fix: replace NaNs with None in some backends when loading from pandas…

… dataframe (#9094)

## Description of changes

Examples were broken on MySQL backend and Postgres backend when there are
null values in a numeric column.

Druid, PySpark, RW don't support examples.

- Exasol - did not test
- Flink - broken
- Impala - did not test
- MSSQL - broken #9095 
- MySQL - fixed
- Oracle - did not test
- PostgreSQL - fixed
- Snowflake - did not test

## Issues closed

#8792

---------

Co-authored-by: Chloe He <[email protected]>

ibis/backends/mysql/__init__.py

@@ -11,6 +11,7 @@
 from typing import TYPE_CHECKING, Any
 from urllib.parse import parse_qs, urlparse
 
+import numpy as np
 import pymysql
 import sqlglot as sg
 import sqlglot.expressions as sge
@@ -481,6 +482,9 @@ def _register_in_memory_table(self, op: ops.InMemoryTable) -> None:
 
             columns = schema.keys()
             df = op.data.to_frame()
+            # nan can not be used with MySQL
+            df = df.replace(np.nan, None)
+
             data = df.itertuples(index=False)
             cols = ", ".join(
                 ident.sql(self.name)

ibis/backends/postgres/__init__.py

@@ -11,8 +11,11 @@
 from typing import TYPE_CHECKING, Any, Callable
 from urllib.parse import parse_qs, urlparse
 
+import numpy as np
+import pandas as pd
 import sqlglot as sg
 import sqlglot.expressions as sge
+from pandas.api.types import is_float_dtype
 
 import ibis
 import ibis.common.exceptions as com
@@ -144,6 +147,16 @@ def _register_in_memory_table(self, op: ops.InMemoryTable) -> None:
 
             columns = schema.keys()
             df = op.data.to_frame()
+            # nan gets compiled into 'NaN'::float which throws errors in non-float columns
+            # In order to hold NaN values, pandas automatically converts integer columns
+            # to float columns if there are NaN values in them. Therefore, we need to convert
+            # them to their original dtypes (that support pd.NA) to figure out which columns
+            # are actually non-float, then fill the NaN values in those columns with None.
+            convert_df = df.convert_dtypes()
+            for col in convert_df.columns:
+                if not is_float_dtype(convert_df[col]):
+                    df[col] = df[col].replace(np.nan, None)
+
             data = df.itertuples(index=False)
             cols = ", ".join(
                 ident.sql(self.dialect)