Why does “SELECT POWER(10.0, 38.0);” throw an arithmetic overflow error?

Question

I'm updating my IDENTITY overflow check script to account for DECIMAL and NUMERIC IDENTITY columns.

As part of the check I compute the size of the data type's range for every IDENTITY column; I use that to calculate what percentage of that range has been exhausted. For DECIMAL and NUMERIC the size of that range is 2 * 10^p - 2 where p is the precision.

I created a bunch of test tables with DECIMAL and NUMERIC IDENTITY columns and attempted to calculate their ranges as follows:

SELECT POWER(10.0, precision)
FROM sys.columns
WHERE 
       is_identity = 1
   AND type_is_decimal_or_numeric
;

This threw the following error:

Msg 8115, Level 16, State 6, Line 1
Arithmetic overflow error converting float to data type numeric. 

I narrowed it down to the IDENTITY columns of type DECIMAL(38, 0) (i.e. with the maximum precision), so I then tried the POWER() calculation directly on that value.

All of the following queries

SELECT POWER(10.0, 38.0);
SELECT CONVERT(FLOAT, (POWER(10.0, 38.0)));
SELECT CAST(POWER(10.0, 38.0) AS FLOAT);

also resulted in the same error.

• Why does SQL Server try to convert the output of POWER(), which is of type FLOAT, to NUMERIC (especially when FLOAT has a higher precedence)?
• How can I dynamically calculate the range of a DECIMAL or NUMERIC column for all possible precisions (including p = 38, of course)?

Answer 1

It seems that despite the implication in BOL that the left hand operand will be implicitly cast to float that this is not the case. The output of POWER() is cast to the type of the left hand operand, which is DECIMAL if you use 10.0. Using an explicit float works fine.

SELECT POWER(1e1, 38);
SELECT POWER(CAST(10 as float), 38.0);

Answer 2

Instead of meddling with Martin's answer any further, I'll add the rest of my findings regarding POWER() here.

Hold on to your knickers.

Preamble

First, I present to you exhibit A, the MSDN documentation for POWER():

Syntax

POWER ( float_expression , y )

Arguments

float_expression Is an expression of type float or of a type that can be implicitly converted to float.

Return Types

Same as float_expression.

You may conclude from reading that last line that POWER()'s return type is FLOAT, but read again. float_expression is "of type float or of a type that can be implicitly converted to float". So, despite its name, float_expression may actually be a FLOAT, a DECIMAL, or an INT. Since the output of POWER() is the same as that of float_expression, it too may also be one of those types.

So we have a scalar function with return types that depend on the input. Could it be?

Observations

I present to you exhibit B, a test demonstrating that POWER() casts its output to different data types depending on its input. Please read this script with a faux Jamaican accent for maximum derived benefit.

SELECT 
    POWER(10, 3)             AS int_man
  , POWER(1000000000000, 3)  AS numeric0_man     -- one trillion
  , POWER(10.0, 3)           AS numeric1_man
  , POWER(10.12305, 3)       AS numeric5_man
  , POWER(1e1, 3)            AS float_man
INTO power_test_man;

EXECUTE sp_help power_test_man;

DROP TABLE power_test_man;

The relevant results are:

Column_name    Type      Length    Prec     Scale
-------------------------------------------------
int_man        int       4         10       0
numeric0_man   numeric   17        38       0
numeric1_man   numeric   17        38       1
numeric5_man   numeric   17        38       5
float_man      float     8         53       NULL

What appears to be happening is that POWER() casts float_expression into the smallest type that fits it, not including BIGINT.

Therefore, SELECT POWER(10.0, 38); fails with an overflow error because 10.0 gets cast to NUMERIC(38, 1) which isn't big enough to hold the result of 10³⁸. That's because 10³⁸ expands to take 39 digits before the decimal, whereas NUMERIC(38, 1) can store 37 digits before the decimal plus one after it. Therefore, the maximum value NUMERIC(38, 1) can hold is 10³⁷ - 0.1.

Armed with this understanding I can concoct another overflow failure as follows.

SELECT POWER(1000000000, 3);    -- one billion

One billion (as opposed to the one trillion from the first example, which is cast to NUMERIC(38, 0)) is just small enough to fit in an INT. One billion raised to the third power, however, is too big for INT, hence the overflow error.

Several other functions exhibit similar behavior, where their output type is dependent on their input:

• Mathematical functions: POWER(), CEILING(), FLOOR(), RADIANS(), DEGREES(), and ABS()
• System functions and expressions: NULLIF(), ISNULL(), COALESCE(), IIF(), CHOOSE(), and CASE expressions
• Arithmetic operators: Both SELECT 2 * @MAX_INT; and SELECT @MAX_SMALLINT + @MAX_SMALLINT;, for example, result in arithmetic overflows when the variables are of the named data type.

Conclusion

In this particular case, the solution is to use SELECT POWER(1e1, precision).... This will work for all possible precisions since 1e1 gets cast to FLOAT, which can hold ridiculously large numbers.

Since these functions are so commonplace, it's important to understand that your results may be rounded or may cause overflow errors due to their behavior. If you expect or rely on a specific data type for your output, explicitly cast the relevant input as necessary.

So kids, now that you know this, you may go forth and prosper.

Yeah man.