Horton

Thread

Commits

  1. jit: Re-allow JIT compilation of execGrouping.c hashtable comparisons.

  2. Fix determination when slot types for upper executor nodes are fixed.

  1. JIT performance bug/regression & JIT EXPLAIN

    Andres Freund <andres@anarazel.de> — 2019-09-27T07:20:53Z

    Hi,

Unfortunately I found a performance regression for JITed query
compilation introduced in 12, compared to 11. Fixed in one of the
attached patches (v1-0009-Fix-determination-when-tuple-deforming-can-be-JIT.patch
- which needs a better commit message).

The first question is when to push that fix. I'm inclined to just do so
now - as we still do JITed tuple deforming in most cases, as well as
doing so in 11 in the places this patch fixes, the risk of that seems
low. But I can also see an arguments for waiting after 12.0.


For me the bigger question is about how to make sure we can write tests
determining which parts of the querytree are JIT compiled and which are
not. There's the above bug, and I'm also hunting a regression introduced
somewhere during 11's lifetime, which suggests to me that we need better
coverage. I also want to add new JIT logic, making this even more
important.


The reason that 11 didn't have tests verifying that certain parts of the
plan tree are JIT compiled is that EXPLAIN doesn't currently show the
relevant information, and it's not that trivial to do so.

What I'd like to do is to add additional, presumably optional, output to
EXPLAIN showing additional information about expressions.

There's two major parts to doing so:

1) Find a way to represent the additional information attached to
   expressions, and provide show_expression et al with the ExprState to
   be able to do so.  The additional information I think is necessary is
   a) is expression jit compiled
   b-d) is scan/outer/inner tuple deforming necessary, and if so, JIT
   compiled.

   We can't unconditionally JIT compile for tuple deforming, because
   there's a number of cases where the source slot doesn't have
   precisely the same tuple desc, and/or doesn't have the same type.

2) Expand EXPLAIN output to show expressions that currently aren't
   shown. Performance-wise the ones most critical that aren't currently
   visible, and that I know about, are:
   - Agg's combined transition function, we also currently don't display
     in any understandable way how many passes over the input we do (for
     grouping sets), nor how much memory is needed.
   - Agg's hash comparator (separate regression referenced above)
   - Hash/HashJoin's hashkeys/hjclauses


For 1) think we need to change show_expression()/show_qual() etc to also
pass down the corresponding ExprState if available (not available in
plenty of cases, most of which are not particularly important). That's
fairly mechanical.

Then we need to add information about JIT to individual expressions. In
the attached WIP patchset I've made that dependent on the new
"jit_details" EXPLAIN option. When specified, new per-expression
information is shown:
- JIT-Expr: whether the expression was JIT compiled (might e.g. not be
  the case because no parent was provided)
- JIT-Deform-{Scan,Outer,Inner}: wether necessary, and whether JIT accelerated.

I don't like these names much, but ...

For the deform cases I chose to display
a) the function name if JIT compiled
b) "false" if the expression is JIT compiled, deforming is
   necessary, but deforming is not JIT compiled (e.g. because the slot type
   wasn't fixed)
c) "null" if not necessary, with that being omitted in text mode.

So e.g in json format this looks like:

"Filter": {
  "Expr": "(lineitem.l_shipdate <= '1998-09-18 00:00:00'::timestamp without time zone)",
  "JIT-Expr": "evalexpr_0_2",
  "JIT-Deform-Scan": "deform_0_3",
  "JIT-Deform-Outer": null,
  "JIT-Deform-Inner": null
}
and in text mode:

Filter: (lineitem.l_shipdate <= '1998-09-18 00:00:00'::timestamp without time zone); JIT-Expr: evalexpr_0_2, JIT-Deform-Scan: deform_0_3

For now I chose to make Filter a group when both, not in text mode and
jit_details on - otherwise it's unclear what the JIT fields would apply
to. But that's pretty crappy, because it means that the 'shape' of the
output depends on the jit_details option.  I think if we were starting
from scratch it'd make sense to alway have the Expression as it's own
sub-node, so interpreting code doesn't have to know all the places an
expression can be referenced from. But it's probably not too attractive
to change that today?

Somewhat independently the series also contains a patch that renames
verbose mode's "Output" to project if the node projects. I find it
pretty hard to interpret whether a node projects otherwise, and it's
confusing when jit_details shows details only for some node's Output,
but not for others. But the compat break due to that change is not small
- perhaps we could instead mark that in another way?


For 2) I've only started to improve the situation, but it's a pretty
number of pretty crucial pieces.

I first focussed adding information for Agg nodes, as a) those are
typically performance sensitive in cases where JIT is beneficial b) the
current instrumentation is really insufficient, especially in cases
where multiple grouping sets are computed at the same time - I think
it's effectilvey not interpretable.

In verbose mode explain now shows per-phase output about the transition
computation. E.g. for a grouping set query that can't be computed in one
pass, it now displays something like

MixedAggregate  (cost=6083420.07..14022888.98 rows=10011685 width=64)
  Project: avg((l_linenumber)::bigint), count((l_partkey)::bigint), sum(l_quantity), l_linenumber, l_partkey, l_quantity
  Filter: (sum(lineitem.l_quantity) IS NOT NULL)
  Phase 2 using strategy "Sort":
    Sort Key: lineitem.l_partkey, lineitem.l_quantity
    Transition Function: 2 * int8_avg_accum(TRANS, (l_linenumber)::bigint), 2 * int8inc_any(TRANS, (l_partkey)::bigint), 2 * float8pl(TRANS, l_quantity)
    Sorted Group: lineitem.l_partkey, lineitem.l_quantity
    Sorted Group: lineitem.l_partkey
  Phase 1 using strategy "Sorted Input & All & Hash":
    Transition Function: 6 * int8_avg_accum(TRANS, (l_linenumber)::bigint), 6 * int8inc_any(TRANS, (l_partkey)::bigint), 6 * float8pl(TRANS, l_quantity)
    Sorted Input Group: lineitem.l_linenumber, lineitem.l_partkey, lineitem.l_quantity
    Sorted Input Group: lineitem.l_linenumber, lineitem.l_partkey
    Sorted Input Group: lineitem.l_linenumber
    All Group
    Hash Group: lineitem.l_quantity
    Hash Group: lineitem.l_quantity, lineitem.l_linenumber
  ->  Sort  (cost=6083420.07..6158418.50 rows=29999372 width=16)
        ...

The N * indicates how many of the same transition functions are computed
during that phase.

I'm not sure that 'TRANS' is the best placeholder for the transition
value here. Maybe $TRANS would be clearer?

For a parallel aggregate the upper level looks like:

Finalize HashAggregate  (cost=610681.93..610682.02 rows=9 width=16)
  Project: l_tax, sum(l_quantity)
  Phase 0 using strategy "Hash":
    Transition Function: float8pl(TRANS, (PARTIAL sum(l_quantity)))
    Hash Group: lineitem.l_tax
  ->  Gather  (cost=610677.11..610681.70 rows=45 width=16)
        Output: l_tax, (PARTIAL sum(l_quantity))
        Workers Planned: 5
        ->  Partial HashAggregate  (cost=609677.11..609677.20 rows=9 width=16)
              Project: l_tax, PARTIAL sum(l_quantity)

I've not done that yet, but I think it's way past time that we also add
memory usage information to Aggregate nodes (both for the hashtable(s),
and for internal sorts if those are performed for grouping sets). Which
would also be very hard in the "current" format, as there's no
representation of passes.

With jit_details enabled, we then can show information about the
aggregation function, and grouping functions:
  Phase 0 using strategy "Hash":
    Transition Function: float8pl(TRANS, (PARTIAL sum(l_quantity))); JIT-Expr: evalexpr_0_11, JIT-Deform-Outer: false
    Hash Group: lineitem.l_tax; JIT-Expr: evalexpr_0_8, JIT-Deform-Outer: deform_0_10, JIT-Deform-Inner: deform_0_9


Currently the "new" format is used when either grouping sets are in use
(as the previous explain output was not particularly useful, and
information about the passes is important), or if VERBOSE or JIT_DETAILS
are specified.


For HashJoin/Hash I've added 'Outer Hash Key' and 'Hash Key' for each
key, but only in verbose mode. That's somewhat important because for
HashJoins those currently are often the performance critical bit,
because they'll commonly be the expressions that deform the slots from
below.  That display is somewhat redundant with HashJoins "Hash Cond",
but they're evaluated separately. Under verbose that seems OK to me.

With jit_details enabled, this e.g. looks like this:

 Hash Join  (cost=271409.60..2326739.51 rows=30000584 width=250)
   Project: lineitem.l_orderkey, lineitem.l_partkey, lineitem.l_suppkey, lineitem.l_linenumber, lineitem.l_quantity, lineitem.l_extendedprice, lineitem.l_discount, lineitem.l_tax,
   Inner Unique: true
   Hash Cond: ((lineitem.l_partkey = partsupp.ps_partkey) AND (lineitem.l_suppkey = partsupp.ps_suppkey)); JIT-Expr: evalexpr_0_7, JIT-Deform-Outer: deform_0_9, JIT-Deform-Inner:
   Outer Hash Key: lineitem.l_partkey; JIT-Expr: evalexpr_0_10, JIT-Deform-Outer: deform_0_11
   Outer Hash Key: lineitem.l_suppkey; JIT-Expr: evalexpr_0_12, JIT-Deform-Outer: deform_0_13
   ->  Seq Scan on public.lineitem  (cost=0.00..819684.84 rows=30000584 width=106)
         Output: lineitem.l_orderkey, lineitem.l_partkey, lineitem.l_suppkey, lineitem.l_linenumber, lineitem.l_quantity, lineitem.l_extendedprice, lineitem.l_discount, lineitem.l
   ->  Hash  (cost=129384.24..129384.24 rows=3999824 width=144)
         Output: partsupp.ps_partkey, partsupp.ps_suppkey, partsupp.ps_availqty, partsupp.ps_supplycost, partsupp.ps_comment
         Hash Key: partsupp.ps_partkey; JIT-Expr: evalexpr_0_0, JIT-Deform-Outer: deform_0_1
         Hash Key: partsupp.ps_suppkey; JIT-Expr: evalexpr_0_2, JIT-Deform-Outer: deform_0_3
         ->  Seq Scan on public.partsupp  (cost=0.00..129384.24 rows=3999824 width=144)
               Output: partsupp.ps_partkey, partsupp.ps_suppkey, partsupp.ps_availqty, partsupp.ps_supplycost, partsupp.ps_comment
 JIT:
   Functions: 14 (6 for expression evaluation, 8 for tuple deforming)
   Options: Inlining true, Optimization true, Expressions true, Deforming true

this also highlights the sad fact that we currently use a separate
ExprState to compute each of the hash keys, and then "manually" invoke
the hash function itself. That's bad both for interpreted execution, as
we repeatedly pay executor startup overhead and don't even hit the
fastpath, as well as for JITed execution, because we have more code to
optimize (some of it pretty redundant, in particular the deforming). In
both cases we suffer from the problem that we deform the tuple
incrementally.


A later patch in the series then uses the new explain output to add some
tests for JIT, and then fixes two bugs, showing that the test output
changes.

Additionally I've also included a small improvement to the expression
evaluation logic, which also changes output in the JIT test, as it
should.

Comments?

Greetings,

Andres Freund

  2. Re: JIT performance bug/regression & JIT EXPLAIN

    Andres Freund <andres@anarazel.de> — 2019-09-29T23:30:29Z

    Hi,

On 2019-09-27 00:20:53 -0700, Andres Freund wrote:
> Unfortunately I found a performance regression for JITed query
> compilation introduced in 12, compared to 11. Fixed in one of the
> attached patches (v1-0009-Fix-determination-when-tuple-deforming-can-be-JIT.patch
> - which needs a better commit message).
> 
> The first question is when to push that fix. I'm inclined to just do so
> now - as we still do JITed tuple deforming in most cases, as well as
> doing so in 11 in the places this patch fixes, the risk of that seems
> low. But I can also see an arguments for waiting after 12.0.

Since nobody opined, I now have pushed that, and the other fix mentioned
later in that email.

I'd appreciate comments on the rest of the email, it's clear that we
need to improve the test infrastructure here. And also the explain
output for grouping sets...

Greetings,

Andres Freund

  3. Re: JIT performance bug/regression & JIT EXPLAIN

    Maciek Sakrejda <m.sakrejda@gmail.com> — 2019-10-28T18:27:02Z

    >But that's pretty crappy, because it means that the 'shape' of the output depends on the jit_details option.

Yeah, that would be hard to work with. What about adding it as a sibling group?

"Filter": "(lineitem.l_shipdate <= '1998-09-18 00:00:00'::timestamp
without time zone)",
"Filter JIT": {
  "Expr": "evalexpr_0_2",
  "Deform Scan": "deform_0_3",
  "Deform Outer": null,
  "Deform Inner": null
}

Also not that pretty, but at least it's easier to work with (I also
changed the dashes to spaces since that's what the rest of EXPLAIN is
doing as a matter of style).

>But the compat break due to that change is not small- perhaps we could instead mark that in another way?

We could add a "Projects" boolean key instead? Of course that's more
awkward in text mode. Maybe compat break is less of an issue in text
mode and we can treat this differently?

>I'm not sure that 'TRANS' is the best placeholder for the transition value here. Maybe $TRANS would be clearer?

+1, I think the `$` makes it clearer that this is not a literal expression.

>For HashJoin/Hash I've added 'Outer Hash Key' and 'Hash Key' for each key, but only in verbose mode.

That reads pretty well to me. What does the structured output look like?

  4. Re: JIT performance bug/regression & JIT EXPLAIN

    Robert Haas <robertmhaas@gmail.com> — 2019-10-28T19:05:01Z

    On Fri, Sep 27, 2019 at 3:21 AM Andres Freund <andres@anarazel.de> wrote:
> - JIT-Expr: whether the expression was JIT compiled (might e.g. not be
>   the case because no parent was provided)
> - JIT-Deform-{Scan,Outer,Inner}: wether necessary, and whether JIT accelerated.
>
> I don't like these names much, but ...
>
> For the deform cases I chose to display
> a) the function name if JIT compiled
> b) "false" if the expression is JIT compiled, deforming is
>    necessary, but deforming is not JIT compiled (e.g. because the slot type
>    wasn't fixed)
> c) "null" if not necessary, with that being omitted in text mode.

I mean, why not just omit in all modes if it's not necessary? I don't
see that making the information we produce randomly inconsistent
between modes is buying us anything.

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

  5. Re: JIT performance bug/regression & JIT EXPLAIN

    Andres Freund <andres@anarazel.de> — 2019-10-28T23:21:45Z

    Hi,

On 2019-10-28 15:05:01 -0400, Robert Haas wrote:
> On Fri, Sep 27, 2019 at 3:21 AM Andres Freund <andres@anarazel.de> wrote:
> > - JIT-Expr: whether the expression was JIT compiled (might e.g. not be
> >   the case because no parent was provided)
> > - JIT-Deform-{Scan,Outer,Inner}: wether necessary, and whether JIT accelerated.
> >
> > I don't like these names much, but ...
> >
> > For the deform cases I chose to display
> > a) the function name if JIT compiled
> > b) "false" if the expression is JIT compiled, deforming is
> >    necessary, but deforming is not JIT compiled (e.g. because the slot type
> >    wasn't fixed)
> > c) "null" if not necessary, with that being omitted in text mode.
> 
> I mean, why not just omit in all modes if it's not necessary? I don't
> see that making the information we produce randomly inconsistent
> between modes is buying us anything.

Because that's the normal way to represent something non-existing for
formats like json? There's a lot of information we show always for !text
format, even if not really applicable to the context (e.g. Triggers for
select statements). I think there's an argument to made to deviate in
this case, but I don't think it's obvious.

Abstract formatting reasons aside, it's actually useful to see where we
know we're dealing with tuples that don't need to be deformed and thus
overhead due to that cannot be relevant. Not sure if there's sufficient
consumers for that, but ...  We e.g. should verify that the "none"
doesn't suddenly vanish, because we broke the information that let us
infer that we don't need tuple deforming - and that's easier to
understand if there's an explicit field, rather than reasining from
absence. IMO.

Greetings,

Andres Freund

  6. Re: JIT performance bug/regression & JIT EXPLAIN

    Andres Freund <andres@anarazel.de> — 2019-10-29T00:02:29Z

    Hi,

On 2019-10-28 11:27:02 -0700, Maciek Sakrejda wrote:
> >But that's pretty crappy, because it means that the 'shape' of the output depends on the jit_details option.
> 
> Yeah, that would be hard to work with. What about adding it as a sibling group?
> 
> "Filter": "(lineitem.l_shipdate <= '1998-09-18 00:00:00'::timestamp
> without time zone)",
> "Filter JIT": {
>   "Expr": "evalexpr_0_2",
>   "Deform Scan": "deform_0_3",
>   "Deform Outer": null,
>   "Deform Inner": null
> }
> 
> Also not that pretty, but at least it's easier to work with

What I dislike about that is that it basically again is introducing
something that requires either pattern matching on key names (i.e. a key
of '(.*) JIT' is one that has information about JIT, and the associated
expresssion is in key $1), or knowing all the potential keys an
expression could be in.


> (I also
> changed the dashes to spaces since that's what the rest of EXPLAIN is
> doing as a matter of style).

That makes sense.


> >But the compat break due to that change is not small- perhaps we could instead mark that in another way?
> 
> We could add a "Projects" boolean key instead? Of course that's more
> awkward in text mode. Maybe compat break is less of an issue in text
> mode and we can treat this differently?

Yea, I think projects as a key for each node makes sense. For text mode
I guess we could just display the key on the same line when es->verbose
is set?  Still not sure if not just changing the output is the better
approach.

Another alternative would be to just remove the 'Output' line when a
node doesn't project - it can't really carry meaning in those cases
anyway?


> >For HashJoin/Hash I've added 'Outer Hash Key' and 'Hash Key' for each key, but only in verbose mode.
> 
> That reads pretty well to me. What does the structured output look
> like?

Just a new "Outer Hash Key" for the HashJoin node, and "Hash Key" for
the Hash node. Perhaps the latter should be 'Inner Hash Key' - while
that's currently a bit confusing because of Hash's subtree being the
outer tree, it'd reduce changes when merging Hash into HashJoin [1], and
it's clearer when looking at the HashJoin node itself.

Here's an example query:

EXPLAIN (VERBOSE, FORMAT JSON, COSTS OFF) SELECT pc.oid::regclass, pc.relkind, pc.relfilenode, pc_t.oid::regclass as toast_rel, pc_t.relfilenode as toast_relfilenode FROM pg_class pc LEFT OUTER JOIN pg_class pc_t ON (pc.reltoastrelid = pc_t.oid);
[
  {
    "Plan": {
      "Node Type": "Hash Join",
      "Parallel Aware": false,
      "Join Type": "Left",
      "Project": ["(pc.oid)::regclass", "pc.relkind", "pc.relfilenode", "(pc_t.oid)::regclass", "pc_t.relfilenode"],
      "Inner Unique": true,
      "Hash Cond": "(pc.reltoastrelid = pc_t.oid)",
      "Outer Hash Key": "pc.reltoastrelid",
      "Plans": [
        {
          "Node Type": "Seq Scan",
          "Parent Relationship": "Outer",
          "Parallel Aware": false,
          "Relation Name": "pg_class",
          "Schema": "pg_catalog",
          "Alias": "pc",
          "Output": ["pc.oid", "pc.relname", "pc.relnamespace", "pc.reltype", "pc.reloftype", "pc.relowner", "pc.relam", "pc.relfilenode", "pc.reltablespace", "pc.relpages", "pc.reltuples", "pc.relallvisible", "pc.reltoastrelid", "pc.relhasindex", "pc.relisshared", "pc.relpersistence", "pc.relkind", "pc.relnatts", "pc.relchecks", "pc.relhasrules", "pc.relhastriggers", "pc.relhassubclass", "pc.relrowsecurity", "pc.relforcerowsecurity", "pc.relispopulated", "pc.relreplident", "pc.relispartition", "pc.relrewrite", "pc.relfrozenxid", "pc.relminmxid", "pc.relacl", "pc.reloptions", "pc.relpartbound"]
        },
        {
          "Node Type": "Hash",
          "Parent Relationship": "Inner",
          "Parallel Aware": false,
          "Output": ["pc_t.oid", "pc_t.relfilenode"],
          "Hash Key": "pc_t.oid",
          "Plans": [
            {
              "Node Type": "Seq Scan",
              "Parent Relationship": "Outer",
              "Parallel Aware": false,
              "Relation Name": "pg_class",
              "Schema": "pg_catalog",
              "Alias": "pc_t",
              "Project": ["pc_t.oid", "pc_t.relfilenode"]
            }
          ]
        }
      ]
    }
  }
]

and in plain text:

Hash Left Join
  Project: (pc.oid)::regclass, pc.relkind, pc.relfilenode, (pc_t.oid)::regclass, pc_t.relfilenode
  Inner Unique: true
  Hash Cond: (pc.reltoastrelid = pc_t.oid)
  Outer Hash Key: pc.reltoastrelid
  ->  Seq Scan on pg_catalog.pg_class pc
        Output: pc.oid, pc.relname, pc.relnamespace, pc.reltype, pc.reloftype, pc.relowner, pc.relam, pc.relfilenode, pc.reltablespace, pc.relpages, pc.reltuples, pc.relallvisible, pc.reltoastrelid, pc.relhasindex, pc.relisshared, pc.relpersistence, pc.relkind, pc.relnatts, pc.relchecks, pc.relhasrules, pc.relhastriggers, pc.relhassubclass, pc.relrowsecurity, pc.relforcerowsecurity, pc.relispopulated, pc.relreplident, pc.relispartition, pc.relrewrite, pc.relfrozenxid, pc.relminmxid, pc.relacl, pc.reloptions, pc.relpartbound
  ->  Hash
        Output: pc_t.oid, pc_t.relfilenode
        Hash Key: pc_t.oid
        ->  Seq Scan on pg_catalog.pg_class pc_t
              Project: pc_t.oid, pc_t.relfilenode

which also serves as an example about my previous point about
potentially just hiding the 'Output: ' bit when no projection is done:
It's very verbose, without adding much, while hiding that there's
actually nothing being done at the SeqScan level.

I've attached a rebased version of the patcheset. No changes except for
a minor conflict, and removing some already applied bugfixes.

Greetings,

Andres Freund

[1] https://www.postgresql.org/message-id/20191028231526.wcnwag7lllkra4qt%40alap3.anarazel.de

  7. Re: JIT performance bug/regression & JIT EXPLAIN

    Maciek Sakrejda <m.sakrejda@gmail.com> — 2019-11-12T21:42:10Z

    On Mon, Oct 28, 2019 at 5:02 PM Andres Freund <andres@anarazel.de> wrote:
> What I dislike about that is that it basically again is introducing

"again"? Am I missing some history here? I'd love to read up on this
if there are mistakes to learn from.

> something that requires either pattern matching on key names (i.e. a key
> of '(.*) JIT' is one that has information about JIT, and the associated
> expresssion is in key $1), or knowing all the potential keys an
> expression could be in.

That still seems less awkward than having to handle a Filter field
that's either scalar or a group. Most current EXPLAIN options just add
additional fields to the structured plan instead of modifying it, no?
If that output is better enough, though, maybe we should just always
make Filter a group and go with the breaking change? If tooling
authors need to treat this case specially anyway, might as well evolve
the format.

> Another alternative would be to just remove the 'Output' line when a
> node doesn't project - it can't really carry meaning in those cases
> anyway?

¯\_(ツ)_/¯

For what it's worth, I certainly wouldn't miss it.

  8. Re: JIT performance bug/regression & JIT EXPLAIN

    Andres Freund <andres@anarazel.de> — 2019-11-12T22:21:31Z

    Hi,

On 2019-11-12 13:42:10 -0800, Maciek Sakrejda wrote:
> On Mon, Oct 28, 2019 at 5:02 PM Andres Freund <andres@anarazel.de> wrote:
> > What I dislike about that is that it basically again is introducing
> 
> "again"? Am I missing some history here? I'd love to read up on this
> if there are mistakes to learn from.

I think I was mostly referring to mistakes we've made for the json etc
key names. By e.g. having expressions as "Function Call", "Table
Function Call", "Filter", "TID Cond", ... a tool that wants to interpret
the output needs awareness of all of these different names, rather than
knowing that everything with a sub-group "Expression" has to be an
expression.

I.e. instead of

    "Plan": {
      "Node Type": "Seq Scan",
      "Parallel Aware": false,
      "Relation Name": "pg_class",
      "Schema": "pg_catalog",
      "Alias": "pg_class",
      "Startup Cost": 0.00,
      "Total Cost": 17.82,
      "Plan Rows": 385,
      "Plan Width": 68,
      "Output": ["relname", "tableoid"],
      "Filter": "(pg_class.relname <> 'foo'::name)"
    }

we ought to have gone for

    "Plan": {
      "Node Type": "Seq Scan",
      "Parallel Aware": false,
      "Relation Name": "pg_class",
      "Schema": "pg_catalog",
      "Alias": "pg_class",
      "Startup Cost": 0.00,
      "Total Cost": 17.82,
      "Plan Rows": 385,
      "Plan Width": 68,
      "Output": ["relname", "tableoid"],
      "Filter": {"Expression" : { "text": (pg_class.relname <> 'foo'::name)"}}
    }

or something like that. Which'd then make it obvious how to add
information about JIT to each expression.


Whereas the proposal of the separate key name perpetuates the
messiness...


> > something that requires either pattern matching on key names (i.e. a key
> > of '(.*) JIT' is one that has information about JIT, and the associated
> > expresssion is in key $1), or knowing all the potential keys an
> > expression could be in.
> 
> That still seems less awkward than having to handle a Filter field
> that's either scalar or a group.

Yea, it's a sucky option :(


> Most current EXPLAIN options just add
> additional fields to the structured plan instead of modifying it, no?
> If that output is better enough, though, maybe we should just always
> make Filter a group and go with the breaking change? If tooling
> authors need to treat this case specially anyway, might as well evolve
> the format.

Yea, maybe that's the right thing to do. Would be nice to have some more
input...

Greetings,

Andres Freund

  9. Re: JIT performance bug/regression & JIT EXPLAIN

    Robert Haas <robertmhaas@gmail.com> — 2019-11-13T19:29:07Z

    On Mon, Oct 28, 2019 at 7:21 PM Andres Freund <andres@anarazel.de> wrote:
> Because that's the normal way to represent something non-existing for
> formats like json? There's a lot of information we show always for !text
> format, even if not really applicable to the context (e.g. Triggers for
> select statements). I think there's an argument to made to deviate in
> this case, but I don't think it's obvious.

I've consistently been of the view that anyone who thinks that the
FORMAT option should affect what information gets displayed doesn't
understand the meaning of the word "format." And I still feel that
way.

I also think that conditionally renaming "Output" to "Project" is a
super-bad idea. The idea of a format like this is that the "keys" stay
constant and the values change. If you need to tell people more, you
add more keys.

I also think that making the Filter field a group conditionally is a
bad idea, for similar reasons. But making it always be a group doesn't
necessarily seem like a bad idea. I think, though, that you could
handle this in other ways, like by suffixing existing keys.  e.g. if
you've got Index-Qual and Filter, just do Index-Qual-JIT and
Filter-JIT and call it good.

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

  10. Re: JIT performance bug/regression & JIT EXPLAIN

    Andres Freund <andres@anarazel.de> — 2019-11-13T20:03:18Z

    Hi,

On 2019-11-13 14:29:07 -0500, Robert Haas wrote:
> On Mon, Oct 28, 2019 at 7:21 PM Andres Freund <andres@anarazel.de> wrote:
> > Because that's the normal way to represent something non-existing for
> > formats like json? There's a lot of information we show always for !text
> > format, even if not really applicable to the context (e.g. Triggers for
> > select statements). I think there's an argument to made to deviate in
> > this case, but I don't think it's obvious.
> 
> I've consistently been of the view that anyone who thinks that the
> FORMAT option should affect what information gets displayed doesn't
> understand the meaning of the word "format." And I still feel that
> way.

Well, it's not been that way since the format option was added, so ...



> I also think that conditionally renaming "Output" to "Project" is a
> super-bad idea. The idea of a format like this is that the "keys" stay
> constant and the values change. If you need to tell people more, you
> add more keys.

Yea, I don't like the compat break either.  But I'm not so convinced
that just continuing to collect cruft because of compatibility is worth
it - I just don't see an all that high reliance interest for explain
output.

I think adding a new key is somewhat ok for !text, but for text that
doesn't seem like an easy solution?

I kind of like my idea somewhere downthread, in a reply to Maciek, of
simply not listing "Output" for nodes that don't project.  While that's
still a format break, it seems that tools already need to deal with
"Output" not being present?


> I also think that making the Filter field a group conditionally is a
> bad idea, for similar reasons.

Oh, yea, it's utterly terrible (I called it crappy in my email :)).


> But making it always be a group doesn't necessarily seem like a bad
> idea. I think, though, that you could handle this in other ways, like
> by suffixing existing keys.  e.g. if you've got Index-Qual and Filter,
> just do Index-Qual-JIT and Filter-JIT and call it good.

Maciek suggested the same. But to me it seems going down that way will
make the format harder and harder to understand? So I think I'd rather
break compat here, and go for a group.

Personally I think the group naming choice for explain makes the the
!text outputs much less useful than they could be - we basically force
every tool to understand all possible keys, to make sense of formatted
output. Instead of something like 'Filter: {"Qual":{"text" : "...",
"JIT": ...}' where a tool only needed to understand that everything that
has a "Qual" inside is a filtering expression, everything that has a
"Project" is a projecting type of expression, ... a tool needs to know
about "Inner Cond", "Order By", "Filter", "Recheck Cond", "TID Cond",
"Join Filter", "Merge Cond", "Hash Cond", "One-Time Filter", ...

Greetings,

Andres Freund

  11. Re: JIT performance bug/regression & JIT EXPLAIN

    Robert Haas <robertmhaas@gmail.com> — 2019-11-15T13:49:02Z

    On Wed, Nov 13, 2019 at 3:03 PM Andres Freund <andres@anarazel.de> wrote:
> Well, it's not been that way since the format option was added, so ...

It was pretty close in the original version, but people keep trying to
be clever.

> > I also think that conditionally renaming "Output" to "Project" is a
> > super-bad idea. The idea of a format like this is that the "keys" stay
> > constant and the values change. If you need to tell people more, you
> > add more keys.
>
> Yea, I don't like the compat break either.  But I'm not so convinced
> that just continuing to collect cruft because of compatibility is worth
> it - I just don't see an all that high reliance interest for explain
> output.
>
> I think adding a new key is somewhat ok for !text, but for text that
> doesn't seem like an easy solution?
>
> I kind of like my idea somewhere downthread, in a reply to Maciek, of
> simply not listing "Output" for nodes that don't project.  While that's
> still a format break, it seems that tools already need to deal with
> "Output" not being present?

Yes, I think leaving out Output for a node that doesn't Project would
be fine, as long as we're consistent about it.

> > But making it always be a group doesn't necessarily seem like a bad
> > idea. I think, though, that you could handle this in other ways, like
> > by suffixing existing keys.  e.g. if you've got Index-Qual and Filter,
> > just do Index-Qual-JIT and Filter-JIT and call it good.
>
> Maciek suggested the same. But to me it seems going down that way will
> make the format harder and harder to understand? So I think I'd rather
> break compat here, and go for a group.

Personally, I don't care very much about backward-compatibility, or
about how hard it is for tools to parse. I want it to be possible, but
if it takes a little extra effort, so be it. My main concern is having
the text output look good to human beings, because that is the primary
format and they are the primary consumers.

> Personally I think the group naming choice for explain makes the the
> !text outputs much less useful than they could be - we basically force
> every tool to understand all possible keys, to make sense of formatted
> output. Instead of something like 'Filter: {"Qual":{"text" : "...",
> "JIT": ...}' where a tool only needed to understand that everything that
> has a "Qual" inside is a filtering expression, everything that has a
> "Project" is a projecting type of expression, ... a tool needs to know
> about "Inner Cond", "Order By", "Filter", "Recheck Cond", "TID Cond",
> "Join Filter", "Merge Cond", "Hash Cond", "One-Time Filter", ...

It's not that long of a list, and I don't know of a tool that tries to
do something in particular with all of those types of things anyway.

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

  12. Re: JIT performance bug/regression & JIT EXPLAIN

    Maciek Sakrejda <m.sakrejda@gmail.com> — 2019-11-16T01:04:52Z

    On Fri, Nov 15, 2019 at 5:49 AM Robert Haas <robertmhaas@gmail.com> wrote:
> Personally, I don't care very much about backward-compatibility, or
> about how hard it is for tools to parse. I want it to be possible, but
> if it takes a little extra effort, so be it.

I think these are two separate issues. I agree on
backward-compatibility (especially if we can embed a server version in
structured EXPLAIN output to make it easier for tools to track format
differences), but not caring how hard it is for tools to parse? What's
the point of structured formats, then?

> My main concern is having
> the text output look good to human beings, because that is the primary
> format and they are the primary consumers.

Structured output is also for human beings, albeit indirectly. That
text is the primary format may be more of a reflection of the
difficulty of building and integrating EXPLAIN tools than its inherent
superiority (that said, I'll concede it's a concise and elegant format
for what it does). What if psql supported an EXPLAINER like it does
EDITOR?

For what it's worth, after thinking about this a bit, I'd like to see
structured EXPLAIN evolve into a more consistent format, even if it
means breaking changes (and I do think a version specifier at the root
of the plan would make this easier).

  13. Re: JIT performance bug/regression & JIT EXPLAIN

    Tom Lane <tgl@sss.pgh.pa.us> — 2020-01-27T17:15:53Z

    Maciek Sakrejda <m.sakrejda@gmail.com> writes:
> On Fri, Nov 15, 2019 at 5:49 AM Robert Haas <robertmhaas@gmail.com> wrote:
>> Personally, I don't care very much about backward-compatibility, or
>> about how hard it is for tools to parse. I want it to be possible, but
>> if it takes a little extra effort, so be it.

> I think these are two separate issues. I agree on
> backward-compatibility (especially if we can embed a server version in
> structured EXPLAIN output to make it easier for tools to track format
> differences), but not caring how hard it is for tools to parse? What's
> the point of structured formats, then?

I'd not been paying any attention to this thread, but Andres just
referenced it in another discussion, so I went back and read it.
Here's my two cents:

* I agree with Robert that conditionally changing "Output" to "Project" is
an absolutely horrid idea.  That will break every tool that looks at this
stuff, and it just flies in the face of the design principle that the
output schema should be stable, and it'll be a long term pain-in-the-rear
for regression test back-patching, and it will confuse users much more than
it will help them.  The other idea of suppressing "Output" in cases where
no projection is happening might be all right, but only in text format
where we don't worry about schema stability.  Another idea perhaps is
to emit "Output: all columns" (in text formats, less sure what to do in
structured formats).

* In the structured formats, I think it should be okay to convert
expression-ish fields from being raw strings to being {Expression}
sub-nodes with the raw string as one field.  Aside from making it easy
to inject JIT info, that would also open the door to someday showing
expressions in some more-parse-able format than a string, since other
representations could also be added as new fields.  (I have a vague
recollection of wanting a list of all the Vars used in an expression,
for example.)

* Unfortunately that does nothing for the problem of how to show
per-expression JIT info in text format.  Maybe we just shouldn't.
I do not think that the readability-vs-usefulness tradeoff is going
to be all that good there, anyway.  Certainly for testing purposes
it's going to be more useful to examine portions of a structured output.

* I'm not on board with the idea of adding a version number to the
structured output formats.  In the first place, it's too late, since
we didn't leave room for one to begin with.  In the second, an overall
version number just isn't very helpful for this sort of problem.  If a
tool sees a version number higher than the latest thing it knows, what's
it supposed to do, just fail?  In practice it could still extract an awful
lot of info, so that really isn't a desirable answer.  It's better if the
data structure is such that a tool can understand that some sub-part of
the data is something it can't interpret, and just ignore that part.
(This is more or less the same design principle that PNG image format
was built on, FWIW.)  Adding on fields to an existing node type easily
meets that requirement, as does inventing new sub-node types, and that's
all that we've done so far.  But I think that replacing a scalar field
value with a sub-node probably works too (at least for well-written
tools), so the expression change suggested above should be OK.

			regards, tom lane

  14. Re: JIT performance bug/regression & JIT EXPLAIN

    Andres Freund <andres@anarazel.de> — 2020-01-27T17:41:03Z

    Hi,

On 2020-01-27 12:15:53 -0500, Tom Lane wrote:
> Maciek Sakrejda <m.sakrejda@gmail.com> writes:
> > On Fri, Nov 15, 2019 at 5:49 AM Robert Haas <robertmhaas@gmail.com> wrote:
> >> Personally, I don't care very much about backward-compatibility, or
> >> about how hard it is for tools to parse. I want it to be possible, but
> >> if it takes a little extra effort, so be it.
>
> > I think these are two separate issues. I agree on
> > backward-compatibility (especially if we can embed a server version in
> > structured EXPLAIN output to make it easier for tools to track format
> > differences), but not caring how hard it is for tools to parse? What's
> > the point of structured formats, then?
>
> I'd not been paying any attention to this thread, but Andres just
> referenced it in another discussion, so I went back and read it.
> Here's my two cents:
>
> * I agree with Robert that conditionally changing "Output" to "Project" is
> an absolutely horrid idea.

Yea, I think I'm convinced on that front. I never liked the idea, and
the opposition has been pretty unanimous...


> That will break every tool that looks at this stuff, and it just flies
> in the face of the design principle that the output schema should be
> stable, and it'll be a long term pain-in-the-rear for regression test
> back-patching, and it will confuse users much more than it will help
> them.  The other idea of suppressing "Output" in cases where no
> projection is happening might be all right, but only in text format
> where we don't worry about schema stability.  Another idea perhaps is
> to emit "Output: all columns" (in text formats, less sure what to do
> in structured formats).

I think I like the "all columns" idea. Not what I'd do on a green field,
but...

If we were just dealing with the XML format, we could just add a

<Projecting>True/False</Projecting>
to the current
<Output>
   <Item>a</Item>
   <Item>b</Item>
   ...
</Output>

and it'd make plenty sense. but for json's
    "Output": ["a", "b"]
and yaml's
    Output:
      - "a"
      - "b"
that's not an option as far as I can tell. Not sure what to do about
that.



> * In the structured formats, I think it should be okay to convert
> expression-ish fields from being raw strings to being {Expression}
> sub-nodes with the raw string as one field.  Aside from making it easy
> to inject JIT info, that would also open the door to someday showing
> expressions in some more-parse-able format than a string, since other
> representations could also be added as new fields.  (I have a vague
> recollection of wanting a list of all the Vars used in an expression,
> for example.)

Cool. Being extendable seems like a good direction. That's what I
primarily dislike about the various work-arounds for how to associate
information about JIT by a "related" name.

That'd e.g. open the door to have both a normalized and an original
expression in the explain output. Which would be quite valuable for
some monitoring tools.


> * Unfortunately that does nothing for the problem of how to show
> per-expression JIT info in text format.  Maybe we just shouldn't.
> I do not think that the readability-vs-usefulness tradeoff is going
> to be all that good there, anyway.  Certainly for testing purposes
> it's going to be more useful to examine portions of a structured output.

I think I can live with that, I don't think it's going to be a very
commonly used option. It's basically useful for regression tests, JIT
improvements, and people that want to see whether they can change their
query / schema to make better use of JIT - the latter category won't be
many, I think.

Since this is going to be a default off option anyway, I don't think
we'd need to be as concerned with compatibility. But even leaving
compatibility aside, it's not that clear how to best attach information
in the current text format, without being confusing.

Greetings,

Andres Freund

  15. Re: JIT performance bug/regression & JIT EXPLAIN

    Robert Haas <robertmhaas@gmail.com> — 2020-01-27T19:01:09Z

    On Fri, Nov 15, 2019 at 8:05 PM Maciek Sakrejda <m.sakrejda@gmail.com> wrote:
> On Fri, Nov 15, 2019 at 5:49 AM Robert Haas <robertmhaas@gmail.com> wrote:
> > Personally, I don't care very much about backward-compatibility, or
> > about how hard it is for tools to parse. I want it to be possible, but
> > if it takes a little extra effort, so be it.
>
> I think these are two separate issues. I agree on
> backward-compatibility (especially if we can embed a server version in
> structured EXPLAIN output to make it easier for tools to track format
> differences), but not caring how hard it is for tools to parse? What's
> the point of structured formats, then?

To make the data easy to parse. :-)

I mean, it's clear that, on the one hand, having a format like JSON
that, as has recently been pointed out elsewhere, is parsable by a
wide variety of tools, is advantageous. However, I don't think it
really matters whether the somebody's got to look at a tag called
Flump and match it up with the data in another tag called JIT-Flump,
or whether there's a Flump group that has RegularStuff and JIT tags
inside of it. There's just not much difference in the effort involved.
Being able to parse the JSON or XML using generic code is enough of a
win that the details shouldn't matter that much.

I think if you were going to complain about the limitations of our
current EXPLAIN output format, it'd make a lot more sense to focus on
the way we output expressions. If you want to mechanically parse one
of those expressions and figure out what it's doing - what functions
or operators are involved, and to what they are being applied - you
are probably out of luck altogether, and you are certainly not going
to have an easy time of it. I'm not saying we have to solve that
problem, but I believe it's a much bigger nuisance than the sort of
thing we are talking about here.

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

  16. Re: JIT performance bug/regression & JIT EXPLAIN

    Robert Haas <robertmhaas@gmail.com> — 2020-01-27T19:02:26Z

    On Mon, Jan 27, 2020 at 12:41 PM Andres Freund <andres@anarazel.de> wrote:
> > I do not think that the readability-vs-usefulness tradeoff is going
> > to be all that good there, anyway.  Certainly for testing purposes
> > it's going to be more useful to examine portions of a structured output.
>
> I think I can live with that, I don't think it's going to be a very
> commonly used option. It's basically useful for regression tests, JIT
> improvements, and people that want to see whether they can change their
> query / schema to make better use of JIT - the latter category won't be
> many, I think.

I intensely dislike having information that we can't show in the text
format, or really, that we can't show in every format.

I might be outvoted, but I stand by that position.

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

  17. Re: JIT performance bug/regression & JIT EXPLAIN

    Tom Lane <tgl@sss.pgh.pa.us> — 2020-01-27T21:18:33Z

    Robert Haas <robertmhaas@gmail.com> writes:
>>> I do not think that the readability-vs-usefulness tradeoff is going
>>> to be all that good there, anyway.  Certainly for testing purposes
>>> it's going to be more useful to examine portions of a structured output.

> I intensely dislike having information that we can't show in the text
> format, or really, that we can't show in every format.

Well, if it's relegated to a "jit = detail" option or some such,
the readability objection could be overcome.  But I'm still not clear
on how you'd physically wedge it into the output, at least not in a way
that matches up with the proposal that non-text modes handle this stuff
by producing sub-nodes for the existing types of expression fields.

			regards, tom lane

  18. Re: JIT performance bug/regression & JIT EXPLAIN

    Maciek Sakrejda <m.sakrejda@gmail.com> — 2020-01-27T21:31:06Z

    On Mon, Jan 27, 2020 at 11:01 AM Robert Haas <robertmhaas@gmail.com> wrote:
> On Fri, Nov 15, 2019 at 8:05 PM Maciek Sakrejda <m.sakrejda@gmail.com> wrote:
> > On Fri, Nov 15, 2019 at 5:49 AM Robert Haas <robertmhaas@gmail.com> wrote:
> > > Personally, I don't care very much about backward-compatibility, or
> > > about how hard it is for tools to parse. I want it to be possible, but
> > > if it takes a little extra effort, so be it.
> >
> > I think these are two separate issues. I agree on
> > backward-compatibility (especially if we can embed a server version in
> > structured EXPLAIN output to make it easier for tools to track format
> > differences), but not caring how hard it is for tools to parse? What's
> > the point of structured formats, then?
>
> To make the data easy to parse. :-)
>
> I mean, it's clear that, on the one hand, having a format like JSON
> that, as has recently been pointed out elsewhere, is parsable by a
> wide variety of tools, is advantageous. However, I don't think it
> really matters whether the somebody's got to look at a tag called
> Flump and match it up with the data in another tag called JIT-Flump,
> or whether there's a Flump group that has RegularStuff and JIT tags
> inside of it. There's just not much difference in the effort involved.
> Being able to parse the JSON or XML using generic code is enough of a
> win that the details shouldn't matter that much.

Having a structured EXPLAIN schema that's semantically consistent is
still valuable. At the end of the day, it's humans who are writing the
tools that consume that structured output. Given the sparse structured
EXPLAIN schema documentation, as someone who currently works on
EXPLAIN tooling, I'd prefer a trend toward consistency at the expense
of backward compatibility. (Of course, we should avoid gratuitous
changes.)

But I take back the version number suggestion after reading Tom's
response; that was naïve.

> I think if you were going to complain about the limitations of our
> current EXPLAIN output format, it'd make a lot more sense to focus on
> the way we output expressions.

That would be nice to have, but for what it's worth, my main complaint
would be about documentation (especially around structured formats).
The "Using EXPLAIN" section covers the basics, but understanding what
node types exist, and what fields show up for what nodes and what they
mean--that seems to be a big missing piece (I don't feel entitled to
this documentation; as a structured format consumer, I'm just pointing
out a deficiency). Contrast that with the great wire protocol
documentation. In some ways it's easier to work on native drivers than
on EXPLAIN tooling because the docs are thorough and well organized.

  19. Re: JIT performance bug/regression & JIT EXPLAIN

    Robert Haas <robertmhaas@gmail.com> — 2020-01-28T18:07:29Z

    On Mon, Jan 27, 2020 at 4:18 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
> >>> I do not think that the readability-vs-usefulness tradeoff is going
> >>> to be all that good there, anyway.  Certainly for testing purposes
> >>> it's going to be more useful to examine portions of a structured output.
>
> > I intensely dislike having information that we can't show in the text
> > format, or really, that we can't show in every format.
>
> Well, if it's relegated to a "jit = detail" option or some such,
> the readability objection could be overcome.  But I'm still not clear
> on how you'd physically wedge it into the output, at least not in a way
> that matches up with the proposal that non-text modes handle this stuff
> by producing sub-nodes for the existing types of expression fields.

Well, remember that the text format was the original format. The whole
idea of "groups" was an anachronism that I imposed on the text format
to make it possible to add other formats. It wasn't entirely natural,
because the text format basically indicated nesting by indentation,
and that wasn't going to work for XML or JSON. The text format also
felt free to repeat elements and assume the reader would figure it
out; repeating elements is OK in XML in general, but in JSON it's only
OK if the surrounding context is an array rather than an object.
Anyway, the point is that I (necessarily) started with whatever we had
and found a way to fit it into a structure. It seems like it ought to
be possible to go the other direction also, and figure out how to make
the structured data look OK as text.

Here's Andres's original example:

"Filter": {
  "Expr": "(lineitem.l_shipdate <= '1998-09-18 00:00:00'::timestamp
without time zone)",
  "JIT-Expr": "evalexpr_0_2",
  "JIT-Deform-Scan": "deform_0_3",
  "JIT-Deform-Outer": null,
  "JIT-Deform-Inner": null
}

Right now we show:

Filter: (lineitem.l_shipdate <= '1998-09-18 00:00:00'::timestamp
without time zone)

Andres proposed:

Filter: (lineitem.l_shipdate <= '1998-09-18 00:00:00'::timestamp
without time zone); JIT-Expr: evalexpr_0_2, JIT-Deform-Scan:
deform_0_3

That's not ideal because it's all on one line, but that could be changed:

Filter: (lineitem.l_shipdate <= '1998-09-18 00:00:00'::timestamp
without time zone)
  JIT-Expr: evalexpr_0_2
  JIT-Deform-Scan: deform_0_3

I would propose either including null all the time or omitting it all
the time, so that we would either change the JSON output to...

"Filter": {
  "Expr": "(lineitem.l_shipdate <= '1998-09-18 00:00:00'::timestamp
without time zone)",
  "JIT-Expr": "evalexpr_0_2",
  "JIT-Deform-Scan": "deform_0_3"
}

Or the text output to:

Filter: (lineitem.l_shipdate <= '1998-09-18 00:00:00'::timestamp
without time zone)
  JIT-Expr: evalexpr_0_2
  JIT-Deform-Scan: deform_0_3
  JIT-Deform-Outer: null
  JIT-Deform-Inner: null

You could argue that this is inconsistent because the JSON format
shows a bunch of keys that are essentially parallel, and this text
format makes the Expr key essentially the primary value and the others
secondary. But since the text format is for human beings, and since
human beings are likely to find the Expr key to be the primary piece
of information, maybe that's totally fine.

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