from arvados.retry import retry_method
import config
import hashlib
+import hashlib
+import threading
+import Queue
def split(path):
"""split(path) -> streamname, filename
class BufferBlock(object):
- def __init__(self, locator, starting_size=2**14):
- self.locator = locator
+ WRITABLE = 0
+ PENDING = 1
+ COMMITTED = 2
+
+ def __init__(self, blockid, starting_size):
+ self.blockid = blockid
self.buffer_block = bytearray(starting_size)
self.buffer_view = memoryview(self.buffer_block)
self.write_pointer = 0
+ self.state = BufferBlock.WRITABLE
+ self._locator = None
def append(self, data):
- while (self.write_pointer+len(data)) > len(self.buffer_block):
- new_buffer_block = bytearray(len(self.buffer_block) * 2)
- new_buffer_block[0:self.write_pointer] = self.buffer_block[0:self.write_pointer]
- self.buffer_block = new_buffer_block
- self.buffer_view = memoryview(self.buffer_block)
- self.buffer_view[self.write_pointer:self.write_pointer+len(data)] = data
- self.write_pointer += len(data)
+ if self.state == BufferBlock.WRITABLE:
+ while (self.write_pointer+len(data)) > len(self.buffer_block):
+ new_buffer_block = bytearray(len(self.buffer_block) * 2)
+ new_buffer_block[0:self.write_pointer] = self.buffer_block[0:self.write_pointer]
+ self.buffer_block = new_buffer_block
+ self.buffer_view = memoryview(self.buffer_block)
+ self.buffer_view[self.write_pointer:self.write_pointer+len(data)] = data
+ self.write_pointer += len(data)
+ self._locator = None
+ else:
+ raise AssertionError("Buffer block is not writable")
def size(self):
return self.write_pointer
- def calculate_locator(self):
- return "%s+%i" % (hashlib.md5(self.buffer_view[0:self.write_pointer]).hexdigest(), self.size())
+ def locator(self):
+ if self._locator is None:
+ self._locator = "%s+%i" % (hashlib.md5(self.buffer_view[0:self.write_pointer]).hexdigest(), self.size())
+ return self._locator
+
+class AsyncKeepWriteErrors(Exception):
+ def __init__(self, errors):
+ self.errors = errors
+
+ def __repr__(self):
+ return "\n".join(self.errors)
+class BlockManager(object):
+ def __init__(self, keep):
+ self._keep = keep
+ self._bufferblocks = {}
+ self._put_queue = None
+ self._put_errors = None
+ self._put_threads = None
+ self._prefetch_queue = None
+ self._prefetch_threads = None
+
+ def alloc_bufferblock(self, blockid=None, starting_size=2**14):
+ if blockid is None:
+ blockid = "bufferblock%i" % len(self._bufferblocks)
+ bb = BufferBlock(blockid, starting_size=starting_size)
+ self._bufferblocks[bb.blockid] = bb
+ return bb
+
+ def stop_threads(self):
+ if self._put_threads is not None:
+ for t in self._put_threads:
+ self._put_queue.put(None)
+ for t in self._put_threads:
+ t.join()
+ self._put_threads = None
+ self._put_queue = None
+ self._put_errors = None
+
+ if self._prefetch_threads is not None:
+ for t in self._prefetch_threads:
+ self._prefetch_queue.put(None)
+ for t in self._prefetch_threads:
+ t.join()
+ self._prefetch_threads = None
+ self._prefetch_queue = None
+
+ def commit_bufferblock(self, block):
+ def worker(self):
+ while True:
+ try:
+ b = self._put_queue.get()
+ if b is None:
+ return
+ b._locator = self._keep.put(b.buffer_view[0:b.write_pointer].tobytes())
+ b.state = BufferBlock.COMMITTED
+ b.buffer_view = None
+ b.buffer_block = None
+ except Exception as e:
+ print e
+ self._put_errors.put(e)
+ finally:
+ if self._put_queue is not None:
+ self._put_queue.task_done()
+
+ if self._put_threads is None:
+ self._put_queue = Queue.Queue(maxsize=2)
+ self._put_errors = Queue.Queue()
+ self._put_threads = [threading.Thread(target=worker, args=(self,)),
+ threading.Thread(target=worker, args=(self,))]
+ for t in self._put_threads:
+ t.daemon = True
+ t.start()
+
+ block.state = BufferBlock.PENDING
+ self._put_queue.put(block)
+
+ def get_block(self, locator, num_retries, cache_only=False):
+ if locator in self._bufferblocks:
+ bb = self._bufferblocks[locator]
+ if bb.state != BufferBlock.COMMITTED:
+ return bb.buffer_view[0:bb.write_pointer].tobytes()
+ else:
+ locator = bb._locator
+ return self._keep.get(locator, num_retries=num_retries, cache_only=cache_only)
+
+ def commit_all(self):
+ for k,v in self._bufferblocks.items():
+ if v.state == BufferBlock.WRITABLE:
+ self.commit_bufferblock(v)
+ if self._put_queue is not None:
+ self._put_queue.join()
+ if not self._put_errors.empty():
+ e = []
+ try:
+ while True:
+ e.append(self._put_errors.get(False))
+ except Queue.Empty:
+ pass
+ raise AsyncKeepWriteErrors(e)
+
+ def block_prefetch(self, locator):
+ def worker(self):
+ while True:
+ try:
+ b = self._prefetch_queue.get()
+ if b is None:
+ return
+ self._keep.get(b)
+ except:
+ pass
+
+ if locator in self._bufferblocks:
+ return
+ if self._prefetch_threads is None:
+ self._prefetch_queue = Queue.Queue()
+ self._prefetch_threads = [threading.Thread(target=worker, args=(self,)),
+ threading.Thread(target=worker, args=(self,))]
+ for t in self._prefetch_threads:
+ t.daemon = True
+ t.start()
+ self._prefetch_queue.put(locator)
class ArvadosFile(object):
- def __init__(self, stream=[], segments=[], keep=None):
+ def __init__(self, parent, stream=[], segments=[]):
'''
stream: a list of Range objects representing a block stream
segments: a list of Range objects representing segments
'''
+ self.parent = parent
self._modified = True
- self._segments = []
+ self.segments = []
for s in segments:
- self.add_segment(stream, s.range_start, s.range_size)
+ self.add_segment(stream, s.locator, s.range_size)
self._current_bblock = None
- self._bufferblocks = None
- self._keep = keep
def set_unmodified(self):
self._modified = False
def truncate(self, size):
new_segs = []
- for r in self._segments:
+ for r in self.segments:
range_end = r.range_start+r.range_size
if r.range_start >= size:
# segment is past the trucate size, all done
else:
new_segs.append(r)
- self._segments = new_segs
+ self.segments = new_segs
self._modified = True
- def _keepget(self, locator, num_retries):
- if self._bufferblocks and locator in self._bufferblocks:
- bb = self._bufferblocks[locator]
- return bb.buffer_view[0:bb.write_pointer].tobytes()
- else:
- return self._keep.get(locator, num_retries=num_retries)
-
def readfrom(self, offset, size, num_retries):
if size == 0 or offset >= self.size():
return ''
- if self._keep is None:
- self._keep = KeepClient(num_retries=num_retries)
data = []
- # TODO: initiate prefetch on all blocks in the range (offset, offset + size + config.KEEP_BLOCK_SIZE)
- for lr in locators_and_ranges(self._segments, offset, size):
- # TODO: if data is empty, wait on block get, otherwise only
- # get more data if the block is already in the cache.
- data.append(self._keepget(lr.locator, num_retries=num_retries)[lr.segment_offset:lr.segment_offset+lr.segment_size])
- return ''.join(data)
+ for lr in locators_and_ranges(self.segments, offset, size + config.KEEP_BLOCK_SIZE):
+ self.parent._my_block_manager().block_prefetch(lr.locator)
- def _init_bufferblock(self):
- if self._bufferblocks is None:
- self._bufferblocks = {}
- self._current_bblock = BufferBlock("bufferblock%i" % len(self._bufferblocks))
- self._bufferblocks[self._current_bblock.locator] = self._current_bblock
+ for lr in locators_and_ranges(self.segments, offset, size):
+ d = self.parent._my_block_manager().get_block(lr.locator, num_retries=num_retries, cache_only=bool(data))
+ if d:
+ data.append(d[lr.segment_offset:lr.segment_offset+lr.segment_size])
+ else:
+ break
+ return ''.join(data)
def _repack_writes(self):
- pass
- # TODO: fixme
'''Test if the buffer block has more data than is referenced by actual segments
(this happens when a buffered write over-writes a file range written in
a previous buffered write). Re-pack the buffer block for efficiency
and to avoid leaking information.
'''
- segs = self._segments
+ segs = self.segments
# Sum up the segments to get the total bytes of the file referencing
# into the buffer block.
- bufferblock_segs = [s for s in segs if s.locator == self._current_bblock.locator]
+ bufferblock_segs = [s for s in segs if s.locator == self._current_bblock.blockid]
write_total = sum([s.range_size for s in bufferblock_segs])
if write_total < self._current_bblock.size():
# There is more data in the buffer block than is actually accounted for by segments, so
# re-pack into a new buffer by copying over to a new buffer block.
- new_bb = BufferBlock(self._current_bblock.locator, starting_size=write_total)
+ new_bb = self.parent._my_block_manager().alloc_bufferblock(self._current_bblock.blockid, starting_size=write_total)
for t in bufferblock_segs:
new_bb.append(self._current_bblock.buffer_view[t.segment_offset:t.segment_offset+t.range_size].tobytes())
t.segment_offset = new_bb.size() - t.range_size
self._current_bblock = new_bb
- self._bufferblocks[self._current_bblock.locator] = self._current_bblock
-
def writeto(self, offset, data, num_retries):
if len(data) == 0:
self._modified = True
- if self._current_bblock is None:
- self._init_bufferblock()
+ if self._current_bblock is None or self._current_bblock.state != BufferBlock.WRITABLE:
+ self._current_bblock = self.parent._my_block_manager().alloc_bufferblock()
- if (self._current_bblock.write_pointer + len(data)) > config.KEEP_BLOCK_SIZE:
+ if (self._current_bblock.size() + len(data)) > config.KEEP_BLOCK_SIZE:
self._repack_writes()
- if (self._current_bblock.write_pointer + len(data)) > config.KEEP_BLOCK_SIZE:
- self._init_bufferblock()
+ if (self._current_bblock.size() + len(data)) > config.KEEP_BLOCK_SIZE:
+ self.parent._my_block_manager().commit_bufferblock(self._current_bblock)
+ self._current_bblock = self.parent._my_block_manager().alloc_bufferblock()
self._current_bblock.append(data)
- replace_range(self._segments, offset, len(data), self._current_bblock.locator, self._current_bblock.write_pointer - len(data))
+ replace_range(self.segments, offset, len(data), self._current_bblock.blockid, self._current_bblock.write_pointer - len(data))
def add_segment(self, blocks, pos, size):
self._modified = True
for lr in locators_and_ranges(blocks, pos, size):
- last = self._segments[-1] if self._segments else Range(0, 0, 0)
+ last = self.segments[-1] if self.segments else Range(0, 0, 0)
r = Range(lr.locator, last.range_start+last.range_size, lr.segment_size, lr.segment_offset)
- self._segments.append(r)
+ self.segments.append(r)
def size(self):
- if self._segments:
- n = self._segments[-1]
+ if self.segments:
+ n = self.segments[-1]
return n.range_start + n.range_size
else:
return 0
projects / arvados.git / blobdiff