I use FastApi with SqlAlchemy as context manager
#contextmanager
def get_session(): # need to patch all over tests files
session_ = SessionLocal()
try:
yield session_
except Exception as e:
session_.rollback()
router.py
#router.get('/get_users'):
def get_users(): # No dependencies
with get_session() as session:
users = session.query(Users).all()
return users
I need to override get_session during all my tests (I use pytest)
I could do it with #patch and patch all test. But it's not most effective way because i need to use decorator to patch in each test file, correct specify full path.
I wonder is there quick way to do it in one place or maybe use fixture?
You could try the approach in this answer to a similar question: define a pytest fixture with arguments scope="session", autouse=True, patching the context manager. If you need, you can also provide a new callable:
from contextlib import contextmanager
from unittest.mock import patch
import pytest
#pytest.fixture(scope="session", autouse=True, new_callable=fake_session)
def default_session_fixture():
with patch("your_filename.get_session"):
yield
#contextmanager
def fake_session():
yield ... # place your replacement session here
As a side note, I would highly recommend using FastAPI's dependency injection for handling your SQLAlchemy session, especially since it has built in support for exactly this kind of patching.
Related
I am trying to deploy a tf.keras image classification model to Google CloudML Engine. Do I have to include code to create serving graph separately from training to get it to serve my models in a web app? I already have my model in SavedModel format (saved_model.pb & variable files), so I'm not sure if I need to do this extra step to get it to work.
e.g. this is code directly from GCP Tensorflow Deploying models documentation
def json_serving_input_fn():
"""Build the serving inputs."""
inputs = {}
for feat in INPUT_COLUMNS:
inputs[feat.name] = tf.placeholder(shape=[None], dtype=feat.dtype)
return tf.estimator.export.ServingInputReceiver(inputs, inputs)
You are probably training your model with actual image files, while it is best to send images as encoded byte-string to a model hosted on CloudML. Therefore you'll need to specify a ServingInputReceiver function when exporting the model, as you mention. Some boilerplate code to do this for a Keras model:
# Convert keras model to TF estimator
tf_files_path = './tf'
estimator =\
tf.keras.estimator.model_to_estimator(keras_model=model,
model_dir=tf_files_path)
# Your serving input function will accept a string
# And decode it into an image
def serving_input_receiver_fn():
def prepare_image(image_str_tensor):
image = tf.image.decode_png(image_str_tensor,
channels=3)
return image # apply additional processing if necessary
# Ensure model is batchable
# https://stackoverflow.com/questions/52303403/
input_ph = tf.placeholder(tf.string, shape=[None])
images_tensor = tf.map_fn(
prepare_image, input_ph, back_prop=False, dtype=tf.float32)
return tf.estimator.export.ServingInputReceiver(
{model.input_names[0]: images_tensor},
{'image_bytes': input_ph})
# Export the estimator - deploy it to CloudML afterwards
export_path = './export'
estimator.export_savedmodel(
export_path,
serving_input_receiver_fn=serving_input_receiver_fn)
You can refer to this very helpful answer for a more complete reference and other options for exporting your model.
Edit: If this approach throws a ValueError: Couldn't find trained model at ./tf. error, you can try it the workaround solution that I documented in this answer.
I am getting the error
django.db.utils.ProgrammingError: (1146, "Table 'db_name.django_content_type' doesn't exist")
when trying to do the initial migration for a django project with a new database that I'm deploying on the production server for the first time.
I suspected the problem might be because one of the the apps had a directory full of old migrations from a SQLite3 development environment; I cleared those out but it didn't help. I also searched and found references to people having the problem with multiple databases, but I only have one.
Django version is 1.11.6 on python 3.5.4, mysqlclient 1.3.12
Some considerations:
Are you calling ContentType.objects manager anywhere in your code that may be called before the db has been built?
I am currently facing this issue and need a way to check the db table has been built before I can look up any ContentTypes
I ended up creating a method to check the tables to see if it had been created, not sure if it will also help you:
def get_content_type(cls):
from django.contrib.contenttypes.models import ContentType
from django.db import connection
if 'django_content_type' in connection.introspection.table_names():
return ContentType.objects.get_for_model(cls)
else:
return None
As for migrations, my understanding is that they should always belong in your version control repo, however you can squash, or edit as required, or even rebuild them, this linked helps me with some migrations problems:
Reset Migrations
Answering my own question:
UMDA's comment was right. I have some initialization code for the django-import-export module that looks at content_types, and evidently I have never deployed the app from scratch in a new environment since I wrote it.
Lessons learned / solution:
will wrap the offending code in an exception block, since I should
only have this exception once when deploying in a new environment
test clean deployments in a new environment more regularly.
(edit to add) consider whether your migrationsdirectories belong in .gitignore. For my purposes they do.
(Relatively new to stackoverflow etiquette - how do I credit UMDA's comment for putting me on the right track?)
I had the same issue when trying to create a generic ModelView (where the model name would be passed as a variable in urls.py). I was handling this in a kind of silly way:
Bad idea: a function that returns a generic class-based view
views.py
from django.contrib.auth.mixins import LoginRequiredMixin
from django.contrib.contenttypes.models import ContentType
from django.views.generic.edit import DeleteView
def get_generic_delete_view(model_name):
model_type = ContentType.objects.get(app_label='myapp', model=model_name)
class _GenericDelete(LoginRequiredMixin, DeleteView):
model = model_type.model_class()
template_name = "confirm_delete.html"
return _GenericDelete.as_view()
urls.py
from django.urls import path, include
from my_app import views
urlpatterns = [
path("mymodels/<name>/delete/", views.get_generic_delete_view("MyModel"),
]
Anyway. Let's not dwell in the past.
This was fixable by properly switching to a class-based view, instead of whatever infernal hybrid is outlined above, since (according to this SO post) a class-based view isn't instantiated until request-time.
Better idea: actual generic class-based view
views.py
from django.contrib.auth.mixins import LoginRequiredMixin
from django.contrib.contenttypes.models import ContentType
from django.views.generic.edit import DeleteView
class GenericDelete(LoginRequiredMixin, DeleteView):
template_name = "confirm_delete.html"
def __init__(self, **kwargs):
model = kwargs.pop("model")
model_type = ContentType.objects.get(app_label='myapp', model=model)
self.model = model_type.model_class()
super().__init__()
urls.py
from django.urls import path, include
from my_app import views
urlpatterns = [
path("mymodels/<name>/delete/", views.GenericDelete.as_view(model="MyModel"),
]
May you make new and better mistakes.
Chipping in because maybe this option will appeal better in some scenarios.
Most of the project's imports usually cascade down from your urls.py. What I usually do, is wrap the urls.py imports in a try/except statement and only create the routes if all imports were successful.
What this accomplishes is to create your project's / app's routes only if the modules were imported. If there is an error because the tables do not yet exist, it will be ignored, and the migrations will be done. In the next run, hopefully, you will have no errors in your imports and everything will run smoothly. But if you do, it's easy to spot because you won't have any URLs. Also, I usually add an error log to guide me through the issue in those cases.
A simplified version would look like this:
# Workaround to avoid programming errors on greenfield migrations
register_routes = True
try:
from myapp.views import CoolViewSet
# More imports...
except Exception as e:
register_routes = False
logger.error("Avoiding creation of routes. Error on import: {}".format(e))
if register_routes:
# Add yout url paterns here
Now, maybe you can combine Omar's answer for a more sensible, less catch-all solution.
I want to persist a trained model in CNTK and found the 'persist' functionality after some amount of searching. However, there seems to be some error in importing it.
from cntk import persist
This is throwing ImportError.
Am I doing something the wrong way? Or is this no longer supported? Is there an alternate way to persist a model?
persist is from an earlier beta. save_model is now a method of every CNTK function. So instead of doing save_model(z, filename) you do z.save_model(filename). Load_model works the same as before but you import it from cntk.ops.functions. For an example, see: https://github.com/Microsoft/CNTK/blob/v2.0.beta7.0/Tutorials/CNTK_203_Reinforcement_Learning_Basics.ipynb or https://github.com/Microsoft/CNTK/blob/v2.0.beta7.0/bindings/python/cntk/tests/persist_test.py
The functionality has moved to cntk functions. The new way is mynetwork.save_model(...) where mynetwork represents the root of your computation (typically the prediction). For loading the model you can just say mynetwork = C.load_model(...)
I'm trying to implement python-social-auth in Flask. I've ironed out tons of kinks whilst trying to interpret about 4 tutorials and a full Flask-book at the same time, and feel I've reached sort of an impasse with Flask-migrate.
I'm currently using the following code to create the tables necessary for python-social-auth to function in a flask-sqlalchemy environment.
from social.apps.flask_app.default import models
models.PSABase.metadata.create_all(db.engine)
Now, they're obviously using some form of their own Base, not related to my actual db-object. This in turn causes Flask-Migrate to completely miss out on these tables and remove them in migrations. Now, obviously I can remove these db-drops from every removal, but I can imagine it being one of those things that at one point is going to get forgotten about and all of a sudden I have no OAuth-ties anymore.
I've gotten this solution to work with the usage (and modification) of the manage.py-command syncdb as suggested by the python-social-auth Flask example
Miguel Grinberg, the author of Flask-Migrate replies here to an issue that seems to very closely resemble mine.
The closest I could find on stack overflow was this, but it doesn't shed too much light on the entire thing for me, and the answer was never accepted (and I can't get it to work, I have tried a few times)
For reference, here is my manage.py:
#!/usr/bin/env python
from flask.ext.script import Server, Manager, Shell
from flask.ext.migrate import Migrate, MigrateCommand
from app import app, db
manager = Manager(app)
manager.add_command('runserver', Server())
manager.add_command('shell', Shell(make_context=lambda: {
'app': app,
'db_session': db.session
}))
migrate = Migrate(app, db)
manager.add_command('db', MigrateCommand)
#manager.command
def syncdb():
from social.apps.flask_app.default import models
models.PSABase.metadata.create_all(db.engine)
db.create_all()
if __name__ == '__main__':
manager.run()
And to clarify, the db init / migrate / upgrade commands only create my user table (and the migration one obviously), but not the social auth ones, while the syncdb command works for the python-social-auth tables.
I understand from the github response that this isn't supported by Flask-Migrate, but I'm wondering if there's a way to fiddle in the PSABase-tables so they are picked up by the db-object sent into Migrate.
Any suggestions welcome.
(Also, first-time poster. I feel I've done a lot of research and tried quite a few solutions before I finally came here to post. If I've missed something obvious in the guidelines of SO, don't hesitate to point that out to me in a private message and I'll happily oblige)
After the helpful answer from Miguel here I got some new keywords to research. I ended up at a helpful github-page which had further references to, amongst others, the Alembic bitbucket site which helped immensely.
In the end I did this to my Alembic migration env.py-file:
from sqlalchemy import engine_from_config, pool, MetaData
[...]
# add your model's MetaData object here
# for 'autogenerate' support
# from myapp import mymodel
# target_metadata = mymodel.Base.metadata
from flask import current_app
config.set_main_option('sqlalchemy.url',
current_app.config.get('SQLALCHEMY_DATABASE_URI'))
def combine_metadata(*args):
m = MetaData()
for metadata in args:
for t in metadata.tables.values():
t.tometadata(m)
return m
from social.apps.flask_app.default import models
target_metadata = combine_metadata(
current_app.extensions['migrate'].db.metadata,
models.PSABase.metadata)
This seems to work absolutely perfectly.
The problem is that you have two sets of models, each with a different SQLAlchemy metadata object. The models from PSA were generated directly from SQLAlchemy, while your own models were generated through Flask-SQLAlchemy.
Flask-Migrate only sees the models that are defined via Flask-SQLAlchemy, because the db object that you give it only knows about the metadata for those models, it knows nothing about these other PSA models that bypassed Flask-SQLAlchemy.
So yeah, end result is that each time you generate a migration, Flask-Migrate/Alembic find these PSA tables in the db and decides to delete them, because it does not see any models for them.
I think the best solution for your problem is to configure Alembic to ignore certain tables. For this you can use the include_object configuration in the env.py module stored in the migrations directory. Basically you are going to write a function that Alembic will call every time it comes upon a new entity while generating a migration script. The function will return False when the object in question is one of these PSA tables, and True for every thing else.
Update: Another option, which you included in the response you wrote, is to merge the two metadata objects into one, then the models from your application and PSA are inspected by Alembic together.
I have nothing against the technique of merging multiple metadata objects into one, but I think it is not a good idea for an application to track migrations in models that aren't yours. Many times Alembic will not be able to capture a migration accurately, so you may need to make minor corrections on the generated script before you apply it. For models that are yours, you are capable of detecting these inaccuracies that sometimes show up in migration scripts, but when the models aren't yours I think you can miss stuff, because you will not be familiar enough with the changes that went into those models to do a good review of the Alembic generated script.
For this reason, I think it is a better idea to use my proposed include_object configuration to leave the third party models out of your migrations. Those models should be migrated according to the third party project's instructions instead.
I use two models as following:-
One which is use using db as
db = SQLAlchemy()
app['SQLALCHEMY_DATABASE_URI'] = 'postgresql://postgres:' + POSTGRES_PASSWORD + '#localhost/Flask'
db.init_app(app)
class User(db.Model):
pass
the other with Base as
Base = declarative_base()
uri = 'postgresql://postgres:' + POSTGRES_PASSWORD + '#localhost/Flask'
engine = create_engine(uri)
metadata = MetaData(engine)
Session = sessionmaker(bind=engine)
session = Session()
class Address(Base):
pass
Since you created user with db.Model you can use flask migrate on User and class Address used Base which handles fetching pre-existing table from the database.
As far as I have understood, dependency injection separates the application wiring logic from the business logic. Additionally, I try to adhere to the law of Demeter by only injecting direct collaborators.
If I understand this article correctly, proper dependency injection means that collaborators should be fully initialized when they are injected, unless lazy instantiation is required. This would mean (and is actually mentioned in the article) that objects like database connections and file streams should be up and ready at injection time.
However, opening files and connections could result in an exception, which should be handled at some point. What is the best way to go about this?
I could handle the exception at 'wire time', like in the following snippet:
class Injector:
def inject_MainHelper(self, args):
return MainHelper(self.inject_Original(args))
def inject_Original(self, args):
return open(args[1], 'rb')
class MainHelper:
def __init__(self, original):
self.original = original
def run(self):
# Do stuff with the stream
if __name__ == '__main__':
injector = Injector()
try:
helper = injector.inject_MainHelper(sys.argv)
except Exception:
print "FAILED!"
else:
helper.run()
This solution, however, starts to mix business logic with wiring logic.
Another solution is using a provider:
class FileProvider:
def __init__(self, filename, load_func, mode):
self._load = load_func
self._filename = filename
self._mode = mode
def get(self):
return self._load(self._filename, self._mode)
class Injector:
def inject_MainHelper(self, args):
return MainHelper(self.inject_Original(args))
def inject_Original(self, args):
return FileProvider(args[1], open, 'rb')
class MainHelper:
def __init__(self, provider):
self._provider = provider
def run(self):
try:
original = self._provider.get()
except Exception:
print "FAILED!"
finally:
# Do stuff with the stream
if __name__ == '__main__':
injector = Injector()
helper = injector.inject_MainHelper(sys.argv)
helper.run()
The drawback here is the added complexity of a provider and a violation of the law of Demeter.
What is the best way to deal with exceptions like this when using a dependency-injection framework as discussed in the article?
SOLUTION, based on the discussion with djna
First, as djna correctly points out, there is no actual mixing of business and wiring logic in my first solution. The wiring is happening in its own, separate class, isolated from other logic.
Secondly, there is the case of scopes. Instead of one, there are two smaller scopes:
The scope where the file is not verified yet. Here, the injection engine cannot assume anything about the file's state yet and cannot build objects that depend on it.
The scope where the file is successfully opened and verified. Here, the injection engine can create objects based on the extracted contents of the file, without the worry of blowing up on file errors.
After entering the first scope and obtaining enough information on opening and validating a file, the business logic tries to actually validate and open the file (harvesting the fruit, as djna puts it). Here, exceptions can be handled accordingly. When it is certain the file is loaded and parsed correctly, the application can enter the second scope.
Thirdly, not really related to the core problem, but still an issue: the first solution embeds business logic in the main loop, instead of the MainHelper. This makes testing harder.
class FileProvider:
def __init__(self, filename, load_func):
self._load = load_func
self._filename = filename
def load(self, mode):
return self._load(self._filename, mode)
class Injector:
def inject_MainHelper(self, args):
return MainHelper(self.inject_Original(args))
def inject_Original(self, args):
return FileProvider(args[1], open)
def inject_StreamEditor(self, stream):
return StreamEditor(stream)
class MainHelper:
def __init__(self, provider):
self._provider = provider
def run(self):
# In first scope
try:
original = self._provider.load('rb')
except Exception:
print "FAILED!"
return
# Entering second scope
editor = Injector().inject_StreamEditor(original)
editor.do_work()
if __name__ == '__main__':
injector = Injector()
helper = injector.inject_MainHelper(sys.argv)
helper.run()
Note that I have cut some corners in the last snippet. Refer to the mentioned article for more information on entering scopes.
I've had discussion about this in the contect of Java EE, EJB 3 and resources.
My understanding is that we need to distinguish between injection of the Reference to a resource and the actual use of a resource.
Take the example of a database connection we have some pseudo-code
InjectedConnectionPool icp;
public void doWork(Stuff someData) throws Exception{
Connection c = icp.getConnection().
c.writeToDb(someData);
c.close(); // return to pool
}
As I understand it:
1). That the injected resource can't be the connection itself, rather it must be a connection pool. We grab connections for a short duration and return them.
2). That any Db connection may be invalidated at any time by a failure in the DB or network. So the connection pooling resource must be able to deal with throwing away bad connections and getting new ones.
3). A failure of injection means that the component will not be started. This could happen if, for example, the injection is actually a JNDI lookup. If there's no JNDI entry we can't find the connection pool definition, can't create the pool and so can't start the component. This is not the same as actually opening a connection to the DB ...
4). ... at the time of initialisation we don't actually need to open any connections, a failure to do so just gives us an empty pool - ie. exactly the same state as if we had been running for a while and the DB went away, the pool would/could/should throw away the stale connections.
This model seems to nicely define a set of responsibilities that Demeter might accept. Injection has respobilitiy to prepare the ground, make sure that when the code needs to do something it can. The code has the responsibility to harvest the fruit, try to use the prepared material and cope with actual resource failures and opposed to failures to find out about resources.