refactor(datafusion): avoid reinitializing memtables on every execute…

… call (#10057)

ibis/backends/datafusion/__init__.py

@@ -9,7 +9,6 @@
 
 import datafusion as df
 import pyarrow as pa
-import pyarrow.dataset as ds
 import pyarrow_hotfix  # noqa: F401
 import sqlglot as sg
 import sqlglot.expressions as sge
@@ -418,14 +417,16 @@ def _register_failure(self):
 
     def _register_in_memory_table(self, op: ops.InMemoryTable) -> None:
         name = op.name
-        schema = op.schema
 
-        self.con.deregister_table(name)
-        if batches := op.data.to_pyarrow(schema).to_batches():
-            self.con.register_record_batches(name, [batches])
-        else:
-            empty_dataset = ds.dataset([], schema=schema.to_pyarrow())
-            self.con.register_dataset(name=name, dataset=empty_dataset)
+        db = self.con.catalog().database()
+
+        try:
+            db.table(name)
+        except Exception:  # noqa: BLE001 because datafusion doesn't have anything better
+            # self.con.register_table is broken, so we do this roundabout thing
+            # of constructing a datafusion DataFrame, which has a side effect
+            # of registering the table
+            self.con.from_arrow_table(op.data.to_pyarrow(op.schema), name)
 
     def read_csv(
         self, path: str | Path, table_name: str | None = None, **kwargs: Any