Pin xgboost_ray<0.1.14 (#3328)

* Pin xgboost_ray<0.1.14

* DataFrameReadSQL compatible with SQLAlchemy > 2

---------

Co-authored-by: 刘宝 <[email protected]>

Author: fyrestone

.github/workflows/platform-ci.yml

@@ -104,7 +104,7 @@ jobs:
               rm -fr /tmp/etcd-$ETCD_VER-linux-amd64.tar.gz /tmp/etcd-download-test
             fi
             if [ -n "$WITH_RAY" ] || [ -n "$WITH_RAY_DAG" ] || [ -n "$WITH_RAY_DEPLOY" ]; then
-              pip install "xgboost_ray" "protobuf<4"
+              pip install "xgboost_ray<0.1.14" "protobuf<4"
               # Ray Datasets need pyarrow>=6.0.1
               pip install "pyarrow>=6.0.1"
             fi

mars/dataframe/datasource/read_sql.py

@@ -355,18 +355,16 @@ def _tile_partition(cls, op: "DataFrameReadSQL"):
             from sqlalchemy import sql
 
             engine = sa.create_engine(op.con, **(op.engine_kwargs or dict()))
-            try:
+            with engine.connect() as connection:
                 part_col = selectable.columns[op.partition_col]
-                range_results = engine.execute(
+                range_results = connection.execute(
                     sql.select(sql.func.min(part_col), sql.func.max(part_col))
                 )
 
                 op.low_limit, op.high_limit = next(range_results)
                 if op.parse_dates and op.partition_col in op.parse_dates:
                     op.low_limit = op._parse_datetime(op.low_limit)
                     op.high_limit = op._parse_datetime(op.high_limit)
-            finally:
-                engine.dispose()
 
         if isinstance(op.low_limit, (datetime.datetime, np.datetime64, pd.Timestamp)):
             seps = pd.date_range(op.low_limit, op.high_limit, op.num_partitions + 1)