numeric type

numeric data expresses an exact number with user-defined precision and scale.

Detail Info
Size ~32 bytes
Aliases dec, decimal
Catalog name pg_catalog.numeric
OID 1700
Precision 39
Scale [0, 39]

Syntax

Numeric values

int . frac e E - exp
Field Use
Eexp Multiply the number preceding E by 10exp

Numeric definitions

numeric ( precision , scale )
Field Use
precision Ignored: All numeric values in Materialize have a precision of 39.
scale The total number of fractional decimal digits to track, e.g. .321 has a scale of 3. scale cannot exceed the maximum precision.

Details

Input

Materialize assumes untyped numeric literals are numeric if they:

  • Contain decimal points or e-notation.
  • Exceed bigint’s maximum or minimum values.

Materialize does not accept any numeric literals that exceed 39 digits of precision.

Output

Materialize trims all trailing zeroes off of numeric values, irrespective of their specified scale. This behavior lets us perform byte-level equality when comparing rows of data, where we would otherwise need to decode rows’ columns' values for comparisons.

Precision

All numeric values have a precision of 39, which cannot be changed. For ease of use, Materialize accepts input of any value <= 39, but ignores the specified value.

Note that the leading zeroes in values between -1 and 1 (i.e. (-1, 1)) are not counted digits of precision.

For details on exceeding the numeric type’s maximum precision, see Rounding.

Scale

By default, numeric values do not have a specified scale, so values can have anywhere between 0 and 39 digits after the decimal point. For example:

CREATE TABLE unscaled (c NUMERIC);
INSERT INTO unscaled VALUES
  (987654321098765432109876543210987654321),
  (9876543210987654321.09876543210987654321),
  (.987654321098765432109876543210987654321);

SELECT c FROM unscaled;

                     c
-------------------------------------------
   987654321098765432109876543210987654321
 0.987654321098765432109876543210987654321
  9876543210987654321.09876543210987654321

However, if you specify a scale on a numeric value, values will be rescaled appropriately. If the resulting value exceeds the maximum precision for numeric types, you’ll receive an error.

CREATE TABLE scaled (c NUMERIC(39, 20));

INSERT INTO scaled VALUES
  (987654321098765432109876543210987654321);
ERROR:  numeric field overflow
INSERT INTO scaled VALUES
  (9876543210987654321.09876543210987654321),
  (.987654321098765432109876543210987654321);

SELECT c FROM scaled;

                    c
------------------------------------------
                   0.98765432109876543211
 9876543210987654321.09876543210987654321

Rounding

numeric operations will always round off fractional values to limit their values to 39 digits of precision.

SELECT 2 * 9876543210987654321.09876543210987654321 AS rounded;

                 rounded
------------------------------------------
 19753086421975308642.1975308642197530864

However, if a value exceeds is >= 1E39 or <= -1E39, it generates an overflow, i.e. it will not be rounded.

Overflow

Operations generating values >= 1E39 or <= -1E39 are considered overflown.

Underflow

Operations generating values within the range (-1E-40, 1E-40) are considered underflown.

Aggregations (sum)

tl;dr If you use sum on numeric values, retrieving values of Infinity or -Infinity from the operation signals you have overflown the numeric type. At this point, you should start retracting values or risk causing Materialize to panic.

Materialize’s dataflow engine (Differential) requires operations to retain their commutativity and associativity––a simple way to think about this invariant is that if you add a series of values, and then subtract the same values in some random order, you must deterministically reach zero.

While other types in Materialize achieve this by simply allowing overflows (i.e. going from the maximum 64-bit signed integer to the minimum 64-bit signed integer), the numeric type does not support an equivalent operation.

To provide the invariants that Differential requires, numeric values are instead aggregated into an even larger type behind the scenes––one that supports twice as many digits of precision. This way, if your aggregation exceeds the numeric type’s bounds, we can continue to track the magnitude of excess.

To signal that the numeric value is in this state, the operation begins returning Infinity or -Infinity. At this point, you must either subtract or retract values to return representable numeric values.

However, because the underlying aggregated type is not of infinite precision, it too can overflow. In this case, Materialize will instead panic and crash. To avoid crashing, always immediately retract or subtract values when encountering infinite results from sum.

Valid casts

From numeric

You can cast numeric to:

To numeric

You can cast from the following types to numeric:

Examples

SELECT 1.23::numeric AS num_v;
 num_v
-------
  1.23

SELECT 1.23::numeric(38,3) AS num_38_3_v;
 num_38_3_v
------------
      1.230

SELECT 1.23e4 AS num_w_exp;
 num_w_exp
-----------
     12300
Back to top ↑