def encode(self, encoding):
return {
'expression': self.expressions[0],
'dtype': encoding.encode('dtype', self.dtype),
'dtype_item': encoding.encode('dtype', self.dtype_item),
'flatten': self.flatten,
'unique_limit': self.unique_limit,
'selection': self.selection
}
def encode(self, encoding):
return {
'expression': self.expressions[0],
'dtype': encoding.encode('dtype', self.dtype),
'dtype_item': encoding.encode('dtype', self.dtype_item),
'flatten': self.flatten,
'limit': self.limit,
'limit_raise': self.limit_raise,
'selection': self.selection,
'return_inverse': self.return_inverse
}
def encode(self, encoding):
return {
'expressions': self.expressions,
'shape': self.shape,
'selections': self.selections,
'op': encoding.encode('_op', self.op),
'weights': self.weights,
'dtype': encoding.encode('dtype', DataType(self.dtype)),
'minima': self.minima,
'maxima': self.maxima,
'edges': self.edges,
'selection_waslist': self.selection_waslist
}
def encode(self, encoding):
# TODO: get rid of dtypes
return {
'binners': encoding.encode_list('binner', self.binners),
'aggregation': encoding.encode("aggregation", self.aggregation_description),
'dtypes': encoding.encode_dict("dtype", self.dtypes)
}
def test_encoding():
encoding = vaex.encoding.Encoding()
data = encoding.encode('blobtest', {'someblob': b'1234'})
wiredata = vaex.encoding.serialize(data, encoding)
encoding = vaex.encoding.Encoding()
data = vaex.encoding.deserialize(wiredata, encoding)
values = encoding.decode('blobtest', data)
assert values['someblob'] == b'1234'
def rebuild_dataframe_vaex(df):
# encoding and decode
encoding = vaex.encoding.Encoding()
data = encoding.encode('dataframe', df)
blob = vaex.encoding.serialize(data, encoding)
encoding = vaex.encoding.Encoding()
data = vaex.encoding.deserialize(blob, encoding)
return encoding.decode('dataframe', data)
def rebuild_dataset_vaex(ds):
# encoding and decode
encoding = vaex.encoding.Encoding()
data = encoding.encode('dataset', ds)
blob = vaex.encoding.serialize(data, encoding)
encoding = vaex.encoding.Encoding()
data = vaex.encoding.deserialize(blob, encoding)
return encoding.decode('dataset', data)
def test_encoding_numpy():
x = np.arange(10, dtype='>f4')
encoding = vaex.encoding.Encoding()
data = encoding.encode('ndarray', x)
wiredata = vaex.encoding.serialize(data, encoding)
encoding = vaex.encoding.Encoding()
data = vaex.encoding.deserialize(wiredata, encoding)
value = encoding.decode('ndarray', data)
assert np.all(value == x)
def test_encoding_arrow(array_factory_arrow):
x = array_factory_arrow(np.arange(10, dtype='f4'))
encoding = vaex.encoding.Encoding()
data = encoding.encode('arrow-array', x)
wiredata = vaex.encoding.serialize(data, encoding)
encoding = vaex.encoding.Encoding()
data = vaex.encoding.deserialize(wiredata, encoding)
value = encoding.decode('arrow-array', data)
assert value.to_pylist() == x.to_pylist()
def encode(self, encoding):
# TODO: get rid of dtypes
encoded = {
'grid': encoding.encode('grid', self.grid),
'aggregations': encoding.encode_list("aggregation", self.aggregation_descriptions),
'dtypes': encoding.encode_dict("dtype", self.dtypes)
}
if self.has_values:
encoded['values'] = encoding.encode_list2('ndarray', self.get_values())
return encoded
def test_encoding_numpy_string_objects():
x = np.array(['vaex', 'is', None, 'fast'])
encoding = vaex.encoding.Encoding()
data = encoding.encode('ndarray', x)
wiredata = vaex.encoding.serialize(data, encoding)
encoding = vaex.encoding.Encoding()
data = vaex.encoding.deserialize(wiredata, encoding)
value = encoding.decode('ndarray', data)
assert np.all(value == x)
def test_encoding_numpy_datetime():
x = np.arange('2001', '2005', dtype='M')
encoding = vaex.encoding.Encoding()
data = encoding.encode('ndarray', x)
wiredata = vaex.encoding.serialize(data, encoding)
encoding = vaex.encoding.Encoding()
data = vaex.encoding.deserialize(wiredata, encoding)
value = encoding.decode('ndarray', data)
assert np.all(value == x)
def test_encoding_dtype():
dtype = np.dtype('>f8')
encoding = vaex.encoding.Encoding()
data = encoding.encode('dtype', dtype)
wiredata = vaex.encoding.serialize(data, encoding)
encoding = vaex.encoding.Encoding()
data = vaex.encoding.deserialize(wiredata, encoding)
print(data)
value = encoding.decode('dtype', data)
assert value == dtype
assert value.is_numpy
def encode(self, encoding):
# TODO: get rid of dtypes
return {
'task':
type(self).name,
'grid':
encoding.encode('grid', self.parent_grid),
'aggregations':
encoding.encode_list("aggregation", self.aggregation_descriptions),
'dtypes':
encoding.encode_dict("dtype", self.dtypes)
}
def test_encoding_numpy_masked():
x = np.arange(10, dtype='>f4')
mask = x > 4
x = np.ma.array(x, mask=mask)
encoding = vaex.encoding.Encoding()
data = encoding.encode('ndarray', x)
wiredata = vaex.encoding.serialize(data, encoding)
encoding = vaex.encoding.Encoding()
data = vaex.encoding.deserialize(wiredata, encoding)
value = encoding.decode('ndarray', data)
assert np.all(value == x)
assert np.all(value.mask == x.mask)
def rebuild_with_skip(ds, skip):
repr(ds) # for coverage
# encoding and decode
encoding = vaex.encoding.Encoding()
encoding.set_object_spec(skip.id, None) # this will cause it to skip serialization
data = encoding.encode('dataset', ds)
assert encoding._object_specs[skip.id] is None
del encoding._object_specs[skip.id]
blob = vaex.encoding.serialize(data, encoding)
encoding = vaex.encoding.Encoding()
encoding.set_object(skip.id, skip)
data = vaex.encoding.deserialize(blob, encoding)
return encoding.decode('dataset', data)
async def execute_async(self):
logger.debug("starting with execute")
with self.lock: # setup thread local initial values
if not hasattr(self.local, 'executing'):
self.local.executing = False
# wo don't allow any thread from our thread pool to enter (a computation should never produce a new task)
# and we explicitly disallow reentry (this usually means a bug in vaex, or bad usage)
chunk_executor_thread = threading.current_thread(
) in self.thread_pool._threads
import traceback
trace = ''.join(traceback.format_stack())
if chunk_executor_thread or self.local.executing:
logger.error("nested execute call")
raise RuntimeError(
"nested execute call: %r %r\nlast trace:\n%s\ncurrent trace:\n%s"
% (chunk_executor_thread, self.local.executing,
self.local.last_trace, trace))
else:
self.local.last_trace = trace
self.local.executing = True
try:
t0 = time.time()
self.local.cancelled = False
self.signal_begin.emit()
cancelled = False
# keep getting a list of tasks
# we currently process tasks (grouped) per df
# but also, tasks can add new tasks
while not cancelled:
tasks = self.local.tasks = self._pop_tasks()
if not tasks:
break
tasks = _merge(tasks, tasks[0].df)
run = Run(tasks)
self.passes += 1
# (re) thrown exceptions as soon as possible to avoid complicated stack traces
for task in tasks:
if task.isRejected:
task.get()
if hasattr(task, "check"):
try:
task.check()
except Exception as e:
task.reject(e)
raise
for task in run.tasks:
task._results = []
if not any(task.signal_progress.emit(0)):
logger.debug("task cancelled immediately")
task.cancelled = True
row_count = run.df._index_end - run.df._index_start
chunk_size = self.chunk_size_for(row_count)
run.block_scopes = [
run.df._block_scope(0, chunk_size)
for i in range(self.thread_pool.nthreads)
]
encoding = vaex.encoding.Encoding()
for task in tasks:
spec = encoding.encode('task', task)
spec['task-part-cpu-type'] = spec.pop('task-type')
def create_task_part():
return encoding.decode('task-part-cpu',
spec,
df=run.df)
# We want at least 1 task part (otherwise we cannot do any work)
# then we ask for the task part how often we should split
# This means that we can have 100 threads, but only 2 task parts
# In this case, evaluation of expressions is still multithreaded,
# but aggregation is reduced to effectively 2 threads.
task_part_0 = create_task_part()
ideal_task_splits = task_part_0.ideal_splits(
self.thread_pool.nthreads)
assert ideal_task_splits <= self.thread_pool.nthreads, f'Cannot have more splits {ideal_task_splits} then threads {self.thread_pool.nthreads}'
if ideal_task_splits == self.thread_pool.nthreads:
# in the simple case, we just use a list
task._parts = [task_part_0] + [
create_task_part()
for i in range(1, self.thread_pool.nthreads)
]
else:
# otherwise a queue
task._parts = queue.Queue()
task._parts.put(task_part_0)
for i in range(1, ideal_task_splits):
task._parts.put(create_task_part())
length = run.df.active_length()
if vaex.cache.is_on():
key_df = run.df.fingerprint()
# TODO: in the future we might want to enable the zigzagging again, but this requires all datasets to implement it
# if self.zigzag:
# self.zig = not self.zig
dataset = run.df.dataset[run.df._index_start:run.df._index_end]
# find the columns from the dataset we need
variables = set()
for expression in run.expressions:
variables |= run.df._expr(expression).expand().variables(
ourself=True)
columns = list(variables - set(run.df.variables) -
set(run.df.virtual_columns))
logger.debug('Using columns %r from dataset, chunk_size=%r',
columns, chunk_size)
for column in columns:
if column not in dataset:
raise RuntimeError(
f'Oops, requesting column {column} from dataset, but it does not exist'
)
async for element in self.thread_pool.map_async(
self.process_part,
dataset.chunk_iterator(columns, chunk_size),
dataset.row_count,
progress=lambda p: all(self.signal_progress.emit(p))
and all([
all(task.signal_progress.emit(p)) for task in tasks
]) and all([not task.cancelled for task in tasks]),
cancel=lambda: self._cancel(run),
unpack=True,
run=run):
pass # just eat all element
duration_wallclock = time.time() - t0
logger.debug("executing took %r seconds", duration_wallclock)
cancelled = self.local.cancelled or any(
task.cancelled for task in tasks) or run.cancelled
logger.debug("cancelled: %r", cancelled)
if cancelled:
logger.debug("execution aborted")
for task in tasks:
task.reject(UserAbort("cancelled"))
# remove references
task._result = None
task._results = None
cancelled = True
if isinstance(task, vaex.tasks.TaskAggregations):
for subtask in task.original_tasks:
subtask.reject(UserAbort("cancelled"))
else:
for task in tasks:
logger.debug("fulfill task: %r", task)
if not task.cancelled:
parts = task._parts
if not isinstance(parts, list):
parts_queue = parts
parts = []
while not parts_queue.empty():
parts.append(parts_queue.get())
parts[0].reduce(parts[1:])
logger.debug("wait for task: %r", task)
task._result = parts[0].get_result()
task.end()
task.fulfill(task._result)
logger.debug("got result for: %r", task)
if task._result is not None and task.cacheable: # we don't want to store None
if vaex.cache.is_on():
# we only want to store the original task results into the cache
tasks_cachable = task.original_tasks if isinstance(
task,
vaex.tasks.TaskAggregations) else [
task
]
for task_cachable in tasks_cachable:
key_task = task_cachable.fingerprint()
# tasks' fingerprints don't include the dataframe
key = f'{key_task}-{key_df}'
previous_result = vaex.cache.get(
key, type='task')
if (
previous_result is not None
): # and (previous_result != task_cachable.get()):
try:
if previous_result != task_cachable.get(
):
# this can happen with multithreading, where two threads enter the same tasks in parallel (IF using different executors)
logger.warning(
"calculated new result: %r, while cache had value: %r",
previous_result,
task_cachable.get())
except ValueError: # when comparing numpy results
if np.array_equal(
previous_result,
task_cachable.get(),
equal_nan=True):
# this can happen with multithreading, where two threads enter the same tasks in parallel (IF using different executors)
logger.warning(
"calculated new result: %r, while cache had value: %r",
previous_result,
task_cachable.get())
vaex.cache.set(key,
task_cachable.get(),
type='task',
duration_wallclock=
duration_wallclock)
logger.info(
"added result: %r in cache under key: %r",
task_cachable.get(), key)
else:
task.reject(UserAbort("Task was cancelled"))
# remove references
cancelled = True
task._result = None
task._results = None
self.signal_end.emit()
except: # noqa
self.signal_cancel.emit()
logger.exception("error in task, flush task queue")
raise
finally:
self.local.executing = False
def execute_generator(self, use_async=False):
logger.debug("starting with execute")
with self.lock: # setup thread local initial values
if not hasattr(self.local, 'executing'):
self.local.executing = False
try:
t0 = time.time()
self.local.cancelled = False
self.signal_begin.emit()
# keep getting a list of tasks
# we currently process tasks (grouped) per df
# but also, tasks can add new tasks
while True:
tasks = self.local.tasks = self._pop_tasks()
# wo don't allow any thread from our thread pool to enter (a computation should never produce a new task)
# and we explicitly disallow reentry (this usually means a bug in vaex, or bad usage)
chunk_executor_thread = threading.current_thread(
) in self.thread_pool._threads
import traceback
trace = ''.join(traceback.format_stack())
if chunk_executor_thread or self.local.executing and (
has_contextvars is False
or self.isnested.get() is True):
logger.error("nested execute call")
raise RuntimeError(
"nested execute call: %r %r\nlast trace:\n%s\ncurrent trace:\n%s"
% (chunk_executor_thread, self.local.executing,
self.local.last_trace, trace))
else:
self.local.last_trace = trace
self.local.executing = True
if has_contextvars:
self.isnested.set(True)
if not tasks:
break
tasks = _merge(tasks)
run = Run(tasks)
self.passes += 1
dataset = run.dataset
run.variables = {}
for df in run.tasks_per_df.keys():
run.variables[df] = {
key: df.evaluate_variable(key)
for key in df.variables.keys()
}
# (re) thrown exceptions as soon as possible to avoid complicated stack traces
for task in tasks:
if task.isRejected:
task.get()
if hasattr(task, "check"):
try:
task.check()
except Exception as e:
task.reject(e)
raise
for task in run.tasks:
task.signal_start.emit(self)
for task in run.tasks:
task._results = []
if not any(task.signal_progress.emit(0)):
logger.debug("task cancelled immediately")
task.cancelled = True
row_count = dataset.row_count
chunk_size = self.chunk_size_for(row_count)
encoding = vaex.encoding.Encoding()
run.nthreads = nthreads = self.thread_pool.nthreads
task_checkers = vaex.tasks.create_checkers()
memory_tracker = vaex.memory.create_tracker()
vaex.memory.local.agg = memory_tracker
# we track this for consistency
memory_usage = 0
for task in tasks:
for task_checker in task_checkers:
task_checker.add_task(task)
spec = encoding.encode('task', task)
spec['task-part-cpu-type'] = spec.pop('task-type')
def create_task_part():
nonlocal memory_usage
task_part = encoding.decode('task-part-cpu',
spec,
df=task.df,
nthreads=nthreads)
memory_usage += task_part.memory_usage()
for task_checker in task_checkers:
task_checker.add_task(task)
if task.requires_fingerprint:
task_part.fingerprint = task.fingerprint()
return task_part
# We want at least 1 task part (otherwise we cannot do any work)
# then we ask for the task part how often we should split
# This means that we can have 100 threads, but only 2 task parts
# In this case, evaluation of expressions is still multithreaded,
# but aggregation is reduced to effectively 2 threads.
task_part_0 = create_task_part()
ideal_task_splits = task_part_0.ideal_splits(
self.thread_pool.nthreads)
assert ideal_task_splits <= self.thread_pool.nthreads, f'Cannot have more splits {ideal_task_splits} then threads {self.thread_pool.nthreads}'
if ideal_task_splits == self.thread_pool.nthreads or task.see_all:
# in the simple case, we just use a list
task._parts = [task_part_0] + [
create_task_part()
for i in range(1, ideal_task_splits)
]
else:
# otherwise a queue
task._parts = queue.Queue()
task._parts.put(task_part_0)
for i in range(1, ideal_task_splits):
task._parts.put(create_task_part())
if memory_usage != memory_tracker.used:
raise RuntimeError(
f"Reported memory usage by tasks was {memory_usage}, while tracker listed {memory_tracker.used}"
)
vaex.memory.local.agg = None
# TODO: in the future we might want to enable the zigzagging again, but this requires all datasets to implement it
# if self.zigzag:
# self.zig = not self.zig
def progress(p):
# no global cancel and at least 1 tasks wants to continue, then we continue
ok_tasks = any([task.progress(p) for task in tasks])
all_stopped = all([task.stopped for task in tasks])
ok_executor = all(self.signal_progress.emit(p))
if all_stopped:
logger.debug(
"Pass cancelled because all tasks are stopped: %r",
tasks)
if not ok_tasks:
logger.debug(
"Pass cancelled because all tasks cancelled: %r",
tasks)
if not ok_executor:
logger.debug(
"Pass cancelled because of the global progress event: %r",
self.signal_progress.callbacks)
return ok_tasks and ok_executor and not all_stopped
yield from self.thread_pool.map(
self.process_part,
dataset.chunk_iterator(run.dataset_deps, chunk_size),
dataset.row_count,
progress=progress,
cancel=lambda: self._cancel(run),
unpack=True,
run=run,
use_async=use_async)
duration_wallclock = time.time() - t0
logger.debug("executing took %r seconds", duration_wallclock)
self.local.executing = False
if has_contextvars:
self.isnested.set(False)
if True: # kept to keep the diff small
for task in tasks:
if not task.cancelled:
logger.debug("fulfill task: %r", task)
parts = task._parts
if not isinstance(parts, list):
parts_queue = parts
parts = []
while not parts_queue.empty():
parts.append(parts_queue.get())
parts[0].reduce(parts[1:])
logger.debug("wait for task: %r", task)
task._result = parts[0].get_result()
task.end()
task.fulfill(task._result)
logger.debug("got result for: %r", task)
if task._result is not None and task.cacheable: # we don't want to store None
if vaex.cache.is_on():
# we only want to store the original task results into the cache
tasks_cachable = task.original_tasks if isinstance(
task,
vaex.tasks.TaskAggregations) else [
task
]
for task_cachable in tasks_cachable:
key = task_cachable.fingerprint()
previous_result = vaex.cache.get(
key, type='task')
if (
previous_result is not None
): # and (previous_result != task_cachable.get()):
try:
if previous_result != task_cachable.get(
):
# this can happen with multithreading, where two threads enter the same tasks in parallel (IF using different executors)
logger.warning(
"calculated new result: %r, while cache had value: %r",
previous_result,
task_cachable.get())
except ValueError: # when comparing numpy results
if not np.array_equal(
previous_result,
task_cachable.get(),
equal_nan=True):
# this can happen with multithreading, where two threads enter the same tasks in parallel (IF using different executors)
logger.warning(
"calculated new result: %r, while cache had value: %r",
previous_result,
task_cachable.get())
vaex.cache.set(key,
task_cachable.get(),
type='task',
duration_wallclock=
duration_wallclock)
logger.info(
"added result: %r in cache under key: %r",
task_cachable.get(), key)
else:
logger.debug("rejecting task: %r", task)
# we now reject, in the main thread
if task._toreject:
task.reject(task._toreject)
else:
task.reject(UserAbort("Task was cancelled"))
# remove references
task._result = None
task._results = None
self.signal_end.emit()
except: # noqa
self.signal_cancel.emit()
raise
finally:
self.local.executing = False
if has_contextvars:
self.isnested.set(False)
