def test_fit_model(self):
if skip_lightning_tests:
self.skipTest(
'Spark PyTorch Lightning tests conflict with Tensorflow 2.5.x: '
'https://github.com/horovod/horovod/pull/3263')
model = create_xor_model()
with spark_session('test_fit_model') as spark:
df = create_noisy_xor_data(spark)
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
num_proc=2,
store=store,
model=model,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
validation=0.2,
batch_size=4,
epochs=2,
random_seed=1,
verbose=2)
torch_model = torch_estimator.fit(df)
trained_model = torch_model.getModel()
pred = trained_model(torch.ones([1, 2], dtype=torch.int32))
assert len(pred) == 1
assert pred.dtype == torch.float32
def test_fit_model(self):
model = create_xor_model()
optimizer = torch.optim.SGD(model.parameters(), lr=0.1)
loss = F.binary_cross_entropy
with spark_session('test_fit_model') as spark:
df = create_xor_data(spark)
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
num_proc=2,
store=store,
model=model,
optimizer=optimizer,
loss=loss,
input_shapes=[[2]],
feature_cols=['features'],
label_cols=['y'],
batch_size=1,
epochs=3,
random_seed=1,
verbose=2,
sample_weight_col='weight')
torch_model = torch_estimator.fit(df)
trained_model = torch_model.getModel()
pred = trained_model(torch.ones([1, 2], dtype=torch.int32))
assert len(pred) == 1
assert pred.dtype == torch.float32
def test_legacy_fit_model(self):
if skip_lightning_tests:
self.skipTest(
'Spark PyTorch Lightning tests conflict with Tensorflow 2.5.x: '
'https://github.com/horovod/horovod/pull/3263')
model = create_legacy_xor_model()
optimizer = torch.optim.SGD(model.parameters(), lr=0.1)
loss = F.binary_cross_entropy
with spark_session('test_fit_model') as spark:
df = create_noisy_xor_data(spark)
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
num_proc=2,
store=store,
model=model,
optimizer=optimizer,
loss=loss,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
batch_size=4,
epochs=2,
verbose=2,
sample_weight_col='weight')
torch_model = torch_estimator.fit(df)
trained_model = torch_model.getModel()
pred = trained_model(torch.ones([1, 2], dtype=torch.int32))
assert len(pred) == 1
assert pred.dtype == torch.float32
def test_fit_model(self):
if sys.version_info < (3, 0, 0) and is_gloo_used():
self.skipTest(
'Horovod on Spark over Gloo only supported on Python3')
model = create_xor_model()
optimizer = torch.optim.SGD(model.parameters(), lr=0.1)
loss = F.binary_cross_entropy
with spark_session('test_fit_model') as spark:
df = create_xor_data(spark)
with local_store() as store:
torch_estimator = hvd.TorchEstimator(
num_proc=2,
store=store,
model=model,
optimizer=optimizer,
loss=loss,
input_shapes=[[2]],
feature_cols=['features'],
label_cols=['y'],
batch_size=1,
epochs=3,
verbose=2,
sample_weight_col='weight')
torch_model = torch_estimator.fit(df)
trained_model = torch_model.getModel()
pred = trained_model(torch.ones([1, 2], dtype=torch.int32))
assert len(pred) == 1
assert pred.dtype == torch.float32
def test_torch_direct_parquet_train(self):
with spark_session('test_torch_direct_parquet_train') as spark:
df = create_xor_data(spark)
backend = CallbackBackend()
with local_store() as store:
store.get_train_data_path = lambda v=None: store._train_path
store.get_val_data_path = lambda v=None: store._val_path
with util.prepare_data(backend.num_processes(),
store,
df,
feature_columns=['features'],
label_columns=['y']):
model = create_xor_model()
optimizer = torch.optim.SGD(model.parameters(), lr=0.1)
loss = nn.BCELoss()
est = hvd_spark.TorchEstimator(backend=backend,
store=store,
model=model,
optimizer=optimizer,
input_shapes=[[2]],
feature_cols=['features'],
label_cols=['y'],
batch_size=1,
epochs=3,
verbose=2)
# To make sure that setLoss works with non-list loss.
est.setLoss(loss)
transformer = est.fit_on_parquet()
predictions = transformer.transform(df)
assert predictions.count() == df.count()
def test_transform_multi_class(self):
model = create_xor_model(output_dim=2)
with spark_session('test_transform_multi_class') as spark:
df = create_xor_data(spark)
metadata = util._get_metadata(df)
torch_model = hvd_spark.TorchModel(history=None,
model=model,
input_shapes=[[2]],
feature_columns=['features'],
label_columns=['y'],
_metadata=metadata)
out_df = torch_model.transform(df)
expected_types = {
'x1': LongType,
'x2': LongType,
'features': VectorUDT,
'weight': DoubleType,
'y': DoubleType,
'y__output': VectorUDT
}
for field in out_df.schema.fields:
assert type(field.dataType) == expected_types[field.name]
def test_model_override_trainer_args(self):
if skip_lightning_tests:
self.skipTest(
'Spark PyTorch Lightning tests conflict with Tensorflow 2.5.x: '
'https://github.com/horovod/horovod/pull/3263')
from pytorch_lightning.callbacks.model_checkpoint import ModelCheckpoint
with spark_session('test_fit_model') as spark:
df = create_noisy_xor_data(spark)
model = create_xor_model()
with tempdir() as dir:
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
num_proc=2,
store=store,
model=model,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
validation=0.2,
batch_size=4,
epochs=2,
verbose=2,
trainer_args={'stochastic_weight_avg': True})
torch_model = torch_estimator.fit(df)
# TODO: Find a way to pass log metrics from remote, and assert base on the logger.
trained_model = torch_model.getModel()
pred = trained_model(torch.ones([1, 2], dtype=torch.int32))
assert len(pred) == 1
assert pred.dtype == torch.float32
def test_restore_from_checkpoint(self):
model = create_xor_model()
with spark_session('test_restore_from_checkpoint') as spark:
df = create_noisy_xor_data(spark)
ctx = CallbackBackend()
run_id = 'run01'
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
backend=ctx,
store=store,
model=model,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
validation=0.2,
batch_size=4,
epochs=2,
verbose=2,
run_id=run_id)
torch_estimator._read_checkpoint = Mock(
side_effect=torch_estimator._read_checkpoint)
ckpt_path = store.get_checkpoint_path(run_id)
assert not store.exists(ckpt_path)
torch_estimator._read_checkpoint.assert_not_called()
torch_estimator.fit(df)
assert store.exists(ckpt_path)
torch_estimator.fit(df)
torch_estimator._read_checkpoint.assert_called()
def test_fit_model(self):
model = create_xor_model()
optimizer = tf.keras.optimizers.SGD(lr=0.1)
loss = 'binary_crossentropy'
with spark_session('test_fit_model') as spark:
df = create_xor_data(spark)
with local_store() as store:
keras_estimator = hvd.KerasEstimator(
num_proc=2,
store=store,
model=model,
optimizer=optimizer,
loss=loss,
feature_cols=['features'],
label_cols=['y'],
batch_size=1,
epochs=3,
verbose=2)
keras_model = keras_estimator.fit(df)
trained_model = keras_model.getModel()
pred = trained_model.predict([np.ones([1, 2], dtype=np.float32)])
assert len(pred) == 1
assert pred.dtype == np.float32
def test_happy_run_elastic_fault_tolerant_fails(self):
self.skipTest(
'elastic horovod does not support shutdown from the spark driver '
'while elastic driver is waiting for hosts to come up')
if not gloo_built():
self.skipTest("Gloo is not available")
with spark_session('test_happy_run_elastic_fault_tolerant_fails',
max_failures=2):
with tempdir() as dir:
# these files make training function fail in given rank, epoch and batch
# we have as many failures as Spark has max_failures (per task / index)
with open(os.path.sep.join([dir, 'rank_1_epoch_2_batch_4_fail']), 'w'), \
open(os.path.sep.join([dir, 'rank_1_epoch_3_batch_1_fail']), 'w'):
pass
res = horovod.spark.run_elastic(
fn,
args=(2, 5, 5, dir),
env={'HOROVOD_LOG_LEVEL': 'DEBUG'},
num_proc=2,
min_num_proc=2,
max_num_proc=2,
start_timeout=5,
verbose=2)
self.assertListEqual([([0, 4, 0, 4, 1, 4, 0, 4], 0),
([0, 4, 0, 4, 1, 4, 0, 4], 1)], res)
def test_train_with_inmemory_cache_all(self):
if skip_lightning_tests:
self.skipTest(
'Spark PyTorch Lightning tests conflict with Tensorflow 2.5.x: '
'https://github.com/horovod/horovod/pull/3263')
with spark_session('test_fit_model') as spark:
df = create_noisy_xor_data(spark)
model = create_xor_model()
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
num_proc=
1, # Normally inmem dataloader is for single worker training with small data
store=store,
model=model,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
validation=0.2,
batch_size=4,
epochs=2,
verbose=2,
inmemory_cache_all=True)
torch_model = torch_estimator.fit(df)
# TODO: Find a way to pass log metrics from remote, and assert base on the logger.
trained_model = torch_model.getModel()
pred = trained_model(torch.ones([1, 2], dtype=torch.int32))
assert len(pred) == 1
assert pred.dtype == torch.float32
def test_fit_model_multiclass(self):
model = create_mnist_model()
optimizer = tf.keras.optimizers.Adadelta(1.0)
loss = tf.keras.losses.categorical_crossentropy
for num_cores in [2, constants.TOTAL_BUFFER_MEMORY_CAP_GIB + 1]:
with spark_session('test_fit_model_multiclass', cores=num_cores) as spark:
df = create_mnist_data(spark)
with local_store() as store:
keras_estimator = hvd.KerasEstimator(
num_proc=num_cores,
store=store,
model=model,
optimizer=optimizer,
loss=loss,
metrics=['accuracy'],
feature_cols=['features'],
label_cols=['label_vec'],
batch_size=2,
epochs=2,
verbose=2)
keras_model = keras_estimator.fit(df).setOutputCols(['label_prob'])
pred_df = keras_model.transform(df)
argmax = udf(lambda v: float(np.argmax(v)), returnType=T.DoubleType())
pred_df = pred_df.withColumn('label_pred', argmax(pred_df.label_prob))
preds = pred_df.collect()
assert len(preds) == df.count()
row = preds[0]
label_prob = row.label_prob.toArray().tolist()
assert label_prob[int(row.label_pred)] == max(label_prob)
def test_direct_parquet_train(self):
with spark_session('test_direct_parquet_train') as spark:
df = create_noisy_xor_data(spark)
backend = CallbackBackend()
with local_store() as store:
store.get_train_data_path = lambda v=None: store._train_path
store.get_val_data_path = lambda v=None: store._val_path
with util.prepare_data(backend.num_processes(),
store,
df,
feature_columns=['features'],
label_columns=['y'],
validation=0.2):
model = create_xor_model()
for inmemory_cache_all in [False, True]:
est = hvd_spark.TorchEstimator(
backend=backend,
store=store,
model=model,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
validation=0.2,
batch_size=1,
epochs=3,
verbose=2,
inmemory_cache_all=inmemory_cache_all)
transformer = est.fit_on_parquet()
predictions = transformer.transform(df)
assert predictions.count() == df.count()
def test_happy_run_elastic_fault_tolerant(self):
if skip_lightning_tests:
self.skipTest(
'Spark PyTorch Lightning tests conflict with Tensorflow 2.5.x: '
'https://github.com/horovod/horovod/pull/3263')
if not gloo_built():
self.skipTest("Gloo is not available")
with spark_session('test_happy_run_elastic_fault_tolerant',
max_failures=3):
with tempdir() as dir:
# these files make training function fail in given rank, epoch and batch
with open(os.path.sep.join([dir, 'rank_1_epoch_2_batch_4_fail']), 'w'), \
open(os.path.sep.join([dir, 'rank_0_epoch_3_batch_1_fail']), 'w'), \
open(os.path.sep.join([dir, 'rank_1_epoch_4_batch_2_fail']), 'w'):
pass
res = horovod.spark.run_elastic(
fn,
args=(2, 5, 5, dir),
env={'HOROVOD_LOG_LEVEL': 'DEBUG'},
num_proc=2,
min_num_proc=2,
max_num_proc=2,
start_timeout=5,
verbose=2)
self.assertListEqual([([0, 4, 0, 4, 1, 4, 0, 4], 0),
([0, 4, 0, 4, 1, 4, 0, 4], 1)], res)
def test_train_with_pytorch_infinite_async_data_loader(self):
from horovod.spark.data_loaders.pytorch_data_loaders import PytorchInfiniteAsyncDataLoader
with spark_session('test_fit_model') as spark:
df = create_noisy_xor_data(spark)
model = create_xor_model()
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
num_proc=2,
store=store,
model=model,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
validation=0.2,
batch_size=4,
epochs=2,
verbose=2,
data_loader_class=PytorchInfiniteAsyncDataLoader)
torch_model = torch_estimator.fit(df)
# TODO: Find a way to pass log metrics from remote, and assert base on the logger.
trained_model = torch_model.getModel()
pred = trained_model(torch.ones([1, 2], dtype=torch.int32))
assert len(pred) == 1
assert pred.dtype == torch.float32
def test_keras_direct_parquet_train(self, mock_fit_fn, mock_pin_gpu_fn):
mock_fit_fn.return_value = get_mock_fit_fn()
mock_pin_gpu_fn.return_value = mock.Mock()
with spark_session('test_keras_direct_parquet_train') as spark:
df = create_xor_data(spark)
backend = CallbackBackend()
with local_store() as store:
store.get_train_data_path = lambda v=None: store._train_path
store.get_val_data_path = lambda v=None: store._val_path
with util.prepare_data(backend.num_processes(),
store,
df,
feature_columns=['features'],
label_columns=['y']):
model = create_xor_model()
optimizer = tf.keras.optimizers.SGD(lr=0.1)
loss = 'binary_crossentropy'
est = hvd.KerasEstimator(backend=backend,
store=store,
model=model,
optimizer=optimizer,
loss=loss,
feature_cols=['features'],
label_cols=['y'],
batch_size=1,
epochs=3,
verbose=2)
transformer = est.fit_on_parquet()
predictions = transformer.transform(df)
assert predictions.count() == df.count()
def test_early_stop_callback(self):
from pytorch_lightning.callbacks.early_stopping import EarlyStopping
with spark_session('test_fit_model') as spark:
df = create_noisy_xor_data(spark)
model = create_xor_model()
early_stop_callback = EarlyStopping(monitor='val_loss',
min_delta=0.00,
patience=3,
verbose=True,
mode='max')
callbacks = [early_stop_callback]
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
num_proc=2,
store=store,
model=model,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
validation=0.2,
batch_size=4,
epochs=2,
verbose=2,
callbacks=callbacks)
torch_model = torch_estimator.fit(df)
# TODO: Find a way to pass log metrics from remote, and assert base on the logger.
trained_model = torch_model.getModel()
pred = trained_model(torch.ones([1, 2], dtype=torch.int32))
assert len(pred) == 1
assert pred.dtype == torch.float32
def test_model_checkpoint_callback(self):
from pytorch_lightning.callbacks.model_checkpoint import ModelCheckpoint
with spark_session('test_fit_model') as spark:
df = create_noisy_xor_data(spark)
model = create_xor_model()
with tempdir() as dir:
checkpoint_callback = ModelCheckpoint(dirpath=dir)
callbacks = [checkpoint_callback]
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
num_proc=2,
store=store,
model=model,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
validation=0.2,
batch_size=4,
epochs=2,
verbose=2,
callbacks=callbacks)
torch_model = torch_estimator.fit(df)
# TODO: Find a way to pass log metrics from remote, and assert base on the logger.
trained_model = torch_model.getModel()
pred = trained_model(torch.ones([1, 2], dtype=torch.int32))
assert len(pred) == 1
assert pred.dtype == torch.float32
def test_fit_model(self):
model = create_xor_model()
with spark_session('test_fit_model') as spark:
df = create_noisy_xor_data(spark)
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
num_proc=2,
store=store,
model=model,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
validation=0.2,
batch_size=4,
epochs=2,
verbose=2)
torch_model = torch_estimator.fit(df)
trained_model = torch_model.getModel()
pred = trained_model(torch.ones([1, 2], dtype=torch.int32))
assert len(pred) == 1
assert pred.dtype == torch.float32
def test_transform_multi_class(self):
# set dim as 2, to mock a multi class model.
model = create_xor_model(output_dim=2)
with spark_session('test_transform_multi_class') as spark:
df = create_xor_data(spark)
metadata = util._get_metadata(df)
torch_model = hvd_spark.TorchModel(history=None,
model=model,
input_shapes=[[2]],
feature_columns=['features'],
label_columns=['y'],
_metadata=metadata)
out_df = torch_model.transform(df)
# in multi class model, model output is a vector but label is number.
expected_types = {
'x1': IntegerType,
'x2': IntegerType,
'features': VectorUDT,
'weight': FloatType,
'y': FloatType,
'y__output': VectorUDT
}
for field in out_df.schema.fields:
assert type(field.dataType) == expected_types[field.name]
def test_fit_model(self):
if sys.version_info < (3, 0, 0) and is_gloo_used():
self.skipTest(
'Horovod on Spark over Gloo only supported on Python3')
model = create_xor_model()
optimizer = tf.keras.optimizers.SGD(lr=0.1)
loss = 'binary_crossentropy'
with spark_session('test_fit_model') as spark:
df = create_xor_data(spark)
with local_store() as store:
keras_estimator = hvd.KerasEstimator(num_proc=2,
store=store,
model=model,
optimizer=optimizer,
loss=loss,
feature_cols=['features'],
label_cols=['y'],
batch_size=1,
epochs=3,
verbose=2)
keras_model = keras_estimator.fit(df)
trained_model = keras_model.getModel()
pred = trained_model.predict(
[np.ones([1, 2], dtype=np.float32)])
assert len(pred) == 1
assert pred.dtype == np.float32
def test_get_col_info_error_bad_size(self):
with spark_session('test_get_col_info_error_bad_size') as spark:
data_bad_size = [[DenseVector([1.0, 1.0])], [DenseVector([1.0])]]
schema = StructType([StructField('data', VectorUDT())])
df = create_test_data_from_schema(spark, data_bad_size, schema)
with pytest.raises(ValueError):
util._get_col_info(df)
def test_happy_run(self):
def fn():
hvd.init()
res = hvd.allgather(torch.tensor([hvd.rank()])).tolist()
return res, hvd.rank()
with spark_session('test_happy_run'):
res = horovod.spark.run(fn, env={'PATH': os.environ.get('PATH')}, verbose=0)
self.assertListEqual([([0, 1], 0), ([0, 1], 1)], res)
def test_get_col_info_error_bad_shape(self):
with spark_session('test_get_col_info_error_bad_shape') as spark:
data_bad_shape = [[SparseVector(2, {0: 1.0})],
[SparseVector(1, {0: 1.0})]]
schema = StructType([StructField('data', VectorUDT())])
df = create_test_data_from_schema(spark, data_bad_shape, schema)
with pytest.raises(ValueError):
util._get_col_info(df)
def test_dummy_callback(self):
from pytorch_lightning.callbacks import Callback
model = create_xor_model()
with spark_session('test_fit_model') as spark:
df = create_noisy_xor_data(spark)
for num_proc in [1, 2]:
for epochs in [2, 3]:
class MyDummyCallback(Callback):
def __init__(self):
self.epcoh_end_counter = 0
self.train_epcoh_end_counter = 0
def on_init_start(self, trainer):
print('Starting to init trainer!')
def on_init_end(self, trainer):
print('Trainer is initialized.')
def on_epoch_end(self, trainer, model):
print('A epoch ended.')
self.epcoh_end_counter += 1
def on_train_epoch_end(self, trainer, model, unused=None):
print('A train epoch ended.')
self.train_epcoh_end_counter += 1
def on_train_end(self, trainer, model):
print('Training ends')
assert self.train_epcoh_end_counter == epochs
dm_callback = MyDummyCallback()
callbacks = [dm_callback]
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
num_proc=num_proc,
store=store,
model=model,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
validation=0.2,
batch_size=4,
epochs=epochs,
verbose=2,
callbacks=callbacks)
torch_model = torch_estimator.fit(df)
# TODO: Find a way to pass log metrics from remote, and assert base on the logger.
trained_model = torch_model.getModel()
pred = trained_model(torch.ones([1, 2], dtype=torch.int32))
assert len(pred) == 1
assert pred.dtype == torch.float32
def test_get_available_devices(self):
def fn():
hvd.init()
devices = get_available_devices()
return devices, hvd.local_rank()
with spark_session('test_get_available_devices', gpus=2):
res = horovod.spark.run(fn, env={'PATH': os.environ.get('PATH')}, verbose=0)
self.assertListEqual([(['0'], 0), (['1'], 1)], res)
def test_spark_run_func_with_non_zero_exit(self):
run_func = MagicMock(return_value=1)
def fn():
return 1
with spark_session('test_spark_run_func', cores=4):
with pytest.raises(Exception, match='^mpirun failed with exit code 1$') as e:
horovod.spark.run(fn, verbose=0, run_func=run_func)
def test_get_metadata(self):
expected_metadata = \
{
'float': {
'spark_data_type': FloatType,
'is_sparse_vector_only': False,
'intermediate_format': constants.NOCHANGE,
'max_size': 1,
'shape': 1
},
'dense': {
'spark_data_type': DenseVector,
'is_sparse_vector_only': False,
'intermediate_format': constants.ARRAY,
'max_size': 2,
'shape': 2
},
'sparse': {
'spark_data_type': SparseVector,
'is_sparse_vector_only': True,
'intermediate_format': constants.CUSTOM_SPARSE,
'max_size': 1,
'shape': 2
},
'mixed': {
'spark_data_type': DenseVector,
'is_sparse_vector_only': False,
'intermediate_format': constants.ARRAY,
'max_size': 2,
'shape': 2
},
}
with spark_session('test_get_metadata') as spark:
data = [[
1.0,
DenseVector([1.0, 1.0]),
SparseVector(2, {0: 1.0}),
DenseVector([1.0, 1.0])
],
[
1.0,
DenseVector([1.0, 1.0]),
SparseVector(2, {1: 1.0}),
SparseVector(2, {1: 1.0})
]]
schema = StructType([
StructField('float', FloatType()),
StructField('dense', VectorUDT()),
StructField('sparse', VectorUDT()),
StructField('mixed', VectorUDT())
])
df = create_test_data_from_schema(spark, data, schema)
metadata = util._get_metadata(df)
self.assertDictEqual(metadata, expected_metadata)
def test_timeout(self):
with spark_session('test_timeout'):
with pytest.raises(
Exception,
match='^Timed out waiting for Spark tasks to start.'):
horovod.spark.run(None,
num_proc=4,
start_timeout=5,
env={'PATH': os.environ.get('PATH')},
verbose=0)
def test_mpirun_not_found(self):
start = time.time()
with spark_session('test_mpirun_not_found'):
with pytest.raises(Exception,
match='^mpirun failed with exit code 127$'):
horovod.spark.run(None,
env={'PATH': '/nonexistent'},
verbose=0)
self.assertLessEqual(time.time() - start, 10,
'Failure propagation took too long')