def put_all(self, items, output=None, options: dict = None):
if options is None:
options = {}
include_key = options.get("include_key", True)
job_id = generate_job_id(self.__session_id, RollPair.PUT_ALL)
# TODO:1: consider multiprocessing scenario. parallel size should be sent to egg_pair to set write signal count
def send_command():
job = ErJob(id=job_id,
name=RollPair.PUT_ALL,
inputs=[self.__store],
outputs=[self.__store],
functors=[])
result = self.__command_client.simple_sync_send(
input=job,
output_type=ErJob,
endpoint=self.ctx.get_roll()._command_endpoint,
command_uri=CommandURI(
f'{RollPair.ROLL_PAIR_URI_PREFIX}/{RollPair.RUN_JOB}'),
serdes_type=SerdesTypes.PROTOBUF)
return result
th = Thread(target=send_command,
name=f'roll_pair-send_command-{job_id}')
th.start()
populated_store = self.ctx.populate_processor(self.__store)
shuffler = TransferPair(job_id)
broker = FifoBroker()
bb = BatchBroker(broker)
scatter_future = shuffler.scatter(broker, self.partitioner,
populated_store)
key_serdes = self.key_serdes
value_serdes = self.value_serdes
try:
if include_key:
for k, v in items:
bb.put(item=(key_serdes.serialize(k),
value_serdes.serialize(v)))
else:
k = 0
for v in items:
bb.put(item=(key_serdes.serialize(k),
value_serdes.serialize(v)))
k += 1
finally:
bb.signal_write_finish()
scatter_results = scatter_future.result()
L.debug(f"scatter_results: {scatter_results}")
th.join()
return RollPair(populated_store, self.ctx)
def put_all(self, items, output=None, options: dict = None):
if options is None:
options = {}
include_key = options.get("include_key", True)
job_id = generate_job_id(self.__session_id, RollPair.PUT_ALL)
# TODO:1: consider multiprocessing scenario. parallel size should be sent to egg_pair to set write signal count
def send_command():
job = ErJob(id=job_id,
name=RollPair.PUT_ALL,
inputs=[self.__store],
outputs=[self.__store],
functors=[])
task_results = self._run_job(job)
return self.__get_output_from_result(task_results)
th = Thread(target=send_command,
name=f'roll_pair-send_command-{job_id}')
th.start()
populated_store = self.ctx.populate_processor(self.__store)
shuffler = TransferPair(job_id)
fifo_broker = FifoBroker()
bb = BatchBroker(fifo_broker)
scatter_future = shuffler.scatter(fifo_broker, self.partitioner,
populated_store)
key_serdes = self.key_serdes
value_serdes = self.value_serdes
try:
if include_key:
for k, v in items:
bb.put(item=(key_serdes.serialize(k),
value_serdes.serialize(v)))
else:
k = 0
for v in items:
bb.put(item=(key_serdes.serialize(k),
value_serdes.serialize(v)))
k += 1
finally:
bb.signal_write_finish()
scatter_results = scatter_future.result()
th.join()
return RollPair(populated_store, self.ctx)
def _run_unary(self, func, task, shuffle=False):
key_serdes = create_serdes(task._inputs[0]._store_locator._serdes)
value_serdes = create_serdes(task._inputs[0]._store_locator._serdes)
with create_adapter(task._inputs[0]) as input_db:
L.debug(f"create_store_adatper: {task._inputs[0]}")
with input_db.iteritems() as rb:
L.debug(f"create_store_adatper_iter: {task._inputs[0]}")
from eggroll.roll_pair.transfer_pair import TransferPair, BatchBroker
if shuffle:
total_partitions = task._inputs[
0]._store_locator._total_partitions
output_store = task._job._outputs[0]
shuffle_broker = FifoBroker()
write_bb = BatchBroker(shuffle_broker)
try:
shuffler = TransferPair(transfer_id=task._job._id)
store_future = shuffler.store_broker(
task._outputs[0], True, total_partitions)
scatter_future = shuffler.scatter(
shuffle_broker,
partitioner(hash_func=hash_code,
total_partitions=total_partitions),
output_store)
func(rb, key_serdes, value_serdes, write_bb)
finally:
write_bb.signal_write_finish()
scatter_results = scatter_future.result()
store_result = store_future.result()
L.debug(f"scatter_result:{scatter_results}")
L.debug(f"gather_result:{store_result}")
else:
# TODO: modification may be needed when store options finished
with create_adapter(task._outputs[0],
options=task._job._options
) as db, db.new_batch() as wb:
func(rb, key_serdes, value_serdes, wb)
L.debug(f"close_store_adatper:{task._inputs[0]}")
def _run_unary(self, func, task, shuffle=False, reduce_op=None):
input_store_head = task._job._inputs[0]
output_store_head = task._job._outputs[0]
input_key_serdes = create_serdes(
input_store_head._store_locator._serdes)
input_value_serdes = create_serdes(
input_store_head._store_locator._serdes)
output_key_serdes = create_serdes(
output_store_head._store_locator._serdes)
output_value_serdes = create_serdes(
output_store_head._store_locator._serdes)
if input_key_serdes != output_key_serdes or \
input_value_serdes != output_value_serdes:
raise ValueError(
f"input key-value serdes:{(input_key_serdes, input_value_serdes)}"
f"differ from output key-value serdes:{(output_key_serdes, output_value_serdes)}"
)
if shuffle:
from eggroll.roll_pair.transfer_pair import TransferPair
input_total_partitions = input_store_head._store_locator._total_partitions
output_total_partitions = output_store_head._store_locator._total_partitions
output_store = output_store_head
my_server_node_id = get_static_er_conf().get(
'server_node_id', None)
shuffler = TransferPair(transfer_id=task._job._id)
if not task._outputs or \
(my_server_node_id is not None
and my_server_node_id != task._outputs[0]._processor._server_node_id):
store_future = None
else:
store_future = shuffler.store_broker(
store_partition=task._outputs[0],
is_shuffle=True,
total_writers=input_total_partitions,
reduce_op=reduce_op)
if not task._inputs or \
(my_server_node_id is not None
and my_server_node_id != task._inputs[0]._processor._server_node_id):
scatter_future = None
else:
shuffle_broker = FifoBroker()
write_bb = BatchBroker(shuffle_broker)
try:
scatter_future = shuffler.scatter(
input_broker=shuffle_broker,
partition_function=partitioner(
hash_func=hash_code,
total_partitions=output_total_partitions),
output_store=output_store)
with create_adapter(task._inputs[0]) as input_db, \
input_db.iteritems() as rb:
func(rb, input_key_serdes, input_value_serdes,
write_bb)
finally:
write_bb.signal_write_finish()
if scatter_future:
scatter_results = scatter_future.result()
else:
scatter_results = 'no scatter for this partition'
if store_future:
store_results = store_future.result()
else:
store_results = 'no store for this partition'
else: # no shuffle
with create_adapter(task._inputs[0]) as input_db, \
input_db.iteritems() as rb, \
create_adapter(task._outputs[0], options=task._job._options) as db, \
db.new_batch() as wb:
func(rb, input_key_serdes, input_value_serdes, wb)
L.trace(f"close_store_adatper:{task._inputs[0]}")