[BUG] Decimal128 uses a fixed precision of 18 when converting to an arrow array

[jihoonson]

Describe the bug
Hi all. When a decimal128 column is converted to an arrow array, the precision is set to 18 in this line, even though the max precision decimal128 supports is 38. This seems like a bug rather than intentional behavior.

[GregoryKimball]

@codereport Would you please let me know if you think this precision setting should change? Why or why not?

[codereport]

So I'm not 💯 percent confident this is the story but I am pretty certain is is the case. TLDR: It was a bug when it was introduced in #9986.

---

Timeline:

• [REVIEW] Initial fixed_point Column Support #5704 was merged Jul 2020
• Add to/from arrow for fixed point columns #7609 was merged Apr 2021 ⬅️ introduced the 18 for decimal64
• Add support for decimal128 #9483 was merged Nov 2021
• decimal128 Support for to/from_arrow #9986 was merged Apr 2022 ⬅️ introduced the 18 for decimal6128

My guess is this was just a copy and paste issue. Note though that libcudf doesn't actually have the concept of precision, but Python CUDF does seem to set it to 38 in the case of 128.

class Decimal32Dtype(DecimalDtype):
    name = "decimal32"
    MAX_PRECISION = np.floor(np.log10(np.iinfo("int32").max))
    ITEMSIZE = 4


@doc_apply(
    decimal_dtype_template.format(
        size="64",
    )
)
class Decimal64Dtype(DecimalDtype):
    name = "decimal64"
    MAX_PRECISION = np.floor(np.log10(np.iinfo("int64").max))
    ITEMSIZE = 8


@doc_apply(
    decimal_dtype_template.format(
        size="128",
    )
)
class Decimal128Dtype(DecimalDtype):
    name = "decimal128"
    MAX_PRECISION = 38
    ITEMSIZE = 16

My suggestion is to "fix" it, and then write a few tests to see if it works when roundtripping. IIRC, we had planned to test most of the overflow / close to overflow use cases from the Python side of things cause it was much easier that way. Probably why it wasn't caught immediately in the C++ test cases (as they are quite bare bones)

[jihoonson]

Thanks @codereport. I created #14230 to fix this issue.