def __init__(
self,
identifier = None, #name under which the cache file is stored. defaults to modulename.objname
environment = None, #object containing information regarding the dependencies on global state of the cached operation
operation = None, #function to be cached. note that the order of arguments is significant for key reuse
hierarchy = None, #key hierarchy; if and how args and kwargs are hierarchically ordered
validate = False, #validation mode. if enabled, all cache retrievals are checked against a recomputed function call.
deferred_timeout = 30, #time to wait before a deferred object is considered obsolete. compilation may take a long time; that said, it may also crash your process...
lock_timeout = 1, #time to wait before a lock is considered obsolete. the lock is needed for pure db transactions only; this makes once second a long time
environment_clear = True, #clear the cache upon connection with a novel environment key
connect_clear = False #clear the cache upon every connection
):
"""
if environment_clear is set to true, the cache is cleared
"""
if identifier: self.identifier = identifier
if operation: self.operation = operation
self.hierarchy = hierarchy
#add some essentials to the environment
import platform
globalenv = platform.architecture(), platform.python_version()
funcenv = inspect.getargspec(self.operation), inspect.getsource(self.operation)
self.environment = globalenv, funcenv, (environment if environment else self.environment())
estr, ehash = process_key(self.environment)
self.validate = validate
self.lock_timeout = lock_timeout
self.deferred_timeout = deferred_timeout
self.filename = os.path.join(cachepath, self.identifier)
self.shelve = Shelve(self.filename, autocommit = True)
self.lock = threading.Lock()
self.lock_file = lockfile.MkdirLockFile(self.filename, timeout = lock_timeout)
with self.lock, self.lock_file:
if connect_clear:
#this isnt right; we are now invalidating the precomputed envrowid of other processes...
self.shelve.clear() #need write lock here
#write environment key to database and obtain its unique rowid
try:
self.envrowid = self.shelve.getrowid(self.environment, estr, ehash)
except:
#connect to the db with a novel environment; probably wont change back again
if environment_clear:
self.shelve.clear() #need write lock here
self.shelve.setitem(self.environment, None, estr, ehash)
self.envrowid = self.shelve.getrowid(self.environment, estr, ehash)
class AbstractCache(object):
"""
abstract base class of a cache object which gracefully handles large arbitrary key objects
"""
def __init__(
self,
identifier = None, #name under which the cache file is stored. defaults to modulename.objname
environment = None, #object containing information regarding the dependencies on global state of the cached operation
operation = None, #function to be cached. note that the order of arguments is significant for key reuse
hierarchy = None, #key hierarchy; if and how args and kwargs are hierarchically ordered
validate = False, #validation mode. if enabled, all cache retrievals are checked against a recomputed function call.
deferred_timeout = 30, #time to wait before a deferred object is considered obsolete. compilation may take a long time; that said, it may also crash your process...
lock_timeout = 1, #time to wait before a lock is considered obsolete. the lock is needed for pure db transactions only; this makes once second a long time
environment_clear = True, #clear the cache upon connection with a novel environment key
connect_clear = False #clear the cache upon every connection
):
"""
if environment_clear is set to true, the cache is cleared
"""
if identifier: self.identifier = identifier
if operation: self.operation = operation
self.hierarchy = hierarchy
#add some essentials to the environment
import platform
globalenv = platform.architecture(), platform.python_version()
funcenv = inspect.getargspec(self.operation), inspect.getsource(self.operation)
self.environment = globalenv, funcenv, (environment if environment else self.environment())
estr, ehash = process_key(self.environment)
self.validate = validate
self.lock_timeout = lock_timeout
self.deferred_timeout = deferred_timeout
self.filename = os.path.join(cachepath, self.identifier)
self.shelve = Shelve(self.filename, autocommit = True)
self.lock = threading.Lock()
self.lock_file = lockfile.MkdirLockFile(self.filename, timeout = lock_timeout)
with self.lock, self.lock_file:
if connect_clear:
#this isnt right; we are now invalidating the precomputed envrowid of other processes...
self.shelve.clear() #need write lock here
#write environment key to database and obtain its unique rowid
try:
self.envrowid = self.shelve.getrowid(self.environment, estr, ehash)
except:
#connect to the db with a novel environment; probably wont change back again
if environment_clear:
self.shelve.clear() #need write lock here
self.shelve.setitem(self.environment, None, estr, ehash)
self.envrowid = self.shelve.getrowid(self.environment, estr, ehash)
def __call__(self, *args, **kwargs):
"""
look up a hierachical key object
fill in the missing parts, and perform the computation at the leaf if so required
"""
if self.hierarchy:
#apply the structure in hierarchy to the arguments
fkey = kwargs.copy()
fkey.update(enumerate(args))
hkey = [[fkey.pop(a) for a in level] for level in self.hierarchy]
if fkey: hkey.append(fkey) #any arguments not part of the hierarchy spec are placed at the end
else:
hkey = [args + ((kwargs,) if kwargs else ())] #put all args in a single key
#preprocess subkeys. this minimizes time spent in locked state
hkey = map(as_deterministic, hkey)
with self.lock: #fairly stupid thread locking. dont use threads; how about that?
while True:
try:
with self.lock_file:
#hierarchical key lookup; first key is prebound environment key
previouskey = Partial(self.envrowid)
for ikey, subkey in enumerate(hkey[:-1]):
partialkey = previouskey, subkey
rowid = self.shelve.getrowid(partialkey, *process_key(partialkey)) #read lock?
previouskey = Partial(rowid)
#leaf iteration
ikey = len(hkey)-1
leafkey = previouskey, hkey[-1]
value = self.shelve[leafkey] #read lock?
if isinstance(value, Deferred):
if value.expired(self.deferred_timeout):
raise Exception()
sleep(0.01)
else:
if self.validate:
#check if recomputed value is identical under deterministic serialization
newvalue = self.operation(*args, **kwargs)
try:
#note; new may differ from old in case aliasing in an ndarray was erased
#by original serialization. is this an error?
#id say so; depending on wether we have a cache hit, downstream code may react diffently
#perhaps its best to use custom serializating for values too
assert(as_deterministic(value)==as_deterministic(newvalue))
except:
print 'Cache returned invalid value!'
print 'arguments:'
print args
print kwargs
print 'cached value'
print value
print 'recomputed value'
print newvalue
quit()
#yes! hitting this return is what we are doing this all for!
return value
except:
#lock for the writing branch. multiprocess does not benefit here, but so be it.
#worst case we make multiple insertions into db, but this should do no harm for behavior
if self.lock_file.is_locked():
#if lock not available, better to go back to waiting for a deferred to appear
sleep(0.001)
else:
with self.lock_file:
#hierarchical key insertion
for subkey in hkey[ikey:-1]:
partialkey = previouskey, subkey
kstr, khash = process_key(partialkey)
self.shelve.setitem(partialkey, None, kstr, khash) #wite lock
rowid = self.shelve.getrowid(partialkey, kstr, khash) #read lock
previouskey = Partial(rowid)
#insert leaf node
leafkey = previouskey, hkey[-1]
kstr, khash = process_key(leafkey)
self.shelve.setitem(leafkey, Deferred(), kstr, khash) #write lock
#dont need lock while doing expensive things
value = self.operation(*args, **kwargs)
with self.lock_file:
self.shelve.setitem(leafkey, value , kstr, khash) #write lock
return value
def operation(self, input):
"""
implements the cached operation; to be invoked upon a cache miss
input is a picklable python object
the returned output should be a pickalble python object as well
"""
raise NotImplementedError()
def environment(self):
"""
returns a pickeable object describing the environment of the cached operation,
or a description of the state of your computer which may influence the relation
between the input and output of Cache.cached
"""
raise NotImplementedError()
