make sure that multiple key join is done on physical typed keys

polars/polars-core/src/frame/hash_join/mod.rs

@@ -26,6 +26,7 @@ use unsafe_unwrap::UnsafeUnwrap;
 #[cfg(feature = "private")]
 pub use self::multiple_keys::private_left_join_multiple_keys;
 use crate::frame::groupby::hashing::HASHMAP_INIT_SIZE;
+use crate::utils::series::to_physical;
 
 /// If Categorical types are created without a global string cache or under
 /// a different global string cache the mapping will be incorrect.
@@ -1113,12 +1114,16 @@ impl DataFrame {
                 self.height()
             }
         }
+        // make sure that we don't have logical types.
+        // we don't overwrite the original selected as that might be used to create a column in the new df
+        let selected_left_physical = to_physical(&selected_left);
+        let selected_right_physical = to_physical(&selected_right);
 
         // multiple keys
         match how {
             JoinType::Inner => {
-                let left = DataFrame::new_no_checks(selected_left);
-                let right = DataFrame::new_no_checks(selected_right.clone());
+                let left = DataFrame::new_no_checks(selected_left_physical);
+                let right = DataFrame::new_no_checks(selected_right_physical);
                 let (left, right, swap) = det_hash_prone_order!(left, right);
                 let join_tuples = inner_join_multiple_keys(&left, &right, swap);
 
@@ -1134,8 +1139,8 @@ impl DataFrame {
                 self.finish_join(df_left, df_right, suffix)
             }
             JoinType::Left => {
-                let left = DataFrame::new_no_checks(selected_left);
-                let right = DataFrame::new_no_checks(selected_right.clone());
+                let left = DataFrame::new_no_checks(selected_left_physical);
+                let right = DataFrame::new_no_checks(selected_right_physical);
                 let join_tuples = left_join_multiple_keys(&left, &right);
 
                 let (df_left, df_right) = POOL.join(
@@ -1152,8 +1157,8 @@ impl DataFrame {
                 self.finish_join(df_left, df_right, suffix)
             }
             JoinType::Outer => {
-                let left = DataFrame::new_no_checks(selected_left.clone());
-                let right = DataFrame::new_no_checks(selected_right.clone());
+                let left = DataFrame::new_no_checks(selected_left_physical);
+                let right = DataFrame::new_no_checks(selected_right_physical);
 
                 let (left, right, swap) = det_hash_prone_order!(left, right);
                 let opt_join_tuples = outer_join_multiple_keys(&left, &right, swap);

polars/polars-core/src/utils.rs → polars/polars-core/src/utils/mod.rs

@@ -1,3 +1,5 @@
+pub(crate) mod series;
+
 use crate::prelude::*;
 use crate::POOL;
 pub use arrow;

polars/polars-core/src/utils/series.rs

@@ -0,0 +1,7 @@
+use crate::prelude::*;
+
+pub(crate) fn to_physical(s: &[Series]) -> Vec<Series> {
+    s.iter()
+        .map(|s| s.to_physical_repr().into_owned())
+        .collect()
+}

py-polars/tests/test_df.py

@@ -478,45 +478,27 @@ def test_join():
 
 def test_joins_dispatch():
     # this just flexes the dispatch a bit
-    dfa = pl.DataFrame(
-        {
-            "a": ["a", "b", "c"],
-            "b": [1, 2, 3],
-            "date": ["2021-01-01", "2021-01-02", "2021-01-03"],
-            "datetime": [13241324, 12341256, 12341234],
-        }
-    ).with_columns(
-        [pl.col("date").str.strptime(pl.Date), pl.col("datetime").cast(pl.Datetime)]
-    )
-    dfa = pl.concat([dfa, dfa], rechunk=False)
 
-    dfb = pl.DataFrame(
+    # don't change the data of this dataframe, this triggered:
+    # https://github.com/pola-rs/polars/issues/1688
+    dfa = pl.DataFrame(
         {
-            "a": ["a", "b", "c"],
-            "b": [1, 2, 3],
-            "date": ["2021-01-01", "2021-01-02", "2021-01-03"],
-            "datetime": [13241324, 12341256, 12341234],
+            "a": ["a", "b", "c", "a"],
+            "b": [1, 2, 3, 1],
+            "date": ["2021-01-01", "2021-01-02", "2021-01-03", "2021-01-01"],
+            "datetime": [13241324, 12341256, 12341234, 13241324],
         }
     ).with_columns(
         [pl.col("date").str.strptime(pl.Date), pl.col("datetime").cast(pl.Datetime)]
     )
 
-    dfa.join(dfb, on=["a", "b", "date", "datetime"], how="left")
-    dfa.join(dfb, on=["a", "b", "date", "datetime"], how="inner")
-    dfa.join(dfb, on=["a", "b", "date", "datetime"], how="outer")
-    dfa.join(dfb, on=["date", "datetime"], how="left")
-    dfa.join(dfb, on=["date", "datetime"], how="inner")
-    dfa.join(dfb, on=["date", "datetime"], how="outer")
-    dfa.join(dfb, on=["date", "datetime", "a"], how="left")
-    dfa.join(dfb, on=["date", "datetime", "a"], how="inner")
-    dfa.join(dfb, on=["date", "a"], how="outer")
-    dfa.join(dfb, on=["date", "a"], how="left")
-    dfa.join(dfb, on=["date", "a"], how="inner")
-    dfa.join(dfb, on=["date", "a"], how="outer")
-    dfa.join(dfb, on=["date"], how="outer")
-    dfa.join(dfb, on=["date"], how="left")
-    dfa.join(dfb, on=["date"], how="inner")
-    dfa.join(dfb, on=["date"], how="outer")
+    for how in ["left", "inner", "outer"]:
+        dfa.join(dfa, on=["a", "b", "date", "datetime"], how=how)
+        dfa.join(dfa, on=["date", "datetime"], how=how)
+        dfa.join(dfa, on=["date", "datetime", "a"], how=how)
+        dfa.join(dfa, on=["date", "a"], how=how)
+        dfa.join(dfa, on=["a", "datetime"], how=how)
+        dfa.join(dfa, on=["date"], how=how)
 
 
 def test_hstack():