Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

POC of Version Store with pluggable backend #586

Open
wants to merge 7 commits into
base: master
Choose a base branch
from
Open
Show file tree
Hide file tree
Changes from 3 commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 4 additions & 0 deletions .travis.yml
Original file line number Diff line number Diff line change
Expand Up @@ -28,6 +28,10 @@ install:
- pip install pytest-timeout --upgrade
- pip install pytest-xdist --upgrade
- pip install setuptools-git --upgrade
- pip install boto3 --upgrade
- pip install moto --upgrade
- pip install s3fs --upgrade
- pip install fastparquet --upgrade
script:
- pip freeze
- python setup.py test --pytest-args=-v
Empty file added arctic/pluggable/__init__.py
Empty file.
331 changes: 331 additions & 0 deletions arctic/pluggable/_kv_ndarray_store.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,331 @@
import logging
import hashlib

from bson.binary import Binary
import numpy as np

from arctic.exceptions import UnhandledDtypeException, DataIntegrityException, ArcticException
from arctic.store._version_store_utils import checksum

from arctic._compression import compress_array, decompress
from six.moves import xrange


logger = logging.getLogger(__name__)

_CHUNK_SIZE = 2 * 1024 * 1024 - 2048 # ~2 MB (a bit less for usePowerOf2Sizes)
_APPEND_SIZE = 1 * 1024 * 1024 # 1MB
_APPEND_COUNT = 60 # 1 hour of 1 min data


def _promote_struct_dtypes(dtype1, dtype2):
if not set(dtype1.names).issuperset(set(dtype2.names)):
raise Exception("Removing columns from dtype not handled")

def _promote(type1, type2):
if type2 is None:
return type1
if type1.shape is not None:
if not type1.shape == type2.shape:
raise Exception("We do not handle changes to dtypes that have shape")
return np.promote_types(type1.base, type2.base), type1.shape
return np.promote_types(type1, type2)
return np.dtype([(n, _promote(dtype1.fields[n][0], dtype2.fields.get(n, (None,))[0])) for n in dtype1.names])


def _resize_with_dtype(arr, dtype):
"""
This function will transform arr into an array with the same type as dtype. It will do this by
filling new columns with zeros (or NaNs, if it is a float column). Also, columns that are not
in the new dtype will be dropped.
"""
structured_arrays = dtype.names is not None and arr.dtype.names is not None
old_columns = set(arr.dtype.names or [])
new_columns = set(dtype.names or [])

# In numpy 1.9 the ndarray.astype method used to handle changes in number of fields. The code below
# should replicate the same behaviour the old astype used to have.
#
# One may be tempted to use np.lib.recfunctions.stack_arrays to implement both this step and the
# concatenate that follows but it 2x slower and it requires providing your own default values (instead
# of np.zeros).
#
# Numpy 1.14 supports doing new_arr[old_columns] = arr[old_columns], which is faster than the code below
# (in benchmarks it seems to be even slightly faster than using the old astype). However, that is not
# supported by numpy 1.9.2.
if structured_arrays and (old_columns != new_columns):
new_arr = np.zeros(arr.shape, dtype)
for c in old_columns & new_columns:
new_arr[c] = arr[c]

# missing float columns should default to nan rather than zero
_is_float_type = lambda _dtype: _dtype.type in (np.float32, np.float64)
_is_void_float_type = lambda _dtype: _dtype.type == np.void and _is_float_type(_dtype.subdtype[0])
_is_float_or_void_float_type = lambda _dtype: _is_float_type(_dtype) or _is_void_float_type(_dtype)
_is_float = lambda column: _is_float_or_void_float_type(dtype.fields[column][0])
for new_column in filter(_is_float, new_columns - old_columns):
new_arr[new_column] = np.nan
else:
new_arr = arr.astype(dtype)

return new_arr


class KeyValueNdarrayStore(object):
"""Chunked store for arbitrary ndarrays, supporting append. Using an arbitrary kv store backend.

for the simple example:
dat = np.empty(10)
library.write('test', dat) #version 1
library.append('test', dat) #version 2

version documents:

[
{u'_id': ObjectId('55fa9a7781f12654382e58b8'),
u'symbol': u'test',
u'version': 1
u'type': u'ndarray',
u'up_to': 10, # no. of rows included in the data for this version
u'append_count': 0,
u'append_size': 0,
u'dtype': u'float64',
u'dtype_metadata': {},
u'segment_keys': [] # sha
u'sha': Binary('.........', 0),
u'shape': [-1],
},

{u'_id': ObjectId('55fa9aa981f12654382e58ba'),
u'symbol': u'test',
u'version': 2
u'type': u'ndarray',
u'up_to': 20, # no. of rows included in the data for this version
u'append_count': 1, # 1 append operation so far
u'append_size': 80, # 80 bytes appended
u'base_version_id': ObjectId('55fa9a7781f12654382e58b8'), # _id of version 1
u'dtype': u'float64',
u'dtype_metadata': {},
u'segment_count': 2, #2 segments included in this version
}
]


segment documents:

[
#first chunk written:
{u'_id': ObjectId('55fa9a778b376a68efdd10e3'),
u'compressed': True, #data is lz4 compressed on write()
u'data': Binary('...........', 0),
u'parent': [ObjectId('55fa9a7781f12654382e58b8')],
u'segment': 9, #10 rows in the data up to this segment, so last row is 9
u'sha': Binary('.............', 0), # checksum of (symbol, {'data':.., 'compressed':.., 'segment':...})
u'symbol': u'test'},

#second chunk appended:
{u'_id': ObjectId('55fa9aa98b376a68efdd10e6'),
u'compressed': False, # no initial compression for append()
u'data': Binary('...........', 0),
u'parent': [ObjectId('55fa9a7781f12654382e58b8')],
u'segment': 19, #20 rows in the data up to this segment, so last row is 19
u'sha': Binary('............', 0), # checksum of (symbol, {'data':.., 'compressed':.., 'segment':...})
u'symbol': u'test'},
]

"""
TYPE = 'ndarray'

@classmethod
def initialize_library(cls, *args, **kwargs):
pass

@staticmethod
def _ensure_index(collection):
pass

def can_delete(self, version, symbol):
return self.can_read(version, symbol)

def can_read(self, version, symbol):
return version['type'] == self.TYPE

def can_write(self, version, symbol, data):
return isinstance(data, np.ndarray) and not data.dtype.hasobject

def _dtype(self, string, metadata=None):
if metadata is None:
metadata = {}
if string.startswith('['):
return np.dtype(eval(string), metadata=metadata)
return np.dtype(string, metadata=metadata)

def _index_range(self, version, symbol, from_version=None, **kwargs):
"""
Tuple describing range to read from the ndarray - closed:open
"""
from_index = None
if from_version:
from_index = from_version['up_to']
return from_index, None

def get_info(self, version):
ret = {}
dtype = self._dtype(version['dtype'], version.get('dtype_metadata', {}))
length = int(version['up_to'])
ret['size'] = dtype.itemsize * length
ret['segment_count'] = version['segment_count']
ret['dtype'] = version['dtype']
ret['type'] = version['type']
ret['handler'] = self.__class__.__name__
ret['rows'] = int(version['up_to'])
return ret

def read(self, backing_store, library_name, version, symbol, **kwargs):
index_range = self._index_range(version, symbol, **kwargs)
return self._do_read(backing_store, library_name, version, symbol, index_range=index_range)

def _do_read(self, backing_store, library_name, version, symbol, index_range=None):
'''
index_range is a 2-tuple of integers - a [from, to) range of segments to be read.
Either from or to can be None, indicating no bound.
'''
from_index = index_range[0] if index_range else None
to_index = version['up_to']
if index_range and index_range[1] and index_range[1] < version['up_to']:
to_index = index_range[1]

segment_keys = version['segment_keys']
filtered_segment_keys = []
for i, segment_index in enumerate(version['raw_segment_index']):
if (from_index is None or segment_index >= from_index) and \
(to_index is None or segment_index <= to_index):
filtered_segment_keys.append(segment_keys[i])

data = bytearray()
for segment in backing_store.read_segments(library_name, filtered_segment_keys):
data.extend(decompress(segment))

dtype = self._dtype(version['dtype'], version.get('dtype_metadata', {}))
rtn = np.frombuffer(data, dtype=dtype).reshape(version.get('shape', (-1)))
return rtn

def _promote_types(self, dtype, dtype_str):
if dtype_str == str(dtype):
return dtype
prev_dtype = self._dtype(dtype_str)
if dtype.names is None:
rtn = np.promote_types(dtype, prev_dtype)
else:
rtn = _promote_struct_dtypes(dtype, prev_dtype)
rtn = np.dtype(rtn, metadata=dict(dtype.metadata or {}))
return rtn

def check_written(self, collection, symbol, version):
# Check all the chunks are in place
seen_chunks = collection.find({'symbol': symbol, 'parent': version['_id']},
).count()

if seen_chunks != version['segment_count']:
segments = [x['segment'] for x in collection.find({'symbol': symbol, 'parent': version['_id']},
projection={'segment': 1},
)]
raise ArcticException("Failed to write all the Chunks. Saw %s expecting %s"
"Parent: %s \n segments: %s" %
(seen_chunks, version['segment_count'], version['_id'], segments))

def checksum(self, item):
sha = hashlib.sha1()
sha.update(item.tostring())
return Binary(sha.digest())

def write(self, backing_store, library_name, version, symbol, item, previous_version, dtype=None):
if item.dtype.hasobject:
raise UnhandledDtypeException()

if not dtype:
dtype = item.dtype
version['dtype'] = str(dtype)
version['shape'] = (-1,) + item.shape[1:]
version['dtype_metadata'] = dict(dtype.metadata or {})
version['type'] = self.TYPE
version['up_to'] = len(item)
version['sha'] = self.checksum(item)

if previous_version:
if 'sha' in previous_version \
and previous_version['dtype'] == version['dtype'] \
and self.checksum(item[:previous_version['up_to']]) == previous_version['sha']:
# TODO handle appends!, currently segments will be reused to but all hashes will be recomputed
pass
# The first n rows are identical to the previous version, so just append.
# Do a 'dirty' append (i.e. concat & start from a new base version) for safety
# self._do_append(backing_store, collection, version, symbol, item[previous_version['up_to']:],
# previous_version, dirty_append=True)

version['base_sha'] = version['sha']
self._do_write(backing_store, library_name, version, symbol, item, previous_version)

def _do_write(self, backing_store, library_name, version, symbol, item, previous_version, segment_offset=0):

previous_segment_keys = []
if previous_version:
previous_segment_keys = previous_version['segment_keys']

if segment_offset > 0 and 'segment_index' in previous_version:
existing_index = previous_version['segment_index']
else:
existing_index = None

sze = int(item.dtype.itemsize * np.prod(item.shape[1:]))
length = len(item)

# chunk and store the data by (uncompressed) size
chunk_size = int(backing_store.chunk_size / sze)

# Compress
idxs = xrange(int(np.ceil(float(length) / chunk_size)))
chunks = [(item[i * chunk_size: (i + 1) * chunk_size]).tostring() for i in idxs]
compressed_segments = compress_array(chunks)

segment_keys = []
raw_segment_index = []
for i, segment_data in zip(idxs, compressed_segments):
segment_idx = min((i + 1) * chunk_size - 1, length - 1) + segment_offset
segment_key = backing_store.write_segment(library_name, symbol,
segment_data, previous_segment_keys)
raw_segment_index.append(segment_idx)
segment_keys.append(segment_key)

segment_index = self._segment_index(item, existing_index=existing_index, start=segment_offset,
new_segments=raw_segment_index)
if segment_index:
version['segment_index'] = segment_index
version['raw_segment_index'] = raw_segment_index
version['segment_count'] = len(segment_keys) # on appends this value is incorrect but is updated later on
version['append_size'] = 0
version['append_count'] = 0
version['segment_keys'] = segment_keys
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

aha, this is the forward-pointer implementation, where the version keeps all the segment keys, cool


#TODO add write check
#self.check_written(collection, symbol, version)

def _segment_index(self, new_data, existing_index, start, new_segments):
"""
Generate a segment index which can be used in subselect data in _index_range.
This function must handle both generation of the index and appending to an existing index

Parameters:
-----------
new_data: new data being written (or appended)
existing_index: index field from the versions document of the previous version
start: first (0-based) offset of the new data
segments: list of offsets. Each offset is the row index of the
the last row of a particular chunk relative to the start of the _original_ item.
array(new_data) - segments = array(offsets in item)

Returns:
--------
Library specific index metadata to be stored in the version document.
"""
pass # numpy arrays have no index
Loading