带SUM聚合的Posgtres CASE条件评估不需要ELSE部分

Question

According to Postgres documentation :

A CASE expression does not evaluate any subexpressions that are not needed to determine the result. For example, this is a possible way of avoiding a division-by-zero failure:

SELECT ... WHERE CASE WHEN x <> 0 THEN y/x > 1.5 ELSE false END;

Why does the following expression return an ERROR: division by zero? - apparently evaluating the else part:

SELECT CASE WHEN SUM(0) = 0 THEN 42 ELSE 43 / 0 END

while

SELECT CASE WHEN SUM(0) = 0 THEN 42 ELSE 43 END

returns 42.

EDIT: So the example above fails because Postgres calculates immutable values (43/0) already in planning phase. Our actual query looks more like this:

case when sum( column1 ) = 0
            then 0
            else round( sum(   price 
                             * hours 
                             / column1 ), 2 )

Although this query doesn't look immutable (depends on actual values), there is still a division by zero error. Of course sum(column1) is actually 0 in our case.

Answer 1

Interesting example. This does have a good explanation. Say you have data like this:

db=# table test;
 column1 | price | hours 
---------+-------+-------
       1 |     2 |     3
       3 |     2 |     1

PostgreSQL executes your SELECT in two passes, first it would calculate all the aggregate functions (like sum() ) present:

db=# select sum(column1) as sum1, sum(price * hours / column1) as sum2 from test;
 sum1 | sum2 
------+------
    4 |    6

And then it would plug those results in your final expression and calculate the actual result:

db=# with temp as (
db(#     select sum(column1) as sum1, sum(price * hours / column1) as sum2 from test
db(# ) select case when sum1 = 0 then 0 else round(sum2, 2) end from temp;
 round 
-------
  6.00

Now clearly if there's an error in the first aggregate pass, it never reaches the CASE statement.

So this isn't really a problem in documentation about the CASE statement -- it applies to all conditional constructs -- but about the way aggregates are processed in a SELECT statement. This kind of problem cannot occur in any other context because aggregates are only allowed in SELECT.

But the documentation does need updating in this case too. The right documentation in this instance is " the general processing of SELECT ". Step #4 there talks about GROUP BY and HAVING clauses, but it actually evaluates any aggregate functions in this step as well, regardless of GROUP BY/HAVING. And your CASE statement is evaluated in step #5.

Solution

The common solution, if you want to ignore aggregate inputs that would otherwise cause a division by zero, use the nullif() construct to turn them into NULLs:

round( sum(   price 
            * hours 
            / nullif(column1, 0) ), 2 )

PostgreSQL 9.4 will introduce a new FILTER clause for aggregates, which can also be used for this purpose:

round( sum(   price 
            * hours 
            / column1
          ) filter (where column1!=0), 2 )