Re: Incorrect accounting (n_tup_ins) of non-inserted rows

Hello,

Seems that accounting of insertions with `n_tup_ins` not correct in
case of insertion errors cause by constraints checking (duplicate key
value violates unique constraint):

EXAMPLE:

postgres=# create table t(name text unique);
CREATE TABLE
postgres=# SELECT n_tup_ins FROM pg_stat_user_tables WHERE relname='t';
n_tup_ins
-----------
0

postgres=# insert into t(name) values('a');
INSERT 0 1
postgres=# SELECT n_tup_ins FROM pg_stat_user_tables WHERE relname='t';
n_tup_ins
-----------
1

postgres=# insert into t(name) values('b');
INSERT 0 1
postgres=# SELECT n_tup_ins FROM pg_stat_user_tables WHERE relname='t';
n_tup_ins
-----------
2

postgres=# insert into t(name) values('a');
ERROR: duplicate key value violates unique constraint "t_name_key"
DETAIL: Key (name)=(a) already exists.
postgres=# SELECT n_tup_ins FROM pg_stat_user_tables WHERE relname='t';
n_tup_ins
-----------
3

name
------
a
b

CODE REFERENCE (src/backend/postmaster/pgstat.c)

/* count attempted actions regardless of commit/abort */
tabstat->t_counts.t_tuples_inserted += trans->tuples_inserted;
tabstat->t_counts.t_tuples_updated += trans->tuples_updated;
tabstat->t_counts.t_tuples_deleted += trans->tuples_deleted;

So, is this behavior normal or probably needs to be fixed?

--
Ilya Matveychikov

Ilya Matveychikov <matvejchikov(at)gmail(dot)com> writes:
> Seems that accounting of insertions with `n_tup_ins` not correct in
> case of insertion errors cause by constraints checking (duplicate key
> value violates unique constraint):

You already found one of the many code comments indicating that this is
the intended behavior:

> CODE REFERENCE (src/backend/postmaster/pgstat.c)
> /* count attempted actions regardless of commit/abort */

> So, is this behavior normal or probably needs to be fixed?

No, it's not a bug, and it's not going to be fixed. For many of the
intended applications of those counts (e.g, determining whether
autovacuum/autoanalyze is needed), this is the correct behavior
and ignoring actions of failed transactions would be incorrect.

The live/dead tuple counts do attempt to take transaction success
into account; perhaps looking at those would be more helpful for
your use-case?

regards, tom lane

2016-03-18 16:25 GMT+03:00 Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>:
> No, it's not a bug, and it's not going to be fixed. For many of the
> intended applications of those counts (e.g, determining whether
> autovacuum/autoanalyze is needed), this is the correct behavior
> and ignoring actions of failed transactions would be incorrect.
>

Thank you for clarifying.

> The live/dead tuple counts do attempt to take transaction success
> into account; perhaps looking at those would be more helpful for
> your use-case?

I'll take a look on that counters, thanks.