gabriel-server 0.1.6

Server for Wearable Cognitive Assistance Applications

Gabriel Server

Process data captured by a mobile device, and respond with results in real time. Gabriel uses flow control to ensure that clients send new input frames to the server at the rate that the server is able to process them. See this page for more information.

Installation

Requires Python >= 3.5

Run pip install gabriel-server

Usage

Data is processed by Cognitive Engines. Each cognitive engine is implemented in a separate class that inherits cognitive_engine.Engine. The handle method is called each time there is a new frame for the engine to process. handle gets passed a FromClient. It must return a ResultWrapper. If there are no results that should be returned to the client (which might be the case for a cognitive engine that writes results to a database), handle should return a ResultWrapper with an empty results list, when the engine is ready to start processing the next frame. The client will get a token back as soon as handle returns a ResultWrapper. Therefore, returning from handle before the engine is ready for the next frame will cause the engine to get saturated with requests faster than they can be processed.

Single Engine Workflows

The simplest possible setup involves a single cognitive engine. In this case, the Gabriel Server and the cognitive engine are run in the same Python program. Start the engine and server as follows:

local_engine.run(engine_factory=lambda: MyEngine(), filter_name='my_filter',
                 input_queue_maxsize=60, port=9099, num_tokens=2)

engine_factory should be a function that runs the constructor for the cognitive engine. A separate process gets created with Python's multiprocessing module, and engine_factory gets executed in this process. Having engine_factory return a reference to an object that was created before local_engine.run was called is not recommended.

Multiple Engine Workflows

When a workflow requires more than one cognitive engine, the Gabriel server must be run as a standalone Python program. Each cognitive engine is run as an additional separate Python program. The cognitive engines can be run on the same computer that the Gabriel server is running on, or a different computer. Under the hood, the server communicates with the cognitive engines using ZeroMQ.

The Gabriel server is run using network_engine.server_runner as follows:

server_runner.run(websocket_port=9099, zmq_address='tcp://*:5555', num_tokens=2,
                  input_queue_maxsize=60)

Cognitive engines are run using network_engine.engine_runner as follows:

engine_runner.run(engine=MyEngine(), filter_name='my_filter',
                  server_address='tcp://localhost:5555')

Note that engine should be a reference to an existing engine, not a function that runs the constructor for the engine. Unlike local_engine, network_engine.engine_runner does not run the engine in a separate process.

The server should be started before the engine runner.

Timeouts

When setting timeout values, consider the following line from ZeroMQ's guide:

If we use a TCP connection that stays silent for a long while, it will, in some networks, just die. Sending something (technically, a "keep-alive" more than a heartbeat), will keep the network alive.

server_runner.run takes an optional timeout argument. The default value of five seconds should be sufficient unless one of your cognitive engines might take more than five seconds to process a frame. This timeout value should be set to the longest amount of time that any of your cognitive engines could take to process a frame. The engine runner will not send or reply to messages while the cognitive engine is in the middle of processing a frame.

engine_runner.run takes optional timeout and request_retries parameters. request_retries specifies the number of attempts that this runner will make to reestablish a lost connection with the Gabriel server. The number of retry attempts do not get replenished at any point during the engine runner's execution. The default timeout and request_retries values should be sufficient for most configurations.

High Level Design

Each early discard filter should send one frame at a time. Every output from an early discard filter should have the same type of data, and this type should not change. For example, if a filter sends images, it should only ever send images, and it should not also include audio along with an image. Audio and images should be sent by two different filters. FromClient messages have an extras field that can be used to send metadata, such as GPS and IMU measurements, or app state. Embedding binary data to circumvent the "one type of media per filter" restriction will likely lead to cognitive engines that are difficult for other people to maintain. Multiple payloads can be sent in a single FromClient message. This is intended for cases where an input to a filter must contain several consecutive images. A single FromClient message should represent one single input to a cognitive engine.

Each client has one set of tokens per early discard filter. This allows the client to send frames that have passed "filter x" at a different rate than it sends frames that have passed "filter y." A cognitive engine can only consume frames that have passed a single filter. A cognitive engine cannot change the filter that it consumes frames from.

The Gabriel server returns a token to the client for "filter x" as soon as the first cognitive engine that consumes frames from "filter x" returns a ResultWrapper for that frame. When a second cognitive engine that also consumes frames from "filter x" returns a ResultWrapper for the same frame, the Gabriel server does not return a second token to the client. If the ResultWrapper from the second cognitive engine has an empty results list, the server will not send anything to the client in response to this ResultWrapper. If the ResultWrapper contains a non-empty results list, the server will send the ResultWrapper to the client, but it will not return a token (because it already returned the token for this frame with the result from the first cognitive engine).

Cognitive engines might not receive every frame sent to the server. In particular, the client will send frames to the server at the rate that the fastest cognitive engine can process them. Slower engines that consume frames from the same filter might miss some of the frames that were given to the fastest engine. After an engine finishes processing its current frame, it will be given the most recent frame that was given to the fastest engine. When the first engine completes the most recent frame, a new frame will be taken off the input queue and given to the fastest engine.

Future Improvements

1. If two filters both send the same payload, the payload will be sent to the server twice. Caching payloads, and referencing the cached item in subsequent FromClient messages would save bandwidth.
2. We allow multiple different cognitive engines to consume frames that have passed the same early discard filter. However, there is no way to have multiple instances of the same engine. In particular, if there were multiple cognitive engines that performed face recognition, we would not want more than one of them to process the same frame. We need some way to decide which instance of an engine should process a given frame. For each group of engines, there should be a way to toggle between the following options:
   1. Each request can go to a different engine. There should be a scheme to load balance individual requests (such as a simple round robin). This is the best option for engines that do not store any state information. Note that if the amount of state needed for each client is small, the client and engine can pass it back and forth to each other in the extras field of FromClient and FromEngine messages. This would allow the client's frames to be processed by any instance of a given engine.
   2. Each client is assigned to a specific instance of the engine. No other instances of the engine will get frames from this client. This setting will be used for engines that store state information for each client.
3. Gabriel does not expose any client identification information to cognitive engines. Clients can include this information in the extras field of FromClient messages. However, this should be added as a part of Gabriel itself at some point.
   1. Should this identity persist when the client disconnects and reconnects?
   2. If support for multiple instances of the same engine is added, should this identity be used when a group is set to assign a client to one specific instance of an engine?
4. local_engine sends results from the process running the cognitive engine to the process running the websocket server using os.pipe(). The early_discard_filter.py script in the Python client does something similar. This isn't the cleanest approach. Perhaps we should switch to one of the following:
   1. Send results to the websocket server process using multiprocessing.pipe(). Reading from this pipe directly in the event loop will block it. But we could watch the appropriate file descriptor using the asyncio event loop's add_reader function. Another option would be to use the asyncio event loop's run_in_executor method with a concurrent.futures.ThreadPoolExecutor to read the pipe. Reading from the pipe in a different OS thread seems like overkill, but I have not profiled it.
   2. Run the cognitive engine using the asyncio event loop's run_in_executor method with a concurrent.futures.ProcessPoolExecutor. This does not seem like a good option because we can only get results when the function passed to run_in_executor returns. Using this method without restarting the cognitive engine each time we want to process a new frame would probably require a hacky solution that run_in_executor was not intended for. Therefore, this seems like a bad option.
   3. You can start a subprocess by calling a python script with asyncio.create_subprocess_exec. Unfortunately you can only communicate with these subprocesses using stdin/stoud or file descriptors that you leave open with the close_fds or pass_fds arguments. However, we need to use multiprocessing.queue() for our inputs to the cognitive engine. Using os.pipe() or multiprocessing.pipe() is not an option because these might get full and block the event loop. Changing file descriptors to non-blocking mode will not work because some individual input frames might be very large. Pipe size can be increased, but there is a limit to this. It's better to use multiprocessing.queue(), which will make a best effort attempt to hold the number of items we specify when we instantiate it. Unless there is some way to pass a multiprocessing.queue() to a subprocess created with asyncio.create_subprocess_exec that isn't some horrible hack, you should not start the cognitive engine process with asyncio.create_subprocess_exec.
   4. Future versions of Python might offer a high level interface for interprocess communication that does not block the asyncio event loop. This might be a good option for sending results from the cognitive engine to the websocket server. Note that sending results in the other direction (from the websocket server to the cognitive engine) should be done using a queue that will not get full (such as multiprocessing.queue()).

Publishing Changes to PyPi

Update the version number in setup.py. Then follow these instructions.