Coverage for pyspark/shuffle.py : 74%
Hot-keys on this page
r m x p toggle line displays
j k next/prev highlighted chunk
0 (zero) top of page
1 (one) first highlighted chunk
|
# # Licensed to the Apache Software Foundation (ASF) under one or more # contributor license agreements. See the NOTICE file distributed with # this work for additional information regarding copyright ownership. # The ASF licenses this file to You under the Apache License, Version 2.0 # (the "License"); you may not use this file except in compliance with # the License. You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. #
CompressedSerializer, AutoBatchedSerializer
""" Return the used memory in MiB """ global process else: info = process.get_memory_info()
""" Return the used memory in MiB """
else: warnings.warn("Please install psutil to have better " "support with spilling") if platform.system() == "Darwin": import resource rss = resource.getrusage(resource.RUSAGE_SELF).ru_maxrss return rss >> 20 # TODO: support windows
return 0
""" Get all the directories """ # different order in different processes and instances rnd = random.Random(os.getpid() + id(dirs)) random.shuffle(dirs, rnd.random)
# global stats
""" Aggregator has tree functions to merge values into combiner.
createCombiner: (value) -> combiner mergeValue: (combine, value) -> combiner mergeCombiners: (combiner, combiner) -> combiner """
""" SimpleAggregator is useful for the cases that combiners have same type with values """
""" Merge shuffled data together by aggregator """
""" Combine the items by creator and combiner """ raise NotImplementedError
""" Merge the combined items by mergeCombiner """ raise NotImplementedError
""" Return the merged items ad iterator """ raise NotImplementedError
# always use PickleSerializer to simplify implementation
""" External merger will dump the aggregated data into disks when memory usage goes above the limit, then merge them together.
This class works as follows:
- It repeatedly combine the items and save them in one dict in memory.
- When the used memory goes above memory limit, it will split the combined data into partitions by hash code, dump them into disk, one file per partition.
- Then it goes through the rest of the iterator, combine items into different dict by hash. Until the used memory goes over memory limit, it dump all the dicts into disks, one file per dict. Repeat this again until combine all the items.
- Before return any items, it will load each partition and combine them separately. Yield them before loading next partition.
- During loading a partition, if the memory goes over limit, it will partition the loaded data and dump them into disks and load them partition by partition again.
`data` and `pdata` are used to hold the merged items in memory. At first, all the data are merged into `data`. Once the used memory goes over limit, the items in `data` are dumped into disks, `data` will be cleared, all rest of items will be merged into `pdata` and then dumped into disks. Before returning, all the items in `pdata` will be dumped into disks.
Finally, if any items were spilled into disks, each partition will be merged into `data` and be yielded, then cleared.
Examples -------- >>> agg = SimpleAggregator(lambda x, y: x + y) >>> merger = ExternalMerger(agg, 10) >>> N = 10000 >>> merger.mergeValues(zip(range(N), range(N))) >>> assert merger.spills > 0 >>> sum(v for k,v in merger.items()) 49995000
>>> merger = ExternalMerger(agg, 10) >>> merger.mergeCombiners(zip(range(N), range(N))) >>> assert merger.spills > 0 >>> sum(v for k,v in merger.items()) 49995000 """
# the max total partitions created recursively
localdirs=None, scale=1, partitions=59, batch=1000): # number of partitions when spill data into disks # check the memory after # of items merged # scale is used to scale down the hash of key for recursive hash map # un-partitioned merged data # partitioned merged data, list of dicts # number of chunks dumped into disks # randomize the hash of key, id(o) is the address of o (aligned by 8)
""" Choose one directory for spill by number n """
""" Return the next memory limit. If the memory is not released after spilling, it will dump the data only when the used memory starts to increase. """
""" Combine the items by creator and combiner """ # speedup attribute lookup
else:
self._spill()
""" Return the partition for key """
""" How much of memory for this obj, assume that all the objects consume similar bytes of memory """
""" Merge (K,V) pair by mergeCombiner """ # speedup attribute lookup
else:
self._spill()
""" dump already partitioned data into disks.
It will dump the data in batch for better performance. """ global MemoryBytesSpilled, DiskBytesSpilled
# The data has not been partitioned, it will iterator the # dataset once, write them into different files, has no # additional memory. It only called when the memory goes # above limit at the first time.
# open all the files for writing for i in range(self.partitions)]
# put one item in batch, make it compatible with load_stream # it will increase the memory if dump them in batch
else: # dump items in batch
""" Return all merged items as iterator """
""" Return all partitioned items as iterator """ # disable partitioning and spilling when merge combiners from disk
# remove the merged partition finally:
# do not check memory during merging
# limit the total partitions and j < self.spills - 1 and get_used_memory() > limit): self.data.clear() # will read from disk again gc.collect() # release the memory as much as possible return self._recursive_merged_items(index)
""" merge the partitioned items and return the as iterator
If one partition can not be fit in memory, then them will be partitioned and merged recursively. """ subdirs = [os.path.join(d, "parts", str(index)) for d in self.localdirs] m = ExternalMerger(self.agg, self.memory_limit, self.serializer, subdirs, self.scale * self.partitions, self.partitions, self.batch) m.pdata = [{} for _ in range(self.partitions)] limit = self._next_limit()
for j in range(self.spills): path = self._get_spill_dir(j) p = os.path.join(path, str(index)) with open(p, 'rb') as f: m.mergeCombiners(self.serializer.load_stream(f), 0)
if get_used_memory() > limit: m._spill() limit = self._next_limit()
return m._external_items()
""" Clean up all the files in disks """
""" ExternalSorter will divide the elements into chunks, sort them in memory and dump them into disks, finally merge them back.
The spilling will only happen when the used memory goes above the limit.
Examples -------- >>> sorter = ExternalSorter(1) # 1M >>> import random >>> l = list(range(1024)) >>> random.shuffle(l) >>> sorted(l) == list(sorter.sorted(l)) True >>> sorted(l) == list(sorter.sorted(l, key=lambda x: -x, reverse=True)) True """
""" Choose one directory for spill by number n """
""" Return the next memory limit. If the memory is not released after spilling, it will dump the data only when the used memory starts to increase. """
""" Sort the elements in iterator, do external sort when the memory goes above the limit. """ global MemoryBytesSpilled, DiskBytesSpilled # pick elements in batch
# sort them inplace will save memory
# close the file explicit once we consume all the items # to avoid ResourceWarning in Python3
""" ExternalList can have many items which cannot be hold in memory in the same time.
Examples -------- >>> l = ExternalList(list(range(100))) >>> len(l) 100 >>> l.append(10) >>> len(l) 101 >>> for i in range(20240): ... l.append(i) >>> len(l) 20341 >>> import pickle >>> l2 = pickle.loads(pickle.dumps(l)) >>> len(l2) 20341 >>> list(l2)[100] 10 """
else:
else:
# read all items from disks first
# dump them into disk if the key is huge
""" dump the values into disk """ global MemoryBytesSpilled, DiskBytesSpilled
""" An external list for list.
Examples -------- >>> l = ExternalListOfList([[i, i] for i in range(100)]) >>> len(l) 200 >>> l.append(range(10)) >>> len(l) 210 >>> len(list(l)) 210 """
# already counted 1 in ExternalList.append
""" Group a sorted iterator as [(k1, it1), (k2, it2), ...]
Examples -------- >>> k = [i // 3 for i in range(6)] >>> v = [[i] for i in range(6)] >>> g = GroupByKey(zip(k, v)) >>> [(k, list(it)) for k, it in g] [(0, [0, 1, 2]), (1, [3, 4, 5])] """
else:
""" Group by the items by key. If any partition of them can not been hold in memory, it will do sort based group by.
This class works as follows:
- It repeatedly group the items by key and save them in one dict in memory.
- When the used memory goes above memory limit, it will split the combined data into partitions by hash code, dump them into disk, one file per partition. If the number of keys in one partitions is smaller than 1000, it will sort them by key before dumping into disk.
- Then it goes through the rest of the iterator, group items by key into different dict by hash. Until the used memory goes over memory limit, it dump all the dicts into disks, one file per dict. Repeat this again until combine all the items. It also will try to sort the items by key in each partition before dumping into disks.
- It will yield the grouped items partitions by partitions. If the data in one partitions can be hold in memory, then it will load and combine them in memory and yield.
- If the dataset in one partition cannot be hold in memory, it will sort them first. If all the files are already sorted, it merge them by heap.merge(), so it will do external sort for all the files.
- After sorting, `GroupByKey` class will put all the continuous items with the same key as a group, yield the values as an iterator. """
assert isinstance(self.serializer, BatchedSerializer) ser = self.serializer return FlattenedValuesSerializer(ser, 20)
return len(obj)
""" dump already partitioned data into disks. """ global MemoryBytesSpilled, DiskBytesSpilled path = self._get_spill_dir(self.spills) if not os.path.exists(path): os.makedirs(path)
used_memory = get_used_memory() if not self.pdata: # The data has not been partitioned, it will iterator the # data once, write them into different files, has no # additional memory. It only called when the memory goes # above limit at the first time.
# open all the files for writing streams = [open(os.path.join(path, str(i)), 'wb') for i in range(self.partitions)]
# If the number of keys is small, then the overhead of sort is small # sort them before dumping into disks self._sorted = len(self.data) < self.SORT_KEY_LIMIT if self._sorted: self.serializer = self.flattened_serializer() for k in sorted(self.data.keys()): h = self._partition(k) self.serializer.dump_stream([(k, self.data[k])], streams[h]) else: for k, v in self.data.items(): h = self._partition(k) self.serializer.dump_stream([(k, v)], streams[h])
for s in streams: DiskBytesSpilled += s.tell() s.close()
self.data.clear() # self.pdata is cached in `mergeValues` and `mergeCombiners` self.pdata.extend([{} for i in range(self.partitions)])
else: for i in range(self.partitions): p = os.path.join(path, str(i)) with open(p, "wb") as f: # dump items in batch if self._sorted: # sort by key only (stable) sorted_items = sorted(self.pdata[i].items(), key=operator.itemgetter(0)) self.serializer.dump_stream(sorted_items, f) else: self.serializer.dump_stream(self.pdata[i].items(), f) self.pdata[i].clear() DiskBytesSpilled += os.path.getsize(p)
self.spills += 1 gc.collect() # release the memory as much as possible MemoryBytesSpilled += max(used_memory - get_used_memory(), 0) << 20
size = sum(os.path.getsize(os.path.join(self._get_spill_dir(j), str(index))) for j in range(self.spills)) # if the memory can not hold all the partition, # then use sort based merge. Because of compression, # the data on disks will be much smaller than needed memory if size >= self.memory_limit << 17: # * 1M / 8 return self._merge_sorted_items(index)
self.data = {} for j in range(self.spills): path = self._get_spill_dir(j) p = os.path.join(path, str(index)) # do not check memory during merging with open(p, "rb") as f: self.mergeCombiners(self.serializer.load_stream(f), 0) return self.data.items()
""" load a partition from disk, then sort and group by key """ def load_partition(j): path = self._get_spill_dir(j) p = os.path.join(path, str(index)) with open(p, 'rb', 65536) as f: for v in self.serializer.load_stream(f): yield v
disk_items = [load_partition(j) for j in range(self.spills)]
if self._sorted: # all the partitions are already sorted sorted_items = heapq.merge(*disk_items, key=operator.itemgetter(0))
else: # Flatten the combined values, so it will not consume huge # memory during merging sort. ser = self.flattened_serializer() sorter = ExternalSorter(self.memory_limit, ser) sorted_items = sorter.sorted(itertools.chain(*disk_items), key=operator.itemgetter(0)) return ((k, vs) for k, vs in GroupByKey(sorted_items))
sys.exit(-1) |