def _check_partitioner(self, map_task):
output_partitions = gfile.ListDirectory(map_task.output_base_dir)
output_partitions = [
x for x in output_partitions if "SUCCESS" not in x
]
self.assertEqual(len(output_partitions), map_task.output_partition_num)
partition_dirs = ["{}/{}".format(map_task.output_base_dir, x) \
for x in output_partitions]
total_cnt = 0
for partition in output_partitions:
dpath = "{}/{}".format(map_task.output_base_dir, partition)
partition_id = partition.split("_")[-1]
partition_id = int(partition_id)
segments = gfile.ListDirectory(dpath)
for segment in segments:
fpath = "{}/{}".format(dpath, segment)
event_time = 0
for record in tf.python_io.tf_record_iterator(fpath):
tf_item = TfExampleItem(record)
self.assertTrue(
tf_item.event_time >= event_time,
"{}, {}".format(tf_item.event_time, event_time))
event_time = tf_item.event_time ## assert order
self.assertEqual(partition_id, CityHash32(tf_item.raw_id) \
% map_task.output_partition_num)
total_cnt += 1
self.assertEqual(total_cnt,
self._partition_item_num * self._input_partition_num)
def _create_data_block(self, partition_id):
dbm = data_block_manager.DataBlockManager(self.data_source, partition_id)
self.assertEqual(dbm.get_dumped_data_block_count(), 0)
self.assertEqual(dbm.get_lastest_data_block_meta(), None)
leader_index = 0
follower_index = 65536
for i in range(64):
builder = DataBlockBuilder(
common.data_source_data_block_dir(self.data_source),
self.data_source.data_source_meta.name,
partition_id, i,
dj_pb.WriterOptions(output_writer='TF_RECORD'), None
)
builder.set_data_block_manager(dbm)
for j in range(4):
feat = {}
example_id = '{}'.format(i * 1024 + j).encode()
feat['example_id'] = tf.train.Feature(
bytes_list=tf.train.BytesList(value=[example_id]))
event_time = random.randint(0, 10)
feat['event_time'] = tf.train.Feature(
int64_list=tf.train.Int64List(value=[event_time]))
feat['leader_index'] = tf.train.Feature(
int64_list=tf.train.Int64List(value=[leader_index]))
feat['follower_index'] = tf.train.Feature(
int64_list=tf.train.Int64List(value=[follower_index]))
example = tf.train.Example(features=tf.train.Features(feature=feat))
builder.append_item(TfExampleItem(example.SerializeToString()),
leader_index, follower_index)
leader_index += 1
follower_index += 1
self.data_block_matas.append(builder.finish_data_block())
def add_data_block(self, partition_id, x, y):
dbm = self._dbms[partition_id]
builder = DataBlockBuilder(
common.data_source_data_block_dir(self._data_source),
self._data_source.data_source_meta.name, partition_id,
dbm.get_dumped_data_block_count(),
dj_pb.WriterOptions(output_writer="TF_RECORD"), None)
builder.set_data_block_manager(dbm)
for i in range(x.shape[0]):
feat = {}
exam_id = '{}'.format(i).encode()
feat['example_id'] = Feature(
bytes_list=BytesList(value=[exam_id]))
feat['event_time'] = Feature(
int64_list = Int64List(value=[i])
)
feat['x'] = Feature(float_list=FloatList(value=list(x[i])))
if y is not None:
feat['y'] = Feature(int64_list=Int64List(value=[y[i]]))
example = Example(features=Features(feature=feat))
builder.append_item(TfExampleItem(example.SerializeToString()), i, 0)
return builder.finish_data_block()
def generate_leader_raw_data(self):
dbm = data_block_manager.DataBlockManager(self.data_source_l, 0)
raw_data_dir = os.path.join(self.data_source_l.raw_data_dir,
common.partition_repr(0))
if gfile.Exists(raw_data_dir):
gfile.DeleteRecursively(raw_data_dir)
gfile.MakeDirs(raw_data_dir)
rdm = raw_data_visitor.RawDataManager(self.etcd, self.data_source_l, 0)
block_index = 0
builder = DataBlockBuilder(
self.data_source_l.raw_data_dir,
self.data_source_l.data_source_meta.name, 0, block_index,
dj_pb.WriterOptions(output_writer='TF_RECORD'), None)
process_index = 0
start_index = 0
for i in range(0, self.leader_end_index + 3):
if (i > 0 and i % 2048 == 0) or (i == self.leader_end_index + 2):
meta = builder.finish_data_block()
if meta is not None:
ofname = common.encode_data_block_fname(
self.data_source_l.data_source_meta.name, meta)
fpath = os.path.join(raw_data_dir, ofname)
self.manifest_manager.add_raw_data(0, [
dj_pb.RawDataMeta(
file_path=fpath,
timestamp=timestamp_pb2.Timestamp(seconds=3))
], False)
process_index += 1
start_index += len(meta.example_ids)
block_index += 1
builder = DataBlockBuilder(
self.data_source_l.raw_data_dir,
self.data_source_l.data_source_meta.name, 0, block_index,
dj_pb.WriterOptions(output_writer='TF_RECORD'), None)
feat = {}
pt = i + 1 << 30
if i % 3 == 0:
pt = i // 3
example_id = '{}'.format(pt).encode()
feat['example_id'] = tf.train.Feature(
bytes_list=tf.train.BytesList(value=[example_id]))
event_time = 150000000 + pt
feat['event_time'] = tf.train.Feature(
int64_list=tf.train.Int64List(value=[event_time]))
example = tf.train.Example(features=tf.train.Features(
feature=feat))
builder.append_item(TfExampleItem(example.SerializeToString()), i,
i)
fpaths = [
os.path.join(raw_data_dir, f)
for f in gfile.ListDirectory(raw_data_dir)
if not gfile.IsDirectory(os.path.join(raw_data_dir, f))
]
for fpath in fpaths:
if not fpath.endswith(common.DataBlockSuffix):
gfile.Remove(fpath)
def _check_merge(self, reduce_task):
dpath = os.path.join(self._merge_output_dir, \
common.partition_repr(reduce_task.partition_id))
fpaths = gfile.ListDirectory(dpath)
fpaths = sorted(fpaths, key=lambda fpath: fpath, reverse=False)
event_time = 0
total_cnt = 0
for fpath in fpaths:
fpath = os.path.join(dpath, fpath)
logging.info("check merge path:{}".format(fpath))
for record in tf.python_io.tf_record_iterator(fpath):
tf_item = TfExampleItem(record)
self.assertTrue(tf_item.event_time >= event_time)
event_time = tf_item.event_time
total_cnt += 1
return total_cnt
def _create_data_block(self, data_source, partition_id, x, y):
data_block_metas = []
dbm = data_block_manager.DataBlockManager(data_source, partition_id)
self.assertEqual(dbm.get_dumped_data_block_count(), 0)
self.assertEqual(dbm.get_lastest_data_block_meta(), None)
N = 200
chunk_size = x.shape[0] // N
leader_index = 0
follower_index = N * chunk_size * 10
for i in range(N):
builder = DataBlockBuilder(
common.data_source_data_block_dir(data_source),
data_source.data_source_meta.name,
partition_id, i,
dj_pb.WriterOptions(output_writer="TF_RECORD"), None
)
builder.set_data_block_manager(dbm)
for j in range(chunk_size):
feat = {}
idx = i * chunk_size + j
exam_id = '{}'.format(idx).encode()
feat['example_id'] = Feature(
bytes_list=BytesList(value=[exam_id]))
evt_time = random.randint(1, 1000)
feat['event_time'] = Feature(
int64_list = Int64List(value=[evt_time])
)
feat['x'] = Feature(float_list=FloatList(value=list(x[idx])))
if y is not None:
feat['y'] = Feature(int64_list=Int64List(value=[y[idx]]))
feat['leader_index'] = Feature(
int64_list = Int64List(value=[leader_index])
)
feat['follower_index'] = Feature(
int64_list = Int64List(value=[follower_index])
)
example = Example(features=Features(feature=feat))
builder.append_item(TfExampleItem(example.SerializeToString()),
leader_index, follower_index)
leader_index += 1
follower_index += 1
data_block_metas.append(builder.finish_data_block())
self.max_index = follower_index
return data_block_metas
def generate_follower_data_block(self):
dbm = data_block_manager.DataBlockManager(self.data_source_f, 0)
self.assertEqual(dbm.get_dumped_data_block_count(), 0)
self.assertEqual(dbm.get_lastest_data_block_meta(), None)
leader_index = 0
follower_index = 65536
self.dumped_metas = []
for i in range(5):
builder = DataBlockBuilder(
common.data_source_data_block_dir(self.data_source_f),
self.data_source_f.data_source_meta.name, 0, i,
dj_pb.WriterOptions(output_writer='TF_RECORD'), None)
builder.set_data_block_manager(dbm)
for j in range(1024):
feat = {}
example_id = '{}'.format(i * 1024 + j).encode()
feat['example_id'] = tf.train.Feature(
bytes_list=tf.train.BytesList(value=[example_id]))
event_time = 150000000 + i * 1024 + j
feat['event_time'] = tf.train.Feature(
int64_list=tf.train.Int64List(value=[event_time]))
feat['leader_index'] = tf.train.Feature(
int64_list=tf.train.Int64List(value=[leader_index]))
feat['follower_index'] = tf.train.Feature(
int64_list=tf.train.Int64List(value=[follower_index]))
example = tf.train.Example(features=tf.train.Features(
feature=feat))
builder.append_item(TfExampleItem(example.SerializeToString()),
leader_index, follower_index)
leader_index += 3
follower_index += 1
meta = builder.finish_data_block()
self.dumped_metas.append(meta)
self.leader_start_index = 0
self.leader_end_index = leader_index
self.assertEqual(dbm.get_dumped_data_block_count(), 5)
for (idx, meta) in enumerate(self.dumped_metas):
self.assertEqual(dbm.get_data_block_meta_by_index(idx), meta)
def generate_raw_data(self, begin_index, item_count):
raw_data_dir = os.path.join(self.raw_data_dir,
common.partition_repr(0))
if not gfile.Exists(raw_data_dir):
gfile.MakeDirs(raw_data_dir)
self.total_raw_data_count += item_count
useless_index = 0
rdm = raw_data_visitor.RawDataManager(self.kvstore, self.data_source,
0)
fpaths = []
for block_index in range(0, item_count // 2048):
builder = DataBlockBuilder(
self.raw_data_dir,
self.data_source.data_source_meta.name, 0, block_index,
dj_pb.WriterOptions(output_writer='TF_RECORD'), None)
cands = list(
range(begin_index + block_index * 2048,
begin_index + (block_index + 1) * 2048))
start_index = cands[0]
for i in range(len(cands)):
if random.randint(1, 4) > 2:
continue
a = random.randint(i - 32, i + 32)
b = random.randint(i - 32, i + 32)
if a < 0:
a = 0
if a >= len(cands):
a = len(cands) - 1
if b < 0:
b = 0
if b >= len(cands):
b = len(cands) - 1
if (abs(cands[a] - i - start_index) <= 32
and abs(cands[b] - i - start_index) <= 32):
cands[a], cands[b] = cands[b], cands[a]
for example_idx in cands:
feat = {}
example_id = '{}'.format(example_idx).encode()
feat['example_id'] = tf.train.Feature(
bytes_list=tf.train.BytesList(value=[example_id]))
event_time = 150000000 + example_idx
feat['event_time'] = tf.train.Feature(
int64_list=tf.train.Int64List(value=[event_time]))
label = random.choice([1, 0])
if random.random() < 0.8:
feat['label'] = tf.train.Feature(
int64_list=tf.train.Int64List(value=[label]))
example = tf.train.Example(features=tf.train.Features(
feature=feat))
builder.append_item(TfExampleItem(example.SerializeToString()),
useless_index, useless_index)
useless_index += 1
meta = builder.finish_data_block()
fname = common.encode_data_block_fname(
self.data_source.data_source_meta.name, meta)
fpath = os.path.join(raw_data_dir, fname)
fpaths.append(
dj_pb.RawDataMeta(
file_path=fpath,
timestamp=timestamp_pb2.Timestamp(seconds=3)))
self.g_data_block_index += 1
all_files = [
os.path.join(raw_data_dir, f)
for f in gfile.ListDirectory(raw_data_dir)
if not gfile.IsDirectory(os.path.join(raw_data_dir, f))
]
for fpath in all_files:
if not fpath.endswith(common.DataBlockSuffix):
gfile.Remove(fpath)
self.manifest_manager.add_raw_data(0, fpaths, False)
def generate_raw_data(self, etcd, rdp, data_source, partition_id,
block_size, shuffle_win_size, feat_key_fmt,
feat_val_fmt):
dbm = data_block_manager.DataBlockManager(data_source, partition_id)
raw_data_dir = os.path.join(data_source.raw_data_dir,
common.partition_repr(partition_id))
if gfile.Exists(raw_data_dir):
gfile.DeleteRecursively(raw_data_dir)
gfile.MakeDirs(raw_data_dir)
useless_index = 0
new_raw_data_fnames = []
for block_index in range(self.total_index // block_size):
builder = DataBlockBuilder(
data_source.raw_data_dir, data_source.data_source_meta.name,
partition_id, block_index,
dj_pb.WriterOptions(output_writer='TF_RECORD'), None)
cands = list(
range(block_index * block_size,
(block_index + 1) * block_size))
start_index = cands[0]
for i in range(len(cands)):
if random.randint(1, 4) > 2:
continue
a = random.randint(i - shuffle_win_size, i + shuffle_win_size)
b = random.randint(i - shuffle_win_size, i + shuffle_win_size)
if a < 0:
a = 0
if a >= len(cands):
a = len(cands) - 1
if b < 0:
b = 0
if b >= len(cands):
b = len(cands) - 1
if (abs(cands[a] - i - start_index) <= shuffle_win_size and
abs(cands[b] - i - start_index) <= shuffle_win_size):
cands[a], cands[b] = cands[b], cands[a]
for example_idx in cands:
feat = {}
example_id = '{}'.format(example_idx).encode()
feat['example_id'] = tf.train.Feature(
bytes_list=tf.train.BytesList(value=[example_id]))
event_time = 150000000 + example_idx
feat['event_time'] = tf.train.Feature(
int64_list=tf.train.Int64List(value=[event_time]))
feat[feat_key_fmt.format(example_idx)] = tf.train.Feature(
bytes_list=tf.train.BytesList(
value=[feat_val_fmt.format(example_idx).encode()]))
example = tf.train.Example(features=tf.train.Features(
feature=feat))
builder.append_item(TfExampleItem(example.SerializeToString()),
useless_index, useless_index)
useless_index += 1
meta = builder.finish_data_block()
fname = common.encode_data_block_fname(
data_source.data_source_meta.name, meta)
new_raw_data_fnames.append(os.path.join(raw_data_dir, fname))
fpaths = [
os.path.join(raw_data_dir, f)
for f in gfile.ListDirectory(raw_data_dir)
if not gfile.IsDirectory(os.path.join(raw_data_dir, f))
]
for fpath in fpaths:
if fpath.endswith(common.DataBlockMetaSuffix):
gfile.Remove(fpath)
rdp.publish_raw_data(partition_id, new_raw_data_fnames)
def test_data_block_manager(self):
data_block_datas = []
data_block_metas = []
leader_index = 0
follower_index = 65536
for i in range(5):
fill_examples = []
builder = DataBlockBuilder(
self.data_source.data_block_dir,
self.data_source.data_source_meta.name, 0, i,
dj_pb.WriterOptions(output_writer='TF_RECORD'), None)
builder.set_data_block_manager(self.data_block_manager)
for j in range(1024):
feat = {}
example_id = '{}'.format(i * 1024 + j).encode()
feat['example_id'] = tf.train.Feature(
bytes_list=tf.train.BytesList(value=[example_id]))
event_time = 150000000 + i * 1024 + j
feat['event_time'] = tf.train.Feature(
int64_list=tf.train.Int64List(value=[event_time]))
feat['leader_index'] = tf.train.Feature(
int64_list=tf.train.Int64List(value=[leader_index]))
feat['follower_index'] = tf.train.Feature(
int64_list=tf.train.Int64List(value=[follower_index]))
example = tf.train.Example(features=tf.train.Features(
feature=feat))
builder.append_item(TfExampleItem(example.SerializeToString()),
leader_index, follower_index)
fill_examples.append((example, {
'example_id': example_id,
'event_time': event_time,
'leader_index': leader_index,
'follower_index': follower_index
}))
leader_index += 1
follower_index += 1
meta = builder.finish_data_block()
data_block_datas.append(fill_examples)
data_block_metas.append(meta)
self.assertEqual(self.data_block_manager.get_dumped_data_block_count(),
5)
self.assertEqual(self.data_block_manager.get_lastest_data_block_meta(),
data_block_metas[-1])
for (idx, meta) in enumerate(data_block_metas):
self.assertEqual(
self.data_block_manager.get_data_block_meta_by_index(idx),
meta)
self.assertEqual(
meta.block_id,
common.encode_block_id(self.data_source.data_source_meta.name,
meta))
self.assertEqual(
self.data_block_manager.get_data_block_meta_by_index(5), None)
data_block_dir = os.path.join(self.data_source.data_block_dir,
common.partition_repr(0))
for (i, meta) in enumerate(data_block_metas):
data_block_fpath = os.path.join(
data_block_dir, meta.block_id) + common.DataBlockSuffix
data_block_meta_fpath = os.path.join(
data_block_dir,
common.encode_data_block_meta_fname(
self.data_source.data_source_meta.name, 0,
meta.data_block_index))
self.assertTrue(gfile.Exists(data_block_fpath))
self.assertTrue(gfile.Exists(data_block_meta_fpath))
fiter = tf.io.tf_record_iterator(data_block_meta_fpath)
remote_meta = text_format.Parse(
next(fiter).decode(), dj_pb.DataBlockMeta())
self.assertEqual(meta, remote_meta)
for (j, record) in enumerate(
tf.io.tf_record_iterator(data_block_fpath)):
example = tf.train.Example()
example.ParseFromString(record)
stored_data = data_block_datas[i][j]
self.assertEqual(example, stored_data[0])
feat = example.features.feature
stored_feat = stored_data[1]
self.assertTrue('example_id' in feat)
self.assertTrue('example_id' in stored_feat)
self.assertEqual(stored_feat['example_id'],
'{}'.format(i * 1024 + j).encode())
self.assertEqual(stored_feat['example_id'],
feat['example_id'].bytes_list.value[0])
self.assertTrue('event_time' in feat)
self.assertTrue('event_time' in stored_feat)
self.assertEqual(stored_feat['event_time'],
feat['event_time'].int64_list.value[0])
self.assertTrue('leader_index' in feat)
self.assertTrue('leader_index' in stored_feat)
self.assertEqual(stored_feat['leader_index'],
feat['leader_index'].int64_list.value[0])
self.assertTrue('follower_index' in feat)
self.assertTrue('follower_index' in stored_feat)
self.assertEqual(stored_feat['follower_index'],
feat['follower_index'].int64_list.value[0])
self.assertEqual(j, 1023)
data_block_manager2 = data_block_manager.DataBlockManager(
self.data_source, 0)
self.assertEqual(self.data_block_manager.get_dumped_data_block_count(),
5)
