[Ray] Basic slow subtask detection (mars-project#3305)

* Basic slow subtask detection

* Replace sklearn.utils._testing.assert_warns with pytest.warns

* Fix load_boston

* Fix rerun_time type

* Remove rerun_time from Task and Subtask

* Fix

* Fix

* Add UT

* Fix lint

* Fix lint

* Refine logs

* Fix

* Refine code

* Add check_slow_subtasks_iqr_ratio config

Co-authored-by: 刘宝 <po.lb@antgroup.com>

mars/contrib/dask/tests/test_dask.py

@@ -75,15 +75,15 @@ def test_unpartitioned_dataframe(setup_cluster):
     from dask import dataframe as dd
     from pandas._testing import assert_frame_equal
     import pandas as pd
-    from sklearn.datasets import load_boston
+    from sklearn.datasets import load_iris
 
-    boston = load_boston()
+    boston = load_iris()
     pd.DataFrame(boston.data, columns=boston["feature_names"]).to_csv(
         "./boston_housing_data.csv"
     )
 
     df = dd.read_csv(r"./boston_housing_data.csv")
-    df["CRIM"] = df["CRIM"] / 2
+    df["sepal length (cm)"] = df["sepal length (cm)"] / 2
 
     dask_res = df.compute()
     assert_frame_equal(dask_res, df.compute(scheduler=mars_scheduler))

mars/learn/cluster/tests/test_k_means.py

@@ -22,7 +22,7 @@
 try:
     from sklearn.datasets import make_blobs
     from sklearn.metrics.cluster import v_measure_score
-    from sklearn.utils._testing import assert_raise_message, assert_warns
+    from sklearn.utils._testing import assert_raise_message
 except ImportError:
     pass
 
@@ -485,14 +485,13 @@ def test_k_means_function(setup):
     assert inertia > 0.0
 
     # check warning when centers are passed
-    assert_warns(
-        RuntimeWarning,
-        k_means,
-        X,
-        n_clusters=n_clusters,
-        sample_weight=None,
-        init=centers,
-    )
+    with pytest.warns(RuntimeWarning):
+        k_means(
+            X,
+            n_clusters=n_clusters,
+            sample_weight=None,
+            init=centers,
+        )
 
     # to many clusters desired
     with pytest.raises(ValueError):

mars/learn/metrics/pairwise/tests/test_pariwise_distances.py

@@ -17,7 +17,7 @@
 from sklearn.metrics import pairwise_distances as sk_pairwise_distances
 from sklearn.neighbors import NearestNeighbors as SkNearestNeighbors
 from sklearn.exceptions import DataConversionWarning
-from sklearn.utils._testing import assert_warns
+
 
 from ..... import tensor as mt
 from .....session import execute, fetch
@@ -63,7 +63,8 @@ def test_pairwise_distances_execution(setup):
     expected = sk_pairwise_distances(raw_x, raw_y, metric=m)
     np.testing.assert_almost_equal(result, expected)
 
-    assert_warns(DataConversionWarning, pairwise_distances, x, y, metric="jaccard")
+    with pytest.warns(DataConversionWarning):
+        pairwise_distances(x, y, metric="jaccard")
 
     with pytest.raises(ValueError):
         _ = pairwise_distances(x, y, metric="unknown")

mars/learn/metrics/tests/test_ranking.py

@@ -27,7 +27,6 @@
 from sklearn.exceptions import UndefinedMetricWarning
 from sklearn.utils import check_random_state
 from sklearn.utils._testing import (
-    assert_warns,
     assert_almost_equal,
     assert_array_almost_equal,
 )
@@ -166,14 +165,16 @@ def test_roc_curve_one_label(setup):
     y_pred = [0, 1, 0, 1, 0, 1, 0, 1, 0, 1]
     # assert there are warnings
     w = UndefinedMetricWarning
-    fpr, tpr, thresholds = assert_warns(w, roc_curve, y_true, y_pred)
+    with pytest.warns(w):
+        fpr, tpr, thresholds = roc_curve(y_true, y_pred)
     # all true labels, all fpr should be nan
     np.testing.assert_array_equal(fpr.fetch(), np.full(len(thresholds), np.nan))
     assert fpr.shape == tpr.shape
     assert fpr.shape == thresholds.shape
 
     # assert there are warnings
-    fpr, tpr, thresholds = assert_warns(w, roc_curve, [1 - x for x in y_true], y_pred)
+    with pytest.warns(w):
+        fpr, tpr, thresholds = roc_curve([1 - x for x in y_true], y_pred)
     # all negative labels, all tpr should be nan
     np.testing.assert_array_equal(tpr.fetch(), np.full(len(thresholds), np.nan))
     assert fpr.shape == tpr.shape

mars/learn/neighbors/tests/test_nearest_neighbors.py

@@ -24,7 +24,6 @@
     from sklearn.neighbors import NearestNeighbors as SkNearestNeighbors
     from sklearn.neighbors import BallTree as SkBallTree
     from sklearn.neighbors import KDTree as SkKDTree
-    from sklearn.utils._testing import assert_warns
 except ImportError:  # pragma: no cover
     SkNearestNeighbors = None
 
@@ -66,7 +65,8 @@ def test_nearest_neighbors(setup):
     with pytest.raises(ValueError):
         _ = NearestNeighbors(algorithm="auto", metric="unknown")
 
-    assert_warns(SyntaxWarning, NearestNeighbors, metric_params={"p": 1})
+    with pytest.warns(SyntaxWarning):
+        NearestNeighbors(metric_params={"p": 1})
 
     with pytest.raises(ValueError):
         _ = NearestNeighbors(metric="wminkowski", p=0)
@@ -105,7 +105,8 @@ def test_nearest_neighbors(setup):
         nn.fit(np.random.rand(0, 10))
 
     nn = NearestNeighbors(algorithm="ball_tree")
-    assert_warns(UserWarning, nn.fit, X_sparse)
+    with pytest.warns(UserWarning):
+        nn.fit(X_sparse)
 
     nn = NearestNeighbors(metric="haversine")
     with pytest.raises(ValueError):

mars/learn/semi_supervised/tests/test_label_propagation.py

@@ -17,7 +17,7 @@
 from sklearn.datasets import make_classification
 from sklearn.model_selection import train_test_split
 from sklearn.exceptions import ConvergenceWarning
-from sklearn.utils._testing import assert_no_warnings, assert_warns
+from sklearn.utils._testing import assert_no_warnings
 
 from .... import tensor as mt
 from ...metrics.pairwise import rbf_kernel
@@ -103,7 +103,8 @@ def test_convergence_warning(setup):
     y = np.array([0, 1, -1])
 
     mdl = LabelPropagation(kernel="rbf", max_iter=1)
-    assert_warns(ConvergenceWarning, mdl.fit, X, y)
+    with pytest.warns(ConvergenceWarning):
+        mdl.fit(X, y)
     assert mdl.n_iter_ == mdl.max_iter
 
     mdl = LabelPropagation(kernel="rbf", max_iter=500)

mars/services/subtask/core.py

@@ -52,7 +52,7 @@ def is_done(self) -> bool:
 
 
 class Subtask(Serializable):
-    __slots__ = ("_repr", "_pure_depend_keys")
+    __slots__ = ("_repr", "_pure_depend_keys", "runtime")
 
     subtask_id: str = StringField("subtask_id")
     subtask_name: str = StringField("subtask_name")
@@ -65,7 +65,6 @@ class Subtask(Serializable):
     virtual: bool = BoolField("virtual")
     retryable: bool = BoolField("retryable")
     priority: Tuple[int, int] = TupleField("priority", FieldTypes.int32)
-    rerun_time: int = Int32Field("rerun_time")
     extra_config: dict = DictField("extra_config")
     stage_id: str = StringField("stage_id")
     # chunks that need meta updated
@@ -95,7 +94,6 @@ def __init__(
         priority: Tuple[int, int] = None,
         virtual: bool = False,
         retryable: bool = True,
-        rerun_time: int = 0,
         extra_config: dict = None,
         stage_id: str = None,
         update_meta_chunks: List[ChunkType] = None,
@@ -116,7 +114,6 @@ def __init__(
             priority=priority,
             virtual=virtual,
             retryable=retryable,
-            rerun_time=rerun_time,
             extra_config=extra_config,
             stage_id=stage_id,
             update_meta_chunks=update_meta_chunks,
@@ -129,11 +126,13 @@ def __init__(
         )
         self._pure_depend_keys = None
         self._repr = None
+        self.runtime = None
 
     def __on_deserialize__(self):
         super(Subtask, self).__on_deserialize__()
         self._pure_depend_keys = None
         self._repr = None
+        self.runtime = None
 
     @property
     def expect_band(self):

mars/services/task/core.py

@@ -22,7 +22,6 @@
     Serializable,
     StringField,
     ReferenceField,
-    Int32Field,
     BoolField,
     AnyField,
     DictField,
@@ -41,7 +40,6 @@ class Task(Serializable):
     session_id: str = StringField("session_id")
     tileable_graph: TileableGraph = ReferenceField("tileable_graph", TileableGraph)
     fuse_enabled: bool = BoolField("fuse_enabled")
-    rerun_time: int = Int32Field("rerun_time")
     extra_config: dict = DictField("extra_config")
 
     def __init__(
@@ -50,15 +48,13 @@ def __init__(
         session_id: str = None,
         tileable_graph: TileableGraph = None,
         fuse_enabled: bool = True,
-        rerun_time: int = 0,
         extra_config: dict = None,
     ):
         super().__init__(
             task_id=task_id,
             session_id=session_id,
             tileable_graph=tileable_graph,
             fuse_enabled=fuse_enabled,
-            rerun_time=rerun_time,
             extra_config=extra_config,
         )
 

mars/services/task/execution/ray/config.py

@@ -28,6 +28,7 @@
 # The default interval seconds to update progress and collect garbage.
 DEFAULT_MONITOR_INTERVAL_SECONDS = 0 if IN_RAY_CI else 1
 DEFAULT_LOG_INTERVAL_SECONDS = 60
+DEFAULT_CHECK_SLOW_SUBTASKS_INTERVAL_SECONDS = 120
 
 
 @register_config_cls
@@ -77,6 +78,19 @@ def get_log_interval_seconds(self):
             "log_interval_seconds", DEFAULT_LOG_INTERVAL_SECONDS
         )
 
+    def get_check_slow_subtasks_interval_seconds(self) -> float:
+        return self._ray_execution_config.get(
+            "check_slow_subtasks_interval_seconds",
+            DEFAULT_CHECK_SLOW_SUBTASKS_INTERVAL_SECONDS,
+        )
+
+    def get_check_slow_subtask_iqr_ratio(self) -> float:
+        # https://en.wikipedia.org/wiki/Box_plot
+        # iqr = q3 - q1
+        # duration_threshold = q3 + check_slow_subtasks_iqr_ratio * (q3 - q1)
+        # So, the value == 3, extremely slow(probably hang); value == 1.5, slow
+        return self._ray_execution_config.get("check_slow_subtasks_iqr_ratio", 3)
+
     def get_shuffle_fetch_type(self) -> ShuffleFetchType:
         return ShuffleFetchType.FETCH_BY_INDEX