Make set_transaction_id use the PK index for its lookup

[maltoe]

This patch fixes a longstanding (yet undiscovered) bug in the set_transaction_id procedure that caused it to not use the provided index on id spanning xact_id. Reason was the use of CURRVAL in the WHERE clause which caused the query to evaluate every row against a new function call, turning the should-be-fast lookup into a seq scan of the transactions table. Consequently, INSERTs scaled linearly with the size of the transactions table, instead of logarithmically.

Huge 🤦

Function diff as its difficult to see otherwise:

--- old 2024-06-27 12:23:40.078326172 +0200
+++ new 2024-06-27 12:23:36.834933301 +0200
@@ -5,9 +5,15 @@
         /* verify that no previous INSERT within current transaction (with same id) */
         IF
           EXISTS(
+            WITH constants AS (
+              SELECT
+                COALESCE(NEW.id, CURRVAL('#{prefix}.transactions_id_seq')) AS id,
+                COALESCE(NEW.xact_id, pg_current_xact_id()) AS xact_id
+            )
             SELECT 1 FROM #{prefix}.transactions
-            WHERE id = COALESCE(NEW.id, CURRVAL('#{prefix}.transactions_id_seq'))
-            AND xact_id = COALESCE(NEW.xact_id, pg_current_xact_id())
+            JOIN constants
+            ON constants.id = transactions.id
+            AND constants.xact_id = transactions.xact_id
           )
         THEN
           NEW.id = COALESCE(NEW.id, CURRVAL('#{prefix}.transactions_id_seq'));