[BUG] partition_on_multi_column_values rejects batch identifiers whose value is 0, "" or False #12239

What is wrong

PandasDataPartitioner.partition_on_multi_column_values and SparkDataPartitioner.partition_on_multi_column_values decide whether a batch identifier was supplied by testing the value for truthiness:

value = batch_identifiers.get(column_name)
if not value:
    raise ValueError(...)

An identifier that is present and legitimate — 0, "", False, 0.0 — is therefore reported as missing, and the batch is never read. The message itself contradicts the situation:

ValueError: In order for PandasExecution to `_partition_on_multi_column_values`, all values in
column_names must also exist in batch_identifiers. m was not found in batch_identifiers.

m was found — its value is 0.

Neither neighbouring implementation behaves this way. partition_on_column_value (pandas_data_partitioner.py:168) indexes the dict with no truthiness gate, and SqlAlchemyDataPartitioner.partition_on_multi_column_values (sqlalchemy_data_partitioner.py:317-328) emits column == value for every identifier it is given. So identical partitioner_kwargs partition correctly on a SQL backend and raise on pandas and Spark.

Reproduction

measured on develop (8bfb2b2), Python 3.11.15, pandas 3.0.5; the Spark case on pyspark 3.5.5 with JVM 11:

import pandas as pd
from great_expectations.core.batch_spec import RuntimeDataBatchSpec
from great_expectations.execution_engine import PandasExecutionEngine

df = pd.DataFrame({"y": [2020, 2020, 2019, 0, 0], "m": [1, 0, 0, 0, 2]})
PandasExecutionEngine().get_batch_data(
    RuntimeDataBatchSpec(
        batch_data=df,
        partitioner_method="_partition_on_multi_column_values",
        partitioner_kwargs={"column_names": ["m"], "batch_identifiers": {"m": 0}},
    )
)

Before the fix:

multi-column  {"m": 0}   -> ValueError: ... m was not found in batch_identifiers.
single-column {"m": 0}   -> 3 rows          (same file, no gate)
multi-column  {"code": ""}  -> ValueError   / single-column -> 2 rows
multi-column  {"flag": False} -> ValueError / single-column -> 2 rows

The Spark partitioner produces the same three ValueErrors on the same data.

The SQLAlchemy partitioner, run against a real (in-memory sqlite) table with the same identifiers, has no such gate:

sqlalchemy identifiers={'y': 2020, 'm': 1} -> rows=1 | WHERE y = ? AND m = ?
sqlalchemy identifiers={'y': 2020, 'm': 0} -> rows=2 | WHERE y = ? AND m = ?
sqlalchemy identifiers={'m': 0}            -> rows=4 | WHERE m = ?
sqlalchemy identifiers={'y': 0}            -> rows=1 | WHERE y = ?

Expected behaviour

A column name that has an entry in batch_identifiers partitions on that entry's value, whatever it is; only a column name with no entry should raise. Zero-indexed and empty-string partitions are ordinary data — a boolean flag, a code = "" "unknown" bucket, a counter whose first bucket is 0 — and a partitioner that names a column present in the identifiers should not call it absent.

Suggested fix

Test for the key instead of the value (if column_name not in batch_identifiers:) in both engines, which keeps the guard's stated purpose — a column name with no matching identifier — and removes the engine divergence.