID in
>> afterTriggers.transition_capture_states[afterTriggers.query_depth] to
>> find the TCS for the given TriggerDesc or create it if not found, so
>> that all wCTEs find the same TransitionCaptureState object.  The all
>> existing callers continue to do what they're doing now, but they'll be
>> sharing TCSs appropriately with other things in the plan.  Note that
>> TransitionCaptureState already holds tuplestores for each operation
>> (INSERT, UPDATE, DELETE) so the OID of the table alone is a suitable
>> key for the hash table (assuming we are ignoring the column-list part
>> of the spec as you suggested).
>
> It seems unsafe to merge the TCS objects themselves, because the callers
> assume that they can munge the tcs_map and tcs_original_insert_tuple
> fields freely without regard for any other callers.  So as I have it,
> we still have a TCS for each caller, but the TCSes point at tuplestores
> that can be shared across multiple callers for the same event type.
> The tuplestores themselves are managed by the AfterTrigger data
> structures.  Also, because the TCS structs are just throwaway per-caller
> data, it's uncool to reference them in the trigger event lists.
> So I replaced ats_transition_capture with two pointers to the actual
> tuplestores.

Ok, that works and yeah it may be conceptually better.  Also, maybe
the tcs_original_insert_tuple as member of TCS is a bit clunky and
could be reconsidered later: I was trying to avoid widening a bunch of
function calls, but that might have been optimising for the wrong
thing...

> That bloats AfterTriggerSharedData a bit but I think it's
> okay; we don't expect a lot of those structs in a normal query.

Yeah, it was bloat avoidance that led me to find a way to use a single
pointer.  But at least it's in the shared part, so I don't think it
matters too much.

> I chose to make the persistent state (AfterTriggersTableData) independent
> for each operation type.  We could have done that differently perhaps, but
> it seemed more complicated and less likely to match the spec's semantics.

OK.  Your GetAfterTriggersTableData(Oid relid, CmdType cmdType) is
mirroring the spec's way of describing what happens (without the
column list).

+       foreach(lc, qs->tables)
+       {
+               table = (AfterTriggersTableData *) lfirst(lc);
+               if (table->relid == relid && table->cmdType == cmdType &&
+                       !table->closed)
+                       return table;
+       }

Yeah, my suggestion of a hash table was overkill.  (Maybe in future if
we change our rules around inheritance this could finish up being
searched for a lot of child tables; we can cross that bridge when we
come to it.)

I'm a little confused about the "closed" flag.  This seems to imply
that it's possible for there to be more than one separate tuplestore
for a given (table, operation) in a given trigger execution context.
Why is that OK?

> The INSERT ON CONFLICT UPDATE mess is handled by creating two separate
> TCSes with two different underlying AfterTriggersTableData structs.
> The insertion tuplestore sees only the inserted tuples, the update
> tuplestores see only the updated-pre-existing tuples.  That adds a little
> code to nodeModifyTable but it seems conceptually much cleaner.

OK.  The resulting behaviour is unchanged.

>> 3.  Merging the invocation after statement firing so that if you
>> updated the same table directly and also via a wCTE and also
>> indirectly via fk ON DELETE/UPDATE trigger then you still only get one
>> invocation of the after statement trigger.  Not sure exactly how...
>
> What I did here was to use the AfterTriggersTableData structs to hold
> a flag saying we'd already queued statement triggers for this rel and
> cmdType.  There's probably more than one way to do that, but this seemed
> convenient.

Seems good.  That implements the following (from whatever random draft
spec I have): "A statement-level trigger that is considered as
executed for a state change SC (in a given trigger execution context)
is not subsequently executed for SC."

> One thing I don't like too much about that is that it means there are
> cases where the single statement trigger firing would occur before some
> AFTER ROW trigger firings.  Not sure if we promise anything about the
> ordering in the docs.  It looks quite expensive/complicated to try to
> make it always happen afterwards, though, and it might well be totally
> impossible if triggers cause more table updates to occur.

I suppose that it might be possible for AfterTriggersTableData to
record the location of a previously queued event so that you can later
disable it and queue a replacement, with the effect of suppressing
earlier firings rather than later ones.  That might make sense if you
think that after statement triggers should fire after all row
triggers.  I can't figure out from the spec whether that's expected
and I'm not sure if it's useful.

> Because MakeTransitionCaptureState now depends on the trigger query
> level being active, I had to relocate the AfterTriggerBeginQuery calls
> to occur before it.

Right.

> In passing, I refactored the AfterTriggers data structures a bit so
> that we don't need to do as many palloc calls to manage them.  Instead
> of several independent arrays there's now one array of structs.

Good improvement.

-- 
Thomas Munro
http://www.enterprisedb.com