Is it safe to insert data from inside of a CTE expression? - postgresql

Question

Background
I have a function that generates detail rows and inserts these into a detail table. I want to automatically ensure that the referenced master row is available before inserting the detail rows.
A BEFORE INSERT trigger does the job, but unfortunately the job is done too well. If a detail row is prevented due to a unique index, the master row will still be inserted, leaving me with a master without children (I don't want that).
I managed to solve this by inserting the master rows inside of a cte and then inserting the detail rows in the actual query. This works, but I'm worried that it's not a safe way of doing it.
INSERT rows from inside of a CTE expression
Here is the code to try this out. The input_cte is just generating static dummy data in the example. This particular example could be built with separate static insert SQLs, but in my real case the input_cte is dynamic.
Is this a safe way to solve this, or is it out of spec and could blow up in the next PG version?
CREATE TABLE master (
id INT NOT NULL GENERATED BY DEFAULT AS IDENTITY,
PRIMARY KEY (id)
);
CREATE TABLE detail (
id INT NOT NULL GENERATED BY DEFAULT AS IDENTITY,
master_id INT,
PRIMARY KEY (id, master_id)
);
ALTER TABLE detail
ADD CONSTRAINT detail_master_id_fkey
FOREIGN KEY (master_id)
REFERENCES master (id);
WITH input_cte AS (
SELECT 1 AS master_id, 1 AS detail_id
UNION SELECT 1, 2
UNION SELECT 2, 1
),
insert_cte AS (
--Ignore conflicts, as the master row could already exist
INSERT INTO master (id)
SELECT DISTINCT master_id
FROM input_cte
ON CONFLICT DO NOTHING
)
INSERT INTO detail (id, master_id)
SELECT detail_id, master_id
FROM input_cte;
SELECT * FROM master;
SELECT * FROM detail;
Edit
Bummer.. I just realised that this method will also insert master rows even if the detail rows would be stopped by my unique index.
I see two options. Which should I choose?
Check exactly what I can insert before trying to insert. I.e. do the same thing as my unique index does.
Go ahead with the above solution or the BEFORE INSERT trigger concept and then clean up unused master rows afterwards with a separate DELETE query.
Edit 2 as a reponse to Haleemur Ali's comment
I agree with Haleemur, but in my case it's a bit more complicated than the small example I created. The unique key on my detail can actually have null values. The index in my detail table (project_sequence) looks like this to enable null values:
CREATE UNIQUE INDEX
project_sequence_unique_combinations ON main.project_sequence
(project_id, controlpoint_type_id, COALESCE(drawing_id, 0), COALESCE(layer_guid, '00000000-0000-0000-0000-000000000000'));
Due to possible NULL values I cannot use these fields in my primary key, so I have a surrogate integer key. I'm calculating these key values in my CTE so they will always be unique. I.e. they can be inserted into the master table (sequence) even if the detail rows gets stopped by the unique index.*
To clarify I've inserted my actual code below. This code works as it should, but it sure would be nice to be able to utilize the new fancy DEFERRED and REFERENCES triggers.
SELECT COALESCE(MAX(id), 0)
INTO _max_sequence_id
FROM main.sequence;
WITH cte AS (
SELECT
d.project_id,
pct.controlpoint_type_id,
d.id as drawing_id,
DENSE_RANK() OVER(ORDER BY d.id, sequence_group_key) + _max_sequence_id + 1 AS new_sequence_id
FROM main.drawing d
CROSS JOIN main.project_controlpoint_type pct
--This left JOIN along with "ps.project_id IS NULL" is my
--current solution, i.e. its "option 1" from above.
LEFT JOIN main.project_sequence ps
ON ps.project_id = d.project_id
AND ps.drawing_id = d.id
AND ps.controlpoint_type_id = pct.controlpoint_type_id
WHERE d.project_id = _project_id
AND pct.project_id = _project_id
AND pct.sequence_level_id = 2
AND ps.project_id IS NULL
),
insert_sequence_cte AS (
INSERT INTO main.sequence
(id, project_id, last_value)
SELECT DISTINCT cte.new_sequence_id, cte.project_id, 0
FROM cte
ON CONFLICT DO NOTHING
)
INSERT INTO main.project_sequence
(project_id, controlpoint_type_id, drawing_id, sequence_id)
SELECT
project_id,
controlpoint_type_id,
drawing_id,
new_sequence_id
FROM cte;

Answer 1

The manual page on WITH queries states that your use case is legitimate and supported:
You can use data-modifying statements (INSERT, UPDATE, or DELETE) in WITH.
and
... data-modifying statements are only allowed in WITH clauses that are attached to the top-level statement. However, normal WITH visibility rules apply, so it is possible to refer to the WITH statement's output from the sub-SELECT.
Further:
If a data-modifying statement in WITH lacks a RETURNING clause, then it forms no temporary table and cannot be referred to in the rest of the query. Such a statement will be executed nonetheless.

Answer 2

If the rule "no master without a detail" is important to your model, you should let the DB enforce it. This will free you from "auto-inserting" master to reduce the potential for error and let you use conventional methods again.
Take a look at CONSTRAINT TRIGGERS. They allow you to just detect and reject violations of your no-master-without-detail rule while leaving the actual INSERTs to application code.
Your use case would need a CONSTRAINT TRIGGER on your master table that is DEFERRABLE INITIALLY DEFERRED. This allows you to INSERT master, then INSERT detail and still be sure that the transaction only commits if all is consistent.
From the manual linked above:
Constraint triggers must be AFTER ROW triggers on plain tables (not foreign tables). They can be fired either at the end of the statement causing the triggering event, or at the end of the containing transaction; in the latter case they are said to be deferred.
You'll need two triggers, one handling INSERT/UPDATE on master and another one handling DELETE/UPDATE on detail:
CREATE CONSTRAINT TRIGGER trigger_assert_master_has_detail
AFTER INSERT OR UPDATE OF id ON master
DEFERRABLE INITIALLY DEFERRED FOR EACH ROW
EXECUTE PROCEDURE assert_master_has_detail();
CREATE CONSTRAINT TRIGGER trigger_assert_no_leftover_master
AFTER DELETE OR UPDATE OF master_id ON detail
DEFERRABLE INITIALLY DEFERRED FOR EACH ROW
EXECUTE PROCEDURE assert_no_leftover_master();
Note that UPDATEs will only fire the trigger if they concern the FK/PK column.
The two trigger functions will then check if there are 1-n details for the master:
CREATE FUNCTION assert_master_has_detail() RETURNS trigger
AS
$$
BEGIN
IF NOT EXISTS (SELECT 1 FROM detail WHERE master_id = NEW.id)
THEN
RAISE EXCEPTION 'no detail for master_id=%', NEW.id;
ELSE
RETURN NEW;
END IF;
END;
$$
LANGUAGE plpgsql;
CREATE FUNCTION assert_no_leftover_master() RETURNS trigger
AS
$$
BEGIN
IF NOT EXISTS (SELECT 1 FROM detail WHERE master_id = OLD.master_id)
AND EXISTS (SELECT 1 FROM master WHERE id = OLD.master_id)
THEN
RAISE EXCEPTION 'last detail for master_id=% removed, but master still exists', OLD.master_id;
ELSE
RETURN NULL;
END IF;
END;
$$
LANGUAGE plpgsql;
Example of a violation:
INSERT INTO master
VALUES (1);
-- ERROR: no detail for master_id=1
-- CONTEXT: PL/pgSQL function assert_master_has_detail() line 5 at RAISE
and a legal scenario:
INSERT INTO master
VALUES (1);
INSERT INTO detail
VALUES (10, 1);
-- trigger fired at end of transaction, finds everything is OK
Here's a complete solution as dbfiddle.