Return pre-UPDATE column values using SQL only

I posted a related question, but this is another part of my puzzle.

I would like to get the OLD value of a column from a row that was UPDATEd - WITHOUT using triggers (nor stored procedures, nor any other extra, non -SQL/-query entities).

I have a query like this:

   UPDATE my_table
      SET processing_by = our_id_info  -- unique to this worker
    WHERE trans_nbr IN (
                        SELECT trans_nbr
                          FROM my_table
                         GROUP BY trans_nbr
                        HAVING COUNT(trans_nbr) > 1
                         LIMIT our_limit_to_have_single_process_grab
                       )
RETURNING row_id;

If I could do FOR UPDATE ON my_table at the end of the subquery, that'd be divine (and fix my other question/problem). But that won't work: can't combine this with GROUP BY (which is necessary for figuring out the count). Then I could just take those trans_nbr's and do a query first to get the (soon-to-be-) former processing_by values.

I've tried doing like:

   UPDATE my_table
      SET processing_by = our_id_info -- unique to this worker
     FROM my_table old_my_table
     JOIN (
             SELECT trans_nbr
               FROM my_table
           GROUP BY trans_nbr
             HAVING COUNT(trans_nbr) > 1
              LIMIT our_limit_to_have_single_process_grab
          ) sub_my_table
       ON old_my_table.trans_nbr = sub_my_table.trans_nbr
    WHERE     my_table.trans_nbr = sub_my_table.trans_nbr
      AND my_table.processing_by = old_my_table.processing_by
RETURNING my_table.row_id, my_table.processing_by, old_my_table.processing_by

But that can't work; old_my_table is not visible outside the join; the RETURNING clause is blind to it.

I've long since lost count of all the attempts I've made; I have been researching this for literally hours.

If I could just find a bullet-proof way to lock the rows in my subquery - and ONLY those rows, and WHEN the subquery happens - all the concurrency issues I'm trying to avoid would disappear ...

UPDATE: I had a typo in the non-generic code of the above. I retried after Erwin Brandstetter suggested it should work. Since it took me so long to find this sort of solution, perhaps my embarrassment is worth it? At least this is on SO for posterity now... :>

What I now have (that works) is like this:

   UPDATE my_table
      SET processing_by = our_id_info -- unique to this worker
     FROM my_table AS old_my_table
    WHERE trans_nbr IN (
                          SELECT trans_nbr
                            FROM my_table
                        GROUP BY trans_nbr
                          HAVING COUNT(*) > 1
                           LIMIT our_limit_to_have_single_process_grab
                       )
      AND my_table.row_id = old_my_table.row_id
RETURNING my_table.row_id, my_table.processing_by, old_my_table.processing_by AS old_processing_by

The COUNT(*) is per a suggestion from Flimzy in a comment on my other (linked above) question.

Please see my other question for correctly implementing concurrency and even a non-blocking version; THIS query merely shows how to get the old and new values from an update, ignore the bad/wrong concurrency bits.

Solution 1:

Problem

The manual explains:

The optional RETURNING clause causes UPDATE to compute and return value(s) based on each row actually updated. Any expression using the table's columns, and/or columns of other tables mentioned in FROM, can be computed. The new (post-update) values of the table's columns are used. The syntax of the RETURNING list is identical to that of the output list of SELECT.

Bold emphasis mine. There is no way to access the old row in a RETURNING clause. You can work around this restriction with a trigger or a separate SELECT before the UPDATE wrapped in a transaction or wrapped in a CTE as was commented.

However, what you are trying to achieve works perfectly fine if you join to another instance of the table in the FROM clause:

Solution without concurrent writes

UPDATE tbl x
SET    tbl_id = 23
     , name = 'New Guy'
FROM   tbl y                -- using the FROM clause
WHERE  x.tbl_id = y.tbl_id  -- must be UNIQUE NOT NULL
AND    x.tbl_id = 3
RETURNING y.tbl_id AS old_id, y.name AS old_name
        , x.tbl_id          , x.name;

Returns:

 old_id | old_name | tbl_id |  name
--------+----------+--------+---------
  3     | Old Guy  | 23     | New Guy

The column(s) used to self-join must be UNIQUE NOT NULL. In the simple example, the WHERE condition is on the same column tbl_id, but that's just coincidence. Works for any conditions.

I tested this with PostgreSQL versions from 8.4 to 13.

It's different for INSERT:

INSERT INTO ... FROM SELECT ... RETURNING id mappings

Solutions with concurrent write load

There are various ways to avoid race conditions with concurrent write operations on the same rows. (Note that concurrent write operations on unrelated rows are no problem at all.) The simple, slow and sure (but expensive) method is to run the transaction with SERIALIZABLE isolation level:

BEGIN ISOLATION LEVEL SERIALIZABLE;
UPDATE ... ;
COMMIT;

But that's probably overkill. And you need to be prepared to repeat the operation in case of a serialization failure.

Simpler and faster (and just as reliable with concurrent write load) is an explicit lock on the one row to be updated:

UPDATE tbl x
SET    tbl_id = 24
     , name = 'New Gal'
FROM  (SELECT tbl_id, name FROM tbl WHERE tbl_id = 4 FOR UPDATE) y 
WHERE  x.tbl_id = y.tbl_id
RETURNING y.tbl_id AS old_id, y.name AS old_name
        , x.tbl_id          , x.name;

Note how the WHERE condition moved to the subquery (again, can be anything), and only the self-join (on UNIQUE NOT NULL column(s)) remains in the outer query. This guarantees that only rows locked by the inner SELECT are processed. The WHERE conditions might resolve to a different set of rows a moment later.

See:

Atomic UPDATE .. SELECT in Postgres

db<>fiddle here
_{Old sqlfiddle}

Solution 2:

You can use a SELECT subquery.

Example: Update a user's email RETURNING the old value.

RETURNING Subquery

UPDATE users SET email = '[email protected]' WHERE id = 1
RETURNING (SELECT email FROM users WHERE id = 1);

PostgreSQL WITH Query (Common Table Expressions)
```
WITH u AS (
    SELECT email FROM users WHERE id = 1
)
UPDATE users SET email = '[email protected]' WHERE id = 1
RETURNING (SELECT email FROM u);
```
This has worked several times on my local database without fail, but I'm not sure if the SELECT in WITH is guaranteed to consistently execute before the UPDATE since "the sub-statements in WITH are executed concurrently with each other and with the main query."

Solution 3:

The CTE variant as proposed by @MattDiPasquale should work too.
With the comfortable means of a CTE I would be more explicit, though:

WITH sel AS (
   SELECT tbl_id, name FROM tbl WHERE tbl_id = 3  -- assuming unique tbl_id
   )
, upd AS (
   UPDATE tbl SET name = 'New Guy' WHERE tbl_id = 3
   RETURNING tbl_id, name
   )
SELECT s.tbl_id AS old_id, s.name As old_name
     , u.tbl_id, u.name
FROM   sel s, upd u;

Without testing I claim this works: SELECT and UPDATE see the same snapshot of the database. The SELECT is bound to return the old values (even if you place the CTE after the CTE with the UPDATE), while the UPDATE returns the new values by definition. Voilá.

But it will be slower than my first answer.

Solution 4:

when faced with this dilemma I added junk columns to the table and then I copy the old values into the junk columns (which I then return) when I update the record. this bloats the table a bit but avoids the need for joins.