HyukjinKwon commented on code in PR #59214:
URL: https://github.com/apache/spark/pull/59214#discussion_r4248728844
##########
python/pyspark/sql/variant_utils.py:
##########
@@ -413,10 +417,127 @@ def handle_array(value_pos_list: List[int]) -> str:
if isinstance(value, bytes):
# decoding simply converts byte array to string
return '"' + base64.b64encode(value).decode("utf-8") + '"'
- if isinstance(value, (datetime.date, datetime.datetime)):
+ if isinstance(value, float):
+ return cls._float_to_json(value, is_float32)
+ if isinstance(value, decimal.Decimal):
+ return cls._decimal_to_json(value)
+ if isinstance(value, datetime.datetime):
+ return '"' + cls._timestamp_to_json(value) + '"'
+ if isinstance(value, datetime.date):
return '"' + str(value) + '"'
return str(value)
+ # The helpers below format scalars the same way as the JVM
`Variant.toJson` on JDK 19+, so
+ # that `VariantVal.toJson` and the `to_json` SQL function produce the same
output. On older
+ # JDKs, `to_json` can print a longer, still round-tripping digit string
for some doubles and
+ # floats (see `_float_to_json`).
+
+ @classmethod
+ def _float_to_json(cls, f: float, is_float32: bool) -> str:
+ """
+ Formats a double or float like Java's `Double.toString` and
`Float.toString` on JDK 19+,
+ which print the shortest digits (JDK-4511638). Older JDKs can print a
longer, still
+ round-tripping digit string for some values, e.g. 2^-24 as
5.9604644775390625E-8 instead
+ of 5.960464477539063E-8. Non-finite values are quoted because they are
not valid JSON
+ numbers.
+ """
+ if math.isnan(f):
+ return '"NaN"'
+ if math.isinf(f):
+ return '"Infinity"' if f > 0 else '"-Infinity"'
+
+ def round_trips(s: str) -> bool:
+ x = float(s)
+ if not is_float32:
+ return x == f
+ try:
+ rounded = struct.unpack("<f", struct.pack("<f", x))[0]
+ except OverflowError:
+ return False
+ if rounded != x:
+ # Rounding the decimal to a double first can change which
float32 it rounds to
+ # only when the double is exactly the midpoint between two
float32 values. Every
+ # such midpoint is an exact double, so resolve that tie with
the exact decimal.
+ bits = struct.unpack("<I", struct.pack("<f", abs(rounded)))[0]
+ bits = bits + 1 if abs(x) > abs(rounded) else bits - 1
+ other = math.copysign(struct.unpack("<f", struct.pack("<I",
bits))[0], rounded)
+ midpoint = (rounded + other) / 2
+ if x == midpoint:
+ exact, exact_midpoint = decimal.Decimal(s),
decimal.Decimal(midpoint)
Review Comment:
**Nit (P3):** `decimal.Decimal(midpoint)` converts a float through the
implicit constructor, which consults the caller's thread-local decimal context,
not the explicit `context` created below. If an application turns on Python's
strict float/Decimal mode (`decimal.getcontext().traps[decimal.FloatOperation]
= True`), `toJson()`/`str()` raise `decimal.FloatOperation` here instead of
returning JSON.
I checked this on the current head:
- The test's own vectors `0x15AE43FE` and `0x15AE43FD` raise.
- 12,609 of 20,000 sampled float32 values in [1e6, 1e9) raise, and none
below 1e6 do.
- Doubles never reach this branch.
Before this PR, floats never touched `decimal`, so this exception path is
new. Even with the trap off, the branch sets the `FloatOperation` flag on the
caller's context.
`decimal.Decimal.from_float(midpoint)` is an exact, explicit conversion that
the trap allows. Swapping it in removed every exception and left the output
unchanged on 100,000 samples. My earlier `decimal.Decimal(float(s))` suggestion
had the same problem, sorry about that.
**Verification:**
- **Regression:** Under a caller context with the FloatOperation trap
enabled, the float32 midpoint vectors and a float32 value in the millions
format to their expected JDK 19+ strings instead of raising.
- **Compatibility:** All existing JVM-format vectors and test_variant_type
expectations still produce the same strings under the default and
lowered-precision contexts.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
