Subject: Re: PoC/WIP: Extended statistics on expressions
To: Zhihong Yu <zyu@yugabyte.com>
Cc: Justin Pryzby <pryzby@telsasoft.com>,
 Dean Rasheed <dean.a.rasheed@gmail.com>,
 PostgreSQL-development <pgsql-hackers@postgresql.org>
References: 
 <CAEZATCXAPrrOKwEsyZKQ4uzzJQWBCt6QAvOcgqRGdWwT1zb+rQ@mail.gmail.com>
 <329203a4-7cad-a0e3-d5a8-ce70872c8156@enterprisedb.com>
 <CAEZATCWGYooC_Tzt8sQo7760ZEsguj0JGoMLhkANrDwJ9CY20Q@mail.gmail.com>
 <2567ae7e-e1dd-48b5-054d-ed6ff81c0335@enterprisedb.com>
 <CAEZATCW_7nfeHmLtV6TzSfmFG+i=io_Qo0NL8BiBnmEqq3CQnQ@mail.gmail.com>
 <CAEZATCUfMM43sv+RFDqyczQpYMP7SZLd4EexFEEUX4buWRNnrw@mail.gmail.com>
 <afcd65d7-7701-a469-b064-9c62d63ef437@enterprisedb.com>
 <20210108023537.GA19743@telsasoft.com>
 <d7c93270-20a0-93cb-f15e-46d7b51b35c1@enterprisedb.com>
 <5305f9cf-3927-06c4-cd5b-e46cbc58bc7f@enterprisedb.com>
 <20210116232208.GB8560@telsasoft.com>
 <2f2a3b3b-93e4-723f-144e-bc21f10d6b61@enterprisedb.com>
 <CALNJ-vRz-hsCFVNQLm7Amxse+J=dLwSuwbJNeYOk-9mwvh_AyA@mail.gmail.com>
From: Tomas Vondra <tomas.vondra@enterprisedb.com>
Message-ID: <01a2e09f-843a-824b-42c4-4d31d173b4a7@enterprisedb.com>
Date: Mon, 18 Jan 2021 00:14:30 +0100
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:78.0) Gecko/20100101
 Thunderbird/78.6.0
MIME-Version: 1.0
In-Reply-To: 
 <CALNJ-vRz-hsCFVNQLm7Amxse+J=dLwSuwbJNeYOk-9mwvh_AyA@mail.gmail.com>
Content-Type: text/plain; charset=utf-8; format=flowed
Content-Language: en-US
Content-Transfer-Encoding: 8bit
Precedence: bulk


On 1/17/21 3:55 AM, Zhihong Yu wrote:
> Hi,
> +    * Check that only the base rel is mentioned.  (This should be dead code
> +    * now that add_missing_from is history.)
> +    */
> +   if (list_length(pstate->p_rtable) != 1)
> 
> If it is dead code, it can be removed, right ?
> 

Maybe. The question is whether it really is dead code. It's copied from 
transformIndexStmt so I kept it for now.

> For statext_mcv_clauselist_selectivity:
> 
> +                   // bms_free(list_attnums[listidx]);
>  > The commented line can be removed.
>

Actually, this should probably do list_free on the list_exprs.

> +bool
> +examine_clause_args2(List *args, Node **exprp, Const **cstp, bool 
> *expronleftp)
> 
> Better add some comment for examine_clause_args2 since there 
> is examine_clause_args() already.
> 

Yeah, this probably needs a better name. I have a feeling this may need 
a refactoring to reuse more of the existing code for the expressions.

> +   if (!ok || stats->compute_stats == NULL || stats->minrows <= 0)
> 
> When would stats->minrows have negative value ?
> 

The typanalyze function (e.g. std_typanalyze) can set it to negative 
value. This is the same check as in examine_attribute, and we need the 
same protections I think.

> For serialize_expr_stats():
> 
> +   sd = table_open(StatisticRelationId, RowExclusiveLock);
> +
> +   /* lookup OID of composite type for pg_statistic */
> +   typOid = get_rel_type_id(StatisticRelationId);
> +   if (!OidIsValid(typOid))
> +       ereport(ERROR,
> 
> Looks like the table_open() call can be made after the typOid check.
> 

Probably, but this failure is unlikely (should never happen) so I don't 
think this makes any real difference.

> +       Datum       values[Natts_pg_statistic];
> +       bool        nulls[Natts_pg_statistic];
> +       HeapTuple   stup;
> +
> +       if (!stats->stats_valid)
> 
> It seems the local arrays can be declared after the validity check.
> 

Nope, that would be against C99.

> +           if (enabled[i] == STATS_EXT_NDISTINCT)
> +               ndistinct_enabled = true;
> +           if (enabled[i] == STATS_EXT_DEPENDENCIES)
> +               dependencies_enabled = true;
> +           if (enabled[i] == STATS_EXT_MCV)
> 
> the second and third if should be preceded with 'else'
> 

Yeah, although this just moves existing code. But you're right it could 
use else.

> +ReleaseDummy(HeapTuple tuple)
> +{
> +   pfree(tuple);
> 
> Since ReleaseDummy() is just a single pfree call, maybe you don't need 
> this method - call pfree in its place.
> 

No, that's not possible because the freefunc callback expects signature

     void (*)(HeapTupleData *)

and pfree() does not match that.


thanks

-- 
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company