You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
recodex-wiki/Overall-architecture.md

17 KiB

Overall Architecture

Description

ReCodEx is designed to be very modular and configurable. One such configuration is sketched in the following picture. There are two separate frontend instances with distinct databases sharing common backend part. This configuration may be suitable for MFF UK -- basic programming course and KSP competition. Note, that connections between components are not fully accurate.

Overall architecture

Web app is main part of whole project from user point of view. It provides nice user interface and it is the only part, that interacts with outside world directly. Web API contains almost all logic of the app including user management and authentication, storing and versioning files (with help of File server), counting and assigning points to users etc. Advanced users may connect to the API directly or may create custom frontends. Broker is essential part of whole architecture. It maintains list of available Workers, receives submissions from the Web API and routes them further and reports progress of evaluations back to the Web app. Worker securely runs each received job and evaluate its results. Monitor resends evaluation progress messages to the Web app in order to be presented to users.

Communication

Detailed communication inside the ReCodEx system is captured in the following image and described in sections below. Red connections are through ZeroMQ sockets, blue are through WebSockets and green are through HTTP(S). All ZeroMQ messages are sent as multipart with one string (command, option) per part, with no empty frames (unles explicitly specified otherwise).

Communication schema

Broker - Worker communication

Broker acts as server when communicating with worker. Listening IP address and port are configurable, protocol family is TCP. Worker socket is of DEALER type, broker one is ROUTER type. Because of that, very first part of every (multipart) message from broker to worker must be target worker's socket identity (which is saved on its init command).

Commands from broker to worker:

  • eval -- evaluate a job. Requires 3 message frames:
    • job_id -- identifier of the job (in ASCII representation -- we avoid endianness issues and also support alphabetic ids)
    • job_url -- URL of the archive with job configuration and submitted source code
    • result_url -- URL where the results should be stored after evaluation
  • intro -- introduce yourself to the broker (with init command) -- this is required when the broker loses track of the worker who sent the command. Possible reasons for such event are e.g. that one of the communicating sides shut down and restarted without the other side noticing.
  • pong -- reply to ping command, no arguments

Commands from worker to broker:

  • init -- introduce self to the broker. Useful on startup or after reestablishing lost connection. Requires at least 2 arguments:

    • hwgroup -- hardware group of this worker
    • header -- additional header describing worker capabilities. Format must be header_name=value, every header shall be in a separate message frame. There is no limit on number of headers.

    There is also an optional third argument -- additional information. If present, it should be separated from the headers with an empty frame. The format is the same as headers. Supported keys for additional information are:

    • description -- a human readable description of the worker for administrators (it will show up in broker logs)
    • current_job -- an identifier of a job the worker is now processing. This is useful when we are reassembling a connection to the broker and need it to know the worker will not accept a new job.
  • done -- notifying of finished job. Contains following message frames:

    • job_id -- identifier of finished job
    • result -- response result, possible values are:
      • OK -- evaluation finished successfully
      • FAILED -- job failed and cannot be reassigned to another worker (e.g. due to error in configuration)
      • INTERNAL_ERROR -- job failed due to internal worker error, but another worker might be able to process it (e.g. downloading a file failed)
    • message -- a human readable error message
  • progress -- notice about current evaluation progress. Contains following message frames:

    • job_id -- identifier of current job
    • state -- what is happening now.
      • DOWNLOADED -- submission successfuly fetched from fileserver
      • FAILED -- something bad happened and job was not executed at all
      • UPLOADED -- results are uploaded to fileserver
      • STARTED -- evaluation of tasks started
      • ENDED -- evaluation of tasks is finished
      • ABORTED -- evaluation of job encountered internal error, job will be rescheduled to another worker
      • FINISHED -- whole execution is finished and worker ready for another job execution
      • TASK -- task state changed -- see below
    • task_id -- only present for "TASK" state -- identifier of task in current job
    • task_state -- only present for "TASK" state -- result of task evaluation. One of:
      • COMPLETED -- task was successfully executed without any error, subsequent task will be executed
      • FAILED -- task ended up with some error, subsequent task will be skipped
      • SKIPPED -- some of the previous dependencies failed to execute, so this task will not be executed at all
  • ping -- tell broker I am alive, no arguments

Heartbeating

It is important for the broker and workers to know if the other side is still working (and connected). This is achieved with a simple heartbeating protocol.

The protocol requires the workers to send a ping command regularly (the interval is configurable on both sides -- future releases might let the worker send its ping interval with the init command). Upon receiving a ping command, the broker responds with pong.

Whenever a heartbeating message doesn't arrive, a counter called liveness is decreased. When this counter drops to zero, the other side is considered disconnected. When a message arrives, the liveness counter is set back to its maximum value, which is configurable for both sides.

When the broker decides a worker disconnected, it tries to reschedule its jobs to other workers.

If a worker thinks the broker crashed, it tries to reconnect periodically, with a bounded, exponentially increasing delay.

This protocol proved great robustness in real world testing. Thus whole backend is reliable and can outlive short term issues with connection without problems. Also, increasing delay of ping messages does not flood the network when there are problems. We experienced no issues since we are using this protocol.

Worker - File Server communication

Worker is communicating with file server only from execution thread. Supported protocol is HTTP optionally with SSL encryption (recommended). If supported by server and used version of libcurl, HTTP/2 standard is also available. File server should be set up to require basic HTTP authentication and worker is capable to send corresponding credentials with each request.

Worker side

Workers comunicate with the file server in both directions -- they download student's submissions and then upload evaluation results. Internally, worker is using libcurl C library with very similar setup. In both cases it can verify HTTPS certificate (on Linux against system cert list, on Windows against downloaded one from CURL website during installation), support basic HTTP authentication, offer HTTP/2 with fallback to HTTP/1.1 and fail on error (returned HTTP status code is >=400). Worker have list of credentials to all available file servers in its config file.

  • download file -- standard HTTP GET request to given URL expecting file content as response
  • upload file -- standard HTTP PUT request to given URL with file data as body -- same as command line tool curl with option --upload-file

File server side

File server has its own internal directory structure, where all the files are stored. It provides simple REST API to get them or create new ones. File server does not provide authentication or secured connection by itself, but it is supposed to run file server as WSGI script inside a web server (like Apache) with proper configuration. Relevant commands for communication with workers:

  • GET /submission_archives/<id>.<ext> -- gets an archive with submitted source code and corresponding configuration of this job evaluation
  • GET /exercises/<hash> -- gets a file, common usage is for input files or reference result files
  • PUT /results/<id>.<ext> -- upload archive with evaluation results under specified name (should be same id as name of submission archive). On successful upload returns JSON { "result": "OK" } as body of returned page.

If not specified otherwise, zip format of archives is used. Symbol / in API description is root of file server's domain. If the domain is for example fs.recodex.org with SSL support, getting input file for one task could look as GET request to https://fs.recodex.org/tasks/8b31e12787bdae1b5766ebb8534b0adc10a1c34c.

Broker - Monitor communication

Broker communicates with monitor also through ZeroMQ over TCP protocol. Type of socket is same on both sides, ROUTER. Monitor is set to act as server in this communication, its IP address and port are configurable in monitor's config file. ZeroMQ socket ID (set on monitor's side) is "recodex-monitor" and must be sent as first frame of every multipart message -- see ZeroMQ ROUTER socket documentation for more info.

Note that the monitor is designed so that it can receive data both from the broker and workers. The current architecture prefers the broker to do all the communication so that the workers do not have to know too many network services.

Monitor is treated as a somewhat optional part of whole solution, so no special effort on communication realibility was made.

Commands from monitor to broker:

Because there is no need for the monitor to communicate with the broker, there are no commands so far. Any message from monitor to broker is logged and discarded.

Commands from broker to monitor:

  • progress -- notification about progress with job evaluation. See Progress callback section for more info.

Broker - Web API communication

Broker communicates with main REST API through ZeroMQ connection over TCP. Socket type on broker side is ROUTER, on frontend part it is DEALER. Broker acts as a server, its IP address and port is configurable in the API.

Commands from API to broker:

  • eval -- evaluate a job. Requires at least 4 frames:
    • job_id -- identifier of this job (in ASCII representation -- we avoid endianness issues and also support alphabetic ids)
    • header -- additional header describing worker capabilities. Format must be header_name=value, every header shall be in a separate message frame. There is no maximum limit on number of headers. There may be also no headers at all.
    • empty frame -- frame which contains only empty string and serves only as breakpoint after headers
    • job_url -- URI location of archive with job configuration and submitted source code
    • result_url -- remote URI where results will be pushed to

Commands from broker to API (all are responses to eval command):

  • ack -- this is first message which is sent back to frontend right after eval command arrives, basically it means "Hi, I am all right and am capable of receiving job requests", after sending this broker will try to find acceptable worker for arrived request
  • accept -- broker is capable of routing request to a worker
  • reject -- broker cannot handle this job (for example when the requirements specified by the headers cannot be met). There are (rare) cases when the broker finds that it cannot handle the job after it was confirmed. In such cases it uses the frontend REST API to mark the job as failed.

Asynchronous communication between broker and API

Only a fraction of the errors that can happen during evaluation can be detected while there is a ZeroMQ connection between the API and broker. To notify the frontend of the rest, we need an asynchronous communication channel that can be used by the broker when the status of a job changes (it's finished, it failed permanently, the only worker capable of processing it disconnected...).

This functionality is supplied by the broker-reports/ API endpoint group -- see its documentation for more details.

File Server - Web API communication

File server has a REST API for interaction with other parts of ReCodEx. Description of communication with workers is in File server side section. On top of that, there are other commands for interaction with the API:

  • GET /results/<id>.<ext> -- download archive with evaluated results of job id
  • POST /submissions/<id> -- upload new submission with identifier id. Expects that the body of the POST request uses file paths as keys and the content of the files as values. On successful upload returns JSON { "archive_path": <archive_url>, "result_path": <result_url> } in response body. From archive_path the submission can be downloaded (by worker) and corresponding evaluation results should be uploaded to result_path.
  • POST /tasks -- upload new files, which will be available by names equal to sha1sum of their content. There can be uploaded more files at once. On successful upload returns JSON { "result": "OK", "files": <file_list> } in response body, where file_list is dictionary of original file name as key and new URL with already hashed name as value.

There are no plans yet to support deleting files from this API. This may change in time.

Web API calls these fileserver endpoints with standard HTTP requests. There are no special commands involved. There is no communication in opposite direction.

Monitor - Web app communication

Monitor interacts with web application through WebSocket connection. Monitor acts as server and browsers are connecting to it. IP address and port are configurable. When client connects to the monitor, it sends a message with string representation of channel id (which messages are interested in, usually id of evaluating job). There can be multiple listeners per channel, even (shortly) delayed connections will receive all messages from the very beginning.

When monitor receives progress message from broker there are two options:

  • there is no WebSocket connection for listed channel (job id) -- message is dropped
  • there is active WebSocket connection for listed channel -- message is parsed into JSON format (see below) and send as string to that established channel. Messages for active connections are queued, so no messages are discarded even on heavy workload.

Message JSON format is dictionary (associative array) with keys:

  • command -- type of progress, one of:
    • DOWNLOADED -- submission successfuly fetched from fileserver
    • FAILED -- something bad happened and job was not executed at all
    • UPLOADED -- results are uploaded to fileserver
    • STARTED -- evaluation of tasks started
    • ENDED -- evaluation of all tasks finished, worker now just have to send results and cleanup after execution
    • ABORTED -- evaluation of job encountered internal error, job will be rescheduled to another worker
    • FINISHED -- whole execution finished and worker is ready for another job execution
    • TASK -- task state changed, further information will be provided -- see below
  • task_id -- id of currently evaluated task. Present only if command is "TASK".
  • task_state -- state of task with id task_id. Present only if command is "TASK". Value is one of "COMPLETED", "FAILED" and "SKIPPED".
    • COMPLETED -- task was successfully executed without any error, subsequent task will be executed
    • FAILED -- task ended up with some error, subsequent task will be skipped
    • SKIPPED -- some of the previous dependencies failed to execute, so this task will not be executed at all

Web app - Web API communication

Provided web application runs as javascript client inside user's browser. It communicates with REST API on the server through standard HTTP requests. Documentation of the main REST API is in separate document due to its extensiveness. Results are returned as JSON payload, which is simply parsed in web application and presented to the users.