def test_beansdb(self):
N = 100
l = list(range(N))
d = list(zip(list(map(lambda x: str(x).encode('utf-8'), l)), l))
num_splits = 10
rdd = self.sc.makeRDD(d, num_splits)
with temppath('beansdb') as root:
def newpath(c):
return os.path.join(root, str(c))
def check_rdd(_rdd, files, num_w, num_r):
self.assertEqual(files,
['%s/%03d.data' % (path, i) for i in range(num_w)])
self.assertEqual(len(_rdd), num_r)
self.assertEqual(_rdd.count(), N)
self.assertEqual(sorted(_rdd.map(lambda k_v: (k_v[0], k_v[1][0])).collect()), sorted(d))
s = _rdd.map(lambda x: x[1][0]).reduce(lambda x, y: x + y)
self.assertEqual(s, sum(l))
path = newpath(0)
files = rdd.saveAsBeansdb(path)
rdd = self.sc.beansdb(path, depth=0, filter=lambda x: x != "")
check_rdd(rdd, files, num_splits, num_splits)
path = newpath(1)
files = rdd.saveAsBeansdb(path, valueWithMeta=True)
rdd = self.sc.beansdb(path, depth=0, fullscan=True, only_latest=True)
num_splits_reduce = int(ceil(num_splits / 4))
check_rdd(rdd, files, num_splits, num_splits_reduce)
path = newpath(num_splits_reduce)
files = rdd.map(lambda k_v1: (k_v1[0], k_v1[1][0])).saveAsBeansdb(path)
rdd = self.sc.beansdb(path, raw=True, depth=0, fullscan=True)
rdd = rdd.mapValue(lambda v: (restore_value(*v[0]), v[1], v[2]))
check_rdd(rdd, files, num_splits_reduce, num_splits_reduce)
def test_beansdb(self):
N = 100
l = list(range(N))
d = list(zip(list(map(lambda x: str(x).encode('utf-8'), l)), l))
num_splits = 10
rdd = self.sc.makeRDD(d, num_splits)
with temppath('beansdb') as root:
def newpath(c):
return os.path.join(root, str(c))
def check_rdd(_rdd, files, num_w, num_r):
self.assertEqual(
files, ['%s/%03d.data' % (path, i) for i in range(num_w)])
self.assertEqual(len(_rdd), num_r)
self.assertEqual(_rdd.count(), N)
self.assertEqual(
sorted(
_rdd.map(lambda k_v: (k_v[0], k_v[1][0])).collect()),
sorted(d))
s = _rdd.map(lambda x: x[1][0]).reduce(lambda x, y: x + y)
self.assertEqual(s, sum(l))
path = newpath(0)
files = rdd.saveAsBeansdb(path)
rdd = self.sc.beansdb(path, depth=0, filter=lambda x: x != "")
check_rdd(rdd, files, num_splits, num_splits)
path = newpath(1)
files = rdd.saveAsBeansdb(path, valueWithMeta=True)
rdd = self.sc.beansdb(path,
depth=0,
fullscan=True,
only_latest=True)
num_splits_reduce = int(ceil(num_splits / 4))
check_rdd(rdd, files, num_splits, num_splits_reduce)
path = newpath(num_splits_reduce)
files = rdd.map(lambda k_v1: (k_v1[0], k_v1[1][0])).saveAsBeansdb(
path)
rdd = self.sc.beansdb(path, raw=True, depth=0, fullscan=True)
rdd = rdd.mapValue(lambda v: (restore_value(*v[0]), v[1], v[2]))
check_rdd(rdd, files, num_splits_reduce, num_splits_reduce)
def beansdb(self,
path,
depth=None,
filter=None,
fullscan=False,
raw=False,
only_latest=False):
"""(Key, (VALUE, Version, Timestamp)) data in beansdb
Data structure:
REC = (Key, TRIPLE)
TRIPLE = (VALUE, Version, Timestamp)
VALUE = RAW_VALUE | REAL_VALUE
RAW_VALUE = (flag, BYTES_VALUE)
Args:
path: beansdb data path
filter: used to filter key
depth: choice = [None, 0, 1, 2]. e.g. depth=2 assume dir tree like:
'path/[0-F]/[0-F]/%03d.data'
If depth is None, dpark will guess.
fullscan: NOT use index files, which contain (key, pos_in_datafile).
pairs.
Better use fullscan unless the filter selectivity is low.
Effect of using index:
inefficient random access
one split(task) for each file instead of each moosefs chunk
Omitted if filter is None.
raw: VALUE = RAW_VALUE if raw else REAL_VALUE.
only_latest: for each key, keeping the REC with the largest
Timestamp. This will append a reduceByKey RDD.
Need this because online beansdb data is log structured.
"""
key_filter = filter
self.init()
if key_filter is None:
fullscan = True
if isinstance(path, (tuple, list)):
rdd = self.union([
self.beansdb(p,
depth,
key_filter,
fullscan,
raw=True,
only_latest=False) for p in path
])
else:
path = os.path.realpath(path)
assert os.path.exists(path), "%s no exists" % path
if os.path.isdir(path):
subs = []
if not depth:
subs = [
os.path.join(path, n) for n in os.listdir(path)
if n.endswith('.data')
]
if subs:
rdd = self.union([
BeansdbFileRDD(self, p, key_filter, fullscan, raw=True)
for p in subs
])
else:
subs = [os.path.join(path, '%x' % i) for i in range(16)]
rdd = self.union([
self.beansdb(p,
depth and depth - 1,
key_filter,
fullscan,
raw=True,
only_latest=False) for p in subs
if os.path.exists(p)
])
else:
rdd = BeansdbFileRDD(self, path, key_filter, fullscan, raw)
# choose only latest version
if only_latest:
num_splits = min(int(ceil(len(rdd) / 4)), 800)
rdd = rdd.reduceByKey(lambda v1, v2: v1[2] > v2[2] and v1 or v2,
numSplits=num_splits)
if not raw:
rdd = rdd.mapValue(lambda v_ver_t: (restore_value(*v_ver_t[0]),
v_ver_t[1], v_ver_t[2]))
return rdd
def beansdb(self, path, depth=None, filter=None,
fullscan=False, raw=False, only_latest=False):
"""(Key, (VALUE, Version, Timestamp)) data in beansdb
Data structure:
REC = (Key, TRIPLE)
TRIPLE = (VALUE, Version, Timestamp)
VALUE = RAW_VALUE | REAL_VALUE
RAW_VALUE = (flag, BYTES_VALUE)
Args:
path: beansdb data path
filter: used to filter key
depth: choice = [None, 0, 1, 2]. e.g. depth=2 assume dir tree like:
'path/[0-F]/[0-F]/%03d.data'
If depth is None, dpark will guess.
fullscan: NOT use index files, which contain (key, pos_in_datafile).
pairs.
Better use fullscan unless the filter selectivity is low.
Effect of using index:
inefficient random access
one split(task) for each file instead of each moosefs chunk
Omitted if filter is None.
raw: VALUE = RAW_VALUE if raw else REAL_VALUE.
only_latest: for each key, keeping the REC with the largest
Timestamp. This will append a reduceByKey RDD.
Need this because online beansdb data is log structured.
"""
key_filter = filter
self.init()
if key_filter is None:
fullscan = True
if isinstance(path, (tuple, list)):
rdd = self.union([self.beansdb(p, depth, key_filter, fullscan,
raw=True, only_latest=False)
for p in path])
else:
path = os.path.realpath(path)
assert os.path.exists(path), "%s no exists" % path
if os.path.isdir(path):
subs = []
if not depth:
subs = [os.path.join(path, n) for n in os.listdir(path)
if n.endswith('.data')]
if subs:
rdd = self.union([BeansdbFileRDD(self, p, key_filter,
fullscan, raw=True)
for p in subs])
else:
subs = [os.path.join(path, '%x' % i) for i in range(16)]
rdd = self.union([self.beansdb(p, depth and depth - 1, key_filter,
fullscan, raw=True, only_latest=False)
for p in subs if os.path.exists(p)])
else:
rdd = BeansdbFileRDD(self, path, key_filter, fullscan, raw)
# choose only latest version
if only_latest:
num_splits = min(int(ceil(len(rdd) / 4)), 800)
rdd = rdd.reduceByKey(lambda v1, v2: v1[2] > v2[2] and v1 or v2,
numSplits=num_splits)
if not raw:
rdd = rdd.mapValue(lambda v_ver_t: (restore_value(*v_ver_t[0]), v_ver_t[1], v_ver_t[2]))
return rdd