Fix XGBoost when some workers do not have evals data (mars-project#2861)

Author: Xuye (Chris) Qin

mars/learn/contrib/xgboost/start_tracker.py

@@ -47,9 +47,9 @@ def execute(cls, ctx, op):
 
         env = {"DMLC_NUM_WORKER": op.n_workers}
         rabit_context = RabitTracker(
-            hostIP=ctx.get_local_host_ip(), nslave=op.n_workers
+            host_ip=ctx.get_local_host_ip(), n_workers=op.n_workers
         )
-        env.update(rabit_context.slave_envs())
+        env.update(rabit_context.worker_envs())
 
         rabit_context.start(op.n_workers)
         thread = Thread(target=rabit_context.join)

mars/learn/contrib/xgboost/tests/test_train.py

@@ -60,16 +60,17 @@ def test_local_train_dataframe(setup):
 
 
 @pytest.mark.skipif(xgboost is None, reason="XGBoost not installed")
-def test_train_evals(setup_cluster):
+@pytest.mark.parametrize("chunk_size", [n_rows // 5, n_rows])
+def test_train_evals(setup_cluster, chunk_size):
     rs = mt.random.RandomState(0)
     # keep 1 chunk for X and y
     X = rs.rand(n_rows, n_columns, chunk_size=(n_rows, n_columns // 2))
     y = rs.rand(n_rows, chunk_size=n_rows)
     base_margin = rs.rand(n_rows, chunk_size=n_rows)
     dtrain = MarsDMatrix(X, y, base_margin=base_margin)
     eval_x = MarsDMatrix(
-        rs.rand(n_rows, n_columns, chunk_size=n_rows // 5),
-        rs.rand(n_rows, chunk_size=n_rows // 5),
+        rs.rand(n_rows, n_columns, chunk_size=chunk_size),
+        rs.rand(n_rows, chunk_size=chunk_size),
     )
     evals = [(eval_x, "eval_x")]
     eval_result = dict()