Fix dask bugs (#507)

* Fix dask bugs

* update

fugue_dask/_utils.py

@@ -41,6 +41,7 @@ def hash_repartition(df: dd.DataFrame, num: int, cols: List[Any]) -> dd.DataFram
         return df
     if num == 1:
         return df.repartition(1)
+    df = df.reset_index(drop=True).clear_divisions()
     idf, ct = _add_hash_index(df, num, cols)
     return _postprocess(idf, ct, num)
 
@@ -63,9 +64,10 @@ def even_repartition(df: dd.DataFrame, num: int, cols: List[Any]) -> dd.DataFram
     """
     if num == 1:
         return df.repartition(1)
+    if len(cols) == 0 and num <= 0:
+        return df
+    df = df.reset_index(drop=True).clear_divisions()
     if len(cols) == 0:
-        if num <= 0:
-            return df
         idf, ct = _add_continuous_index(df)
     else:
         idf, ct = _add_group_index(df, cols, shuffle=False)
@@ -97,6 +99,7 @@ def rand_repartition(
         return df
     if num == 1:
         return df.repartition(1)
+    df = df.reset_index(drop=True).clear_divisions()
     if len(cols) == 0:
         idf, ct = _add_random_index(df, num=num, seed=seed)
     else:

fugue_ray/_utils/io.py

@@ -100,7 +100,7 @@ def _save_csv(
         if "header" in kw:
             kw["include_header"] = kw.pop("header")
 
-        def _fn() -> Dict[str, Any]:
+        def _fn() -> Dict[str, Any]:  # pragma: no cover
             return dict(write_options=pacsv.WriteOptions(**kw))
 
         df.native.write_csv(

tests/fugue_dask/test_execution_engine.py

@@ -2,8 +2,8 @@
 from threading import RLock
 from typing import Any, List, Optional
 
-import dask
 import dask.dataframe as dd
+import numpy as np
 import pandas as pd
 import pytest
 from dask.distributed import Client
@@ -25,7 +25,6 @@
 from fugue_test.builtin_suite import BuiltInTests
 from fugue_test.execution_suite import ExecutionEngineTests
 
-
 _CONF = {
     "fugue.rpc.server": "fugue.rpc.flask.FlaskRPCServer",
     "fugue.rpc.flask_server.host": "127.0.0.1",
@@ -321,6 +320,48 @@ def tr(df: List[List[Any]], add: Optional[callable]) -> List[List[Any]]:
     assert 5 == cb.n
 
 
+def test_multiple_transforms(fugue_dask_client):
+    def t1(df: pd.DataFrame) -> pd.DataFrame:
+        return pd.concat([df, df])
+
+    def t2(df: pd.DataFrame) -> pd.DataFrame:
+        return (
+            df.groupby(["a", "b"], as_index=False, dropna=False)
+            .apply(lambda x: x.head(1))
+            .reset_index(drop=True)
+        )
+
+    def compute(df: pd.DataFrame, engine) -> pd.DataFrame:
+        with fa.engine_context(engine):
+            ddf = fa.as_fugue_df(df)
+            ddf1 = fa.transform(ddf, t1, schema="*", partition=dict(algo="hash"))
+            ddf2 = fa.transform(
+                ddf1,
+                t2,
+                schema="*",
+                partition=dict(by=["a", "b"], presort="c", algo="coarse", num=2),
+            )
+            return (
+                ddf2.as_pandas()
+                .astype("float64")
+                .fillna(float("nan"))
+                .sort_values(["a", "b"])
+            )
+
+    np.random.seed(0)
+    df = pd.DataFrame(
+        dict(
+            a=np.random.randint(1, 5, 1000),
+            b=np.random.choice([1, 2, 3, None], 1000),
+            c=np.random.rand(1000),
+        )
+    )
+
+    actual = compute(df, fugue_dask_client)
+    expected = compute(df, None)
+    assert np.allclose(actual, expected, equal_nan=True)
+
+
 @transformer("ct:long")
 def count_partition(df: List[List[Any]]) -> List[List[Any]]:
     return [[len(df)]]