def _read_side_inputs(self, tags_and_types):
# type: (...) -> Iterator[apache_sideinputs.SideInputMap]
"""Generator reading side inputs in the order prescribed by tags_and_types.
Args:
tags_and_types: List of tuples (tag, type). Each side input has a string
tag that is specified in the worker instruction. The type is actually
a boolean which is True for singleton input (read just first value)
and False for collection input (read all values).
Yields:
With each iteration it yields the result of reading an entire side source
either in singleton or collection mode according to the tags_and_types
argument.
"""
# Only call this on the old path where side_input_maps was not
# provided directly.
assert self.side_input_maps is None
# We will read the side inputs in the order prescribed by the
# tags_and_types argument because this is exactly the order needed to
# replace the ArgumentPlaceholder objects in the args/kwargs of the DoFn
# getting the side inputs.
#
# Note that for each tag there could be several read operations in the
# specification. This can happen for instance if the source has been
# sharded into several files.
for i, (side_tag, view_class, view_options) in enumerate(tags_and_types):
sources = []
# Using the side_tag in the lambda below will trigger a pylint warning.
# However in this case it is fine because the lambda is used right away
# while the variable has the value assigned by the current iteration of
# the for loop.
# pylint: disable=cell-var-from-loop
for si in filter(lambda o: o.tag == side_tag, self.spec.side_inputs):
if not isinstance(si, operation_specs.WorkerSideInputSource):
raise NotImplementedError('Unknown side input type: %r' % si)
sources.append(si.source)
# The tracking of time spend reading and bytes read from side inputs is
# behind an experiment flag to test its performance impact.
si_counter = opcounters.SideInputReadCounter(
self.counter_factory,
self.state_sampler,
declaring_step=self.name_context.step_name,
# Inputs are 1-indexed, so we add 1 to i in the side input id
input_index=i + 1)
iterator_fn = sideinputs.get_iterator_fn_for_sources(
sources, read_counter=si_counter)
# Backwards compatibility for pre BEAM-733 SDKs.
if isinstance(view_options, tuple):
if view_class == pvalue.AsSingleton:
has_default, default = view_options
view_options = {'default': default} if has_default else {}
else:
view_options = {}
yield apache_sideinputs.SideInputMap(
view_class, view_options, sideinputs.EmulatedIterable(iterator_fn))
def run_benchmark(num_runs=50, input_per_source=4000, num_sources=4):
print("Number of runs:", num_runs)
print("Input size:", num_sources * input_per_source)
print("Sources:", num_sources)
times = []
for i in range(num_runs):
counter_factory = CounterFactory()
state_sampler = statesampler.StateSampler('basic', counter_factory)
with state_sampler.scoped_state('step1', 'state'):
si_counter = opcounters.SideInputReadCounter(
counter_factory, state_sampler, 'step1', 1)
si_counter = opcounters.NoOpTransformIOCounter()
sources = [
FakeSource(long_generator(i, input_per_source))
for i in range(num_sources)]
iterator_fn = sideinputs.get_iterator_fn_for_sources(
sources, read_counter=si_counter)
start = time.time()
list(iterator_fn())
time_cost = time.time() - start
times.append(time_cost)
print("Runtimes:", times)
avg_runtime = sum(times) // len(times)
print("Average runtime:", avg_runtime)
print("Time per element:", avg_runtime // (input_per_source *
num_sources))
def test_basic_counters(self):
counter_factory = CounterFactory()
sampler = statesampler.StateSampler('stage1', counter_factory)
sampler.start()
with sampler.scoped_state('step1', 'stateA'):
counter = opcounters.SideInputReadCounter(counter_factory,
sampler,
declaring_step='step1',
input_index=1)
with sampler.scoped_state('step2', 'stateB'):
with counter:
counter.add_bytes_read(10)
counter.update_current_step()
sampler.stop()
sampler.commit_counters()
actual_counter_names = set(
[c.name for c in counter_factory.get_counters()])
expected_counter_names = set([
# Counter names for STEP 1
counters.CounterName('read-sideinput-msecs',
stage_name='stage1',
step_name='step1',
io_target=counters.side_input_id('step1', 1)),
counters.CounterName('read-sideinput-byte-count',
step_name='step1',
io_target=counters.side_input_id('step1', 1)),
# Counter names for STEP 2
counters.CounterName('read-sideinput-msecs',
stage_name='stage1',
step_name='step1',
io_target=counters.side_input_id('step2', 1)),
counters.CounterName('read-sideinput-byte-count',
step_name='step1',
io_target=counters.side_input_id('step2', 1)),
])
self.assertTrue(
actual_counter_names.issuperset(expected_counter_names))
