XP-456: replace the CLI arguments with a config file

Cc: finiteprods <kwok.doc@gmail.com>
Cc: Robert Steiner <robertt.debug@gmail.com>
Cc: janpetschexain <58227040+janpetschexain@users.noreply.github.com>

References
==========

https://xainag.atlassian.net/browse/XP-456
https://xainag.atlassian.net/browse/DO-58

Rationale
=========

The CLI is getting complex so it is worth loading the configuration
from a file instead.

Implementation details
======================

TOML
----

We decided to use TOML for the following reasons:

  - it is human friendly, ie easy to read and write
  - our configuration has a pretty flat structure which makes TOML
    quite adapted
  - it is well specified and has lots of implementation
  - it is well known

The other options we considered:

  - INI: it is quite frequent in the Python ecosystem to use INI for
    config files, and the standard library even provides support for
    this. However, INI is not as powerful as TOML and does not have a
    specification
  - JSON: it is very popular but is not human friendly. For instance,
    it does not support comments, is very verbose, and breaks
    easily (if a trailing comma is forgotten at the end of a list for
    instance)
  - YAML: another popular choice, but is in my opinion more complex
    than TOML.

Validation
----------

We use the third-party `schema` library to validate the
configuration. It provides a convenient way to:

- declare a schema to validate our config
- leverage third-party libraries to validate some inputs (we use the
  `idna` library to validate hostnames)
- define our own validators
- transform data after it has been validated: this can be useful to
  turn a relative path into an absolute one for example
- provide user friendly error message when the configuration is
  invalid

The `Config` class
------------------

By default, the `schema` library returns a dictionary containing a
valid configuration, but that is not convenient to manipulate in
Python. Therefore, we dynamically create a `Config` class from the
configuration schema, and instantiate a `Config` object from the data
returned by the `schema` validator.

Package re-organization
-----------------------

We moved the command line and config file logic into its own `config`
sub-package, and moved the former `xain_fl.cli.main` entrypoint into
the `xain_fl.__main__` module.

Docker infrastructure
---------------------

- Cache the xain_fl dependencies. This considerably reduces
  "edit->build-> debug" cycle, since installing the dependencies takes
  about 30 minutes.
- Move all the docker related files into the `docker/` directory

Current limitations and future work
-----------------------------------

1. The documentation generated for the `ServerConfig`,
   `AiConfig` and `StorageConfig`  classes is wrong. Each attribute is
   documented as "Alias for field number X". This can be fixed by
   having `create_class_from_schema()` setting the `__doc__` attribute
   for each attribute. However, we won't be able to automatically
   document the type of each attribute.

2. When the configuration contains an invalid value, the error message
   we generate does not contain the invalid value in question. I think
   it is possible to enable this in the future but haven't really
   looked into it.

.gitignore

@@ -129,3 +129,8 @@ dmypy.json
 # Editors
 .idea
 .vscode
+
+# In the README, we document how to use custom config. Let's ignore
+# the files mentioned in that section
+custom_config.toml
+custom_initial_weights.npy

Dockerfile.dev

@@ -3,10 +3,28 @@ FROM python:3.6-alpine
 RUN apk update && apk add python3-dev build-base git
 
 WORKDIR /app
+
+# Some dependencies require a very long compilation: protobuf, numpy,
+# grpcio. To avoid having to re-install these packages every time, we
+# pre-install them so that they are cached.
+#
+# However, we cannot just pre-install a few packages, because we don't
+# know exactly which version `pip` will pick for them. Instead, we
+# give `pip` all the package's dependencies, and let it resolve and
+# install them.
 COPY setup.py .
-COPY xain_fl xain_fl/
+RUN mkdir xain_fl && \
+    printf '__version__ = "0"\n__short_version__ = "0"' > xain_fl/__version__.py && \
+    touch README.md && \
+    python setup.py egg_info && \
+    cat *.egg-info/requires.txt | grep -v '^\[' | uniq | pip install -r /dev/stdin
+
+RUN rm -rf xain_fl README.md
 COPY README.md .
+COPY xain_fl xain_fl/
+RUN pip install -e .
 
-RUN pip install -v -e .
+COPY docker/dev/xain-fl.toml /xain-fl.toml
+COPY docker/dev/initial_weights.npy /initial_weights.npy
 
-CMD ["python3", "setup.py", "--fullname"]
+CMD ["coordinator", "--config", "/xain-fl.toml"]

README.md

@@ -68,16 +68,13 @@ $ make show
 
 ### Running the Coordinator locally
 
-To run the Coordinator on your local machine, use the command:
+To run the Coordinator on your local machine, you can use the
+`example-config.toml` file:
 
 ```shell
-$ python xain_fl/cli.py -f test_array.npy
-```
-
-For more information about the CLI and its arguments, run:
-
-```shell
-$ python xain_fl/cli.py --help
+coordinator --config example-config.toml
+# or
+python xain_fl --config example-config.toml
 ```
 
 ### Run the Coordinator from a Docker image
@@ -92,13 +89,29 @@ To run the coordinator's development image, first build the Docker image:
 $ docker build -t xain-fl-dev -f Dockerfile.dev .
 ```
 
-Then run the image, mounting the directory as a Docker volume, and call the
-entrypoint:
+Then run the image, mounting the directory as a Docker volume:
 
 ```shell
 $ docker run -v $(pwd):/app -v '/app/xain_fl.egg-info' xain-fl-dev coordinator
 ```
 
+The command above uses a default configuration but you can also use a
+custom config file and custom initial weights.
+
+For instance, if you have a `./custom_config.toml` and
+`./custom_initial_weights.npy` files that you'd like to use, you can
+mount them in the container and run the coordinator with them:
+
+```shell
+docker run \
+  -v $(pwd)/custom_config.toml:/custom_config.toml \
+  -v $(pwd)/custom_initial_weights.npy:/custom_initial_weights.npy \
+  -v $(pwd):/app \
+  -v '/app/xain_fl.egg-info' \
+  xain-fl-dev \
+  coordinator --config /custom_config.toml
+```
+
 #### Release image
 
 To run the coordinator's release image, first build it:

docker-compose-dev.yml

@@ -48,20 +48,16 @@ services:
       "
 
   xain-fl-dev:
-    environment:
-      MINIO_ACCESS_KEY: minio
-      MINIO_SECRET_KEY: minio123
     build:
       context: .
       dockerfile: Dockerfile.dev
-    command: sh -c "coordinator -f test_array.npy --storage-endpoint http://minio-dev:9000 --storage-key-id $${MINIO_ACCESS_KEY} --storage-secret-access-key $${MINIO_SECRET_KEY} --storage-bucket xain-fl-aggregated-weights"
     volumes:
-      #  don't use the local egg-info, if one exists
-      - /app/xain_fl.egg-info
-      - ./xain_fl:/app/xain_fl
+      - /app/xain_fl.egg-info  #  don't use the local egg-info, if one exists
+      - ${PWD}/xain_fl:/app/xain_fl
       - ${PWD}/setup.py:/app/setup.py
       - ${PWD}/README.md:/app/README.md
-      - ${PWD}/test_array.npy:/app/test_array.npy
+      - ${PWD}/docker/dev/xain-fl.toml:/xain-fl.toml
+      - ${PWD}/docker/dev/initial_weights.npy:/initial_weights.npy
     networks:
       - xain-fl-dev
     ports:

docker/dev/xain-fl.toml

@@ -0,0 +1,29 @@
+[server]
+# Address to listen on for incoming gRPC connections
+host = "0.0.0.0"
+# Port to listen on for incoming gRPC connections
+port = 50051
+
+[ai]
+# Path to a file containing a numpy ndarray to use a initial model weights.
+initial_weights = "/initial_weights.npy"
+# Number of global rounds the model is going to be trained for. This
+# must be a positive integer.
+rounds = 1
+# Number of local epochs per round
+epochs = 1
+# Minimum number of participants to be selected for a round.
+min_participants = 1
+# Fraction of total clients that participate in a training round. This
+# must be a float between 0 and 1.
+fraction_participants = 1.0
+
+[storage]
+# URL to the storage service to use
+endpoint = "http://minio-dev:9000"
+# Name of the bucket for storing the aggregated models
+bucket = "xain-fl-aggregated-weights"
+# AWS access key ID to use to authenticate to the storage service
+access_key_id = "minio"
+# AWS secret access to use to authenticate to the storage service
+secret_access_key = "minio123"

example-config.toml

@@ -0,0 +1,29 @@
+[server]
+# Address to listen on for incoming gRPC connections
+host = "localhost"
+# Port to listen on for incoming gRPC connections
+port = 50051
+
+[ai]
+# Path to a file containing a numpy ndarray to use a initial model weights.
+initial_weights = "./docker/dev/initial_weights.npy"
+# Number of global rounds the model is going to be trained for. This
+# must be a positive integer.
+rounds = 1
+# Number of local epochs per round
+epochs = 1
+# Minimum number of participants to be selected for a round.
+min_participants = 1
+# Fraction of total clients that participate in a training round. This
+# must be a float between 0 and 1.
+fraction_participants = 1.0
+
+[storage]
+# URL to the storage service to use
+endpoint = "http://minio-dev:9000"
+# Name of the bucket for storing the aggregated models
+bucket = "xain-fl-aggregated-weights"
+# AWS access key ID to use to authenticate to the storage service
+access_key_id = "minio"
+# AWS secret access to use to authenticate to the storage service
+secret_access_key = "minio123"

setup.py

@@ -29,6 +29,9 @@
     "structlog==19.2.0",  # Apache License 2.0
     "xain-proto==0.3.0",  # Apache License 2.0
     "boto3==1.10.48",  # Apache License 2.0
+    "toml==0.10.0",  # MIT
+    "schema~=0.7",  # MIT
+    "idna==2.8",  # BSD
 ]
 
 dev_require = [
@@ -96,5 +99,5 @@
         "docs": docs_require,
         "dev": dev_require + tests_require + docs_require,
     },
-    entry_points={"console_scripts": ["coordinator=xain_fl.cli:main"]},
+    entry_points={"console_scripts": ["coordinator=xain_fl.__main__"]},
 )

tests/store.py

@@ -7,7 +7,8 @@
 
 import numpy as np
 
-from xain_fl.coordinator.store import Store, StoreConfig
+from xain_fl.config import StorageConfig
+from xain_fl.coordinator.store import Store
 
 
 class FakeS3Resource:
@@ -71,7 +72,12 @@ class TestStore(Store):
     #
     # pylint: disable=super-init-not-called
     def __init__(self):
-        self.config = StoreConfig("endpoint_url", "access_key_id", "secret_access_key", "bucket")
+        self.config = StorageConfig(
+            endpoint="endpoint",
+            access_key_id="access_key_id",
+            secret_access_key="secret_access_key",
+            bucket="bucket",
+        )
         self.s3 = FakeS3Resource()
 
     def assert_wrote(self, round: int, weights: np.ndarray):