def _push_rollpair(self,
rp: RollPair,
rs_header: ErRollSiteHeader,
options: dict = None):
if options is None:
options = {}
rs_key = rs_header.get_rs_key()
if L.isEnabledFor(logging.DEBUG):
L.debug(
f"pushing rollpair: rs_key={rs_key}, rs_header={rs_header}, rp.count={rp.count()}"
)
start_time = time.time()
rs_header._total_partitions = rp.get_partitions()
serdes = options.get('serdes', None)
if serdes is not None:
rs_header._options['serdes'] = serdes
wrapee_cls = options.get('wrapee_cls', None)
if serdes is not None:
rs_header._options['wrapee_cls'] = wrapee_cls
batches_per_stream = self.push_batches_per_stream
body_bytes = self.batch_body_bytes
endpoint = self.ctx.proxy_endpoint
max_retry_cnt = self.push_max_retry
long_retry_cnt = self.push_long_retry
per_stream_timeout = self.push_per_stream_timeout
def _push_partition(ertask):
rs_header._partition_id = ertask._inputs[0]._id
L.trace(
f"pushing rollpair partition. rs_key={rs_key}, partition_id={rs_header._partition_id}, rs_header={rs_header}"
)
from eggroll.core.grpc.factory import GrpcChannelFactory
from eggroll.core.proto import proxy_pb2_grpc
grpc_channel_factory = GrpcChannelFactory()
with create_adapter(ertask._inputs[0]) as db, db.iteritems() as rb:
# NOTICE AGAIN: all modifications to rs_header are limited in bs_helper.
# rs_header is shared by bs_helper and here. any modification in bs_helper affects this header.
# Remind that python's object references are passed by value,
# meaning the 'pointer' is copied, while the contents are modificable
bs_helper = _BatchStreamHelper(rs_header)
bin_batch_streams = bs_helper._generate_batch_streams(
pair_iter=rb,
batches_per_stream=batches_per_stream,
body_bytes=body_bytes)
channel = grpc_channel_factory.create_channel(endpoint)
stub = proxy_pb2_grpc.DataTransferServiceStub(channel)
for batch_stream in bin_batch_streams:
batch_stream_data = list(batch_stream)
cur_retry = 0
exception = None
while cur_retry < max_retry_cnt:
L.trace(
f'pushing rollpair partition stream. rs_key={rs_key}, partition_id={rs_header._partition_id}, rs_header={rs_header}, cur_retry={cur_retry}'
)
try:
stub.push(bs_helper.generate_packet(
batch_stream_data, cur_retry),
timeout=per_stream_timeout)
exception = None
break
except Exception as e:
if cur_retry < max_retry_cnt - long_retry_cnt:
retry_interval = round(
min(2 * cur_retry, 20) +
random.random() * 10, 3)
else:
retry_interval = round(
300 + random.random() * 10, 3)
L.warn(
f"push rp partition error. rs_key={rs_key}, partition_id={rs_header._partition_id}, rs_header={rs_header}, max_retry_cnt={max_retry_cnt}, cur_retry={cur_retry}, retry_interval={retry_interval}",
exc_info=e)
time.sleep(retry_interval)
if isinstance(e, RpcError) and e.code(
).name == 'UNAVAILABLE':
channel = grpc_channel_factory.create_channel(
endpoint, refresh=True)
stub = proxy_pb2_grpc.DataTransferServiceStub(
channel)
exception = e
finally:
cur_retry += 1
if exception is not None:
L.exception(
f"push partition failed. rs_key={rs_key}, partition_id={rs_header._partition_id}, rs_header={rs_header}, cur_retry={cur_retry}",
exc_info=exception)
raise exception
L.trace(
f'pushed rollpair partition stream. rs_key={rs_key}, partition_id={rs_header._partition_id}, rs_header={rs_header}, retry count={cur_retry - 1}'
)
L.trace(
f"pushed rollpair partition. rs_key={rs_key}, partition_id={rs_header._partition_id}, rs_header={rs_header}"
)
rp.with_stores(_push_partition, options={"__op": "push_partition"})
if L.isEnabledFor(logging.DEBUG):
L.debug(
f"pushed rollpair: rs_key={rs_key}, rs_header={rs_header}, count={rp.count()}, elapsed={time.time() - start_time}"
)
self.ctx.pushing_latch.count_down()
def _push_bytes(self,
obj,
rs_header: ErRollSiteHeader,
options: dict = None):
if options is None:
options = {}
start_time = time.time()
rs_key = rs_header.get_rs_key()
int_size = 4
if L.isEnabledFor(logging.DEBUG):
L.debug(f"pushing object: rs_key={rs_key}, rs_header={rs_header}")
def _generate_obj_bytes(py_obj, body_bytes):
key_id = 0
obj_bytes = pickle.dumps(py_obj)
obj_bytes_len = len(obj_bytes)
cur_pos = 0
while cur_pos <= obj_bytes_len:
yield key_id.to_bytes(
int_size, "big"), obj_bytes[cur_pos:cur_pos + body_bytes]
key_id += 1
cur_pos += body_bytes
rs_header._partition_id = 0
rs_header._total_partitions = 1
serdes = options.get('serdes', None)
if serdes is not None:
rs_header._options['serdes'] = serdes
wrapee_cls = options.get('wrapee_cls', None)
if serdes is not None:
rs_header._options['wrapee_cls'] = wrapee_cls
# NOTICE: all modifications to rs_header are limited in bs_helper.
# rs_header is shared by bs_helper and here. any modification in bs_helper affects this header.
# Remind that python's object references are passed by value,
# meaning the 'pointer' is copied, while the contents are modificable
bs_helper = _BatchStreamHelper(rs_header=rs_header)
bin_batch_streams = bs_helper._generate_batch_streams(
pair_iter=_generate_obj_bytes(obj, self.batch_body_bytes),
batches_per_stream=self.push_batches_per_stream,
body_bytes=self.batch_body_bytes)
grpc_channel_factory = GrpcChannelFactory()
channel = grpc_channel_factory.create_channel(self.ctx.proxy_endpoint)
stub = proxy_pb2_grpc.DataTransferServiceStub(channel)
max_retry_cnt = self.push_max_retry
long_retry_cnt = self.push_long_retry
per_stream_timeout = self.push_per_stream_timeout
# if use stub.push.future here, retry mechanism is a problem to solve
for batch_stream in bin_batch_streams:
cur_retry = 0
batch_stream_data = list(batch_stream)
exception = None
while cur_retry < max_retry_cnt:
L.trace(
f'pushing object stream. rs_key={rs_key}, rs_header={rs_header}, cur_retry={cur_retry}'
)
try:
stub.push(bs_helper.generate_packet(
batch_stream_data, cur_retry),
timeout=per_stream_timeout)
exception = None
break
except Exception as e:
if cur_retry <= max_retry_cnt - long_retry_cnt:
retry_interval = round(
min(2 * cur_retry, 20) + random.random() * 10, 3)
else:
retry_interval = round(300 + random.random() * 10, 3)
L.warn(
f"push object error. rs_key={rs_key}, partition_id={rs_header._partition_id}, rs_header={rs_header}, max_retry_cnt={max_retry_cnt}, cur_retry={cur_retry}, retry_interval={retry_interval}",
exc_info=e)
time.sleep(retry_interval)
if isinstance(e,
RpcError) and e.code().name == 'UNAVAILABLE':
channel = grpc_channel_factory.create_channel(
self.ctx.proxy_endpoint, refresh=True)
stub = proxy_pb2_grpc.DataTransferServiceStub(channel)
exception = e
finally:
cur_retry += 1
if exception is not None:
L.exception(
f"push object failed. rs_key={rs_key}, partition_id={rs_header._partition_id}, rs_header={rs_header}, cur_retry={cur_retry}",
exc_info=exception)
raise exception
L.trace(
f'pushed object stream. rs_key={rs_key}, rs_header={rs_header}, cur_retry={cur_retry - 1}'
)
L.debug(
f"pushed object: rs_key={rs_key}, rs_header={rs_header}, is_none={obj is None}, elapsed={time.time() - start_time}"
)
self.ctx.pushing_latch.count_down()