diff options
author | Davies Liu <davies.liu@gmail.com> | 2014-09-18 18:11:48 -0700 |
---|---|---|
committer | Josh Rosen <joshrosen@apache.org> | 2014-09-18 18:11:48 -0700 |
commit | e77fa81a61798c89d5a9b6c9dc067d11785254b7 (patch) | |
tree | 2d84f29922e4523f223baff1c84573754c1cf0c7 /python/pyspark/sql.py | |
parent | 9306297d1d888d0430f79b2133ee7377871a3a18 (diff) | |
download | spark-e77fa81a61798c89d5a9b6c9dc067d11785254b7.tar.gz spark-e77fa81a61798c89d5a9b6c9dc067d11785254b7.tar.bz2 spark-e77fa81a61798c89d5a9b6c9dc067d11785254b7.zip |
[SPARK-3554] [PySpark] use broadcast automatically for large closure
Py4j can not handle large string efficiently, so we should use broadcast for large closure automatically. (Broadcast use local filesystem to pass through data).
Author: Davies Liu <davies.liu@gmail.com>
Closes #2417 from davies/command and squashes the following commits:
fbf4e97 [Davies Liu] bugfix
aefd508 [Davies Liu] use broadcast automatically for large closure
Diffstat (limited to 'python/pyspark/sql.py')
-rw-r--r-- | python/pyspark/sql.py | 8 |
1 files changed, 6 insertions, 2 deletions
diff --git a/python/pyspark/sql.py b/python/pyspark/sql.py index 8f6dbab240..42a9920f10 100644 --- a/python/pyspark/sql.py +++ b/python/pyspark/sql.py @@ -27,7 +27,7 @@ import warnings from array import array from operator import itemgetter -from pyspark.rdd import RDD, PipelinedRDD +from pyspark.rdd import RDD from pyspark.serializers import BatchedSerializer, PickleSerializer, CloudPickleSerializer from pyspark.storagelevel import StorageLevel from pyspark.traceback_utils import SCCallSiteSync @@ -975,7 +975,11 @@ class SQLContext(object): command = (func, BatchedSerializer(PickleSerializer(), 1024), BatchedSerializer(PickleSerializer(), 1024)) - pickled_command = CloudPickleSerializer().dumps(command) + ser = CloudPickleSerializer() + pickled_command = ser.dumps(command) + if pickled_command > (1 << 20): # 1M + broadcast = self._sc.broadcast(pickled_command) + pickled_command = ser.dumps(broadcast) broadcast_vars = ListConverter().convert( [x._jbroadcast for x in self._sc._pickled_broadcast_vars], self._sc._gateway._gateway_client) |