I have the same performance issue and dig a bit... ,created a numpy_srv(), similar with numpy_msg()

But it will only improve performance of byte[], not uint8[]. By using this, to transfer 100MBytes bytes[]:

• numpy_srv() with bytes[] uses: 1.2s
• uint8[] uses: 2.5s

I think uint8[] is already very fast for most uses.

import sys

import numpy as np
import rospy
from ax_msgs.srv import GetMapTest, GetMapTestRequest, GetMapTestResponse
from rospy.numpy_msg import numpy_msg

SIZE = 1024 * 1024 * 100

_numpy_srv_types = {}


def numpy_srv(srv_type):
    if srv_type in _numpy_srv_types:
        return _numpy_srv_types[srv_type]

    classdict = {
        "_type": srv_type._type,
        "_md5sum": srv_type._md5sum,
        "_request_class": numpy_msg(srv_type._request_class),
        "_response_class": numpy_msg(srv_type._response_class),
    }

    # create the numpy message type
    srv_type_name = "Numpy_%s" % srv_type._type.replace("/", "__")
    numpy_type = type(srv_type_name, (srv_type,), classdict)
    _numpy_srv_types[srv_type] = numpy_type
    return numpy_type


GetMapTestNumpy = numpy_srv(GetMapTest)  # numpy-lize service


if sys.argv[1] == "server":
    # server

    # create response content
    map_bytes = np.random.randint(0, 128, SIZE, dtype=np.uint8)

    def get_map_bytes_service(req: GetMapTestRequest):
        res = numpy_msg(GetMapTestResponse)()  # Must use numpy_msg() to create the response

        res.map_bytes = map_bytes  # map_bytes is `byte[]`
        return res

    rospy.init_node("my_node_server")
    rospy.Service("/get_map_bytes", GetMapTestNumpy, get_map_bytes_service)
    rospy.spin()
else:
    # client
    rospy.init_node("my_node_client")
    func = rospy.ServiceProxy("/get_map_bytes", GetMapTestNumpy)
    res: GetMapTestResponse = func(GetMapTestRequest())
    print(f"len(res.map_bytes) = {len(res.map_bytes)}")