PostgreSQL 9.3: Primary key violated by a trigger INSERT

Question

My problem

Consider a table t with many frequent updates from users, from which only the last few are relevant.

In order to keep the table size reasonable, whenever a new row is inserted old rows from the same user_id are deleted. In order to keep an archive, the row is also written to t_history.

Both t and t_history have the same schema, in which id is a bigserial with a primary key constraint.

Implementation

Stored procedure

CREATE FUNCTION update_t_history()
RETURNS trigger
AS
$$
declare
BEGIN
 -- Insert the row to the t_history table. `id` is autoincremented
 INSERT INTO t_history (a, b, c, ...)
 VALUES (NEW.a, NEW.b, NEW.c, ...);
 -- Delete old rows from the t table, keep the newest 10 
 DELETE FROM t WHERE id IN (
 SELECT id FROM t 
 WHERE user_id = NEW.user_id 
 ORDER BY id DESC
 OFFSET 9);
 RETURN NEW;
END;
$$
LANGUAGE plpgsql;

Corresponding insertion trigger:

CREATE TRIGGER t_insertion_trigger
AFTER INSERT ON t
FOR EACH ROW
EXECUTE PROCEDURE update_t_history();

The error

The trigger works well, but when I run a few dozen insertions in a single transaction, I get the following error:

BEGIN
ERROR: duplicate key value violates unique constraint "t_history_pkey"
DETAIL: Key (id)=(196) already exists.

Updates

• The id field in both tables (from \d+ t):
  • id|bigint|not null default nextval('t_id_seq'::regclass)
  • "t_pkey" PRIMARY KEY, btree (id)
• PostgreSQL version is 9.3.

Any idea why the stored procedure breaks the primary key constraint in transactions?

• Craig Ringer: Multiple single-value inserts in one transaction, or a single multi-valued insert statement?
• Erwin Brandstetter: Always your version of Postgres, please. Also, how exactly is t_history.id autoincremented? Best provide the table definition you get with \d tbl in psql. And asre you sure you are not copying t.id in the INSERT statement?
• Adam Matan: @ErwinBrandstetter Thanks! Updated my question accordingly.

Answer 1

Why is t_history.id auto-incremented in the first place? If "both t and t_history have the same schema", and t.id is a serial PK, you can just copy whole rows.

I would also suggest you only copy rows you actually delete from t to t_history - in a data-modifying CTE. This way you do not have overlapping rows (which might be part of the problem).

CREATE FUNCTION update_t_history()
 RETURNS trigger AS
$func$
BEGIN
 -- Keep the newest 10, move older rows to t_history
 WITH del AS (
 DELETE FROM t
 USING (
 SELECT id
 FROM t 
 WHERE user_id = NEW.user_id 
 ORDER BY id DESC
 OFFSET 10 -- to keep 10 (not 9)
 FOR UPDATE -- avoid race condition
 ) d
 WHERE t.id = d.id
 RETURNING t.*
 )
 INSERT INTO t_history 
 SELECT * FROM del; -- copy whole row
 RETURN NULL; -- irrelevant in AFTER trigger
END
$func$ LANGUAGE plpgsql;

The new row is already visible in an AFTER trigger.

• Adam Matan: +1 Thanks! I will try your solution and write you back. I don't see how can we have overlapping rows on INSERT, because we INSERT the same row to both tables once, and t never had overlap issues. BTW, I didn't know FOR UPDATE - will read about it.
• Erwin Brandstetter: @AdamMatan: Note, RETURNING was missing in my answer, added now.
• ypercubeᵀᴹ: @AdamMatan if the insert is done on both tables, then any update will have to be done on both as well. The deletes of course will be trivial.
• Adam Matan: I followed your advice, and simply added id and NEW.id to the INSERT statement in my code. It did the trick - I wonder why.