Driveshaft - Gearman Worker Manager

Getting Started

Install

Dependencies

• cmake 2.8.7 or later
• libgearman(-devel)
• log4cxx(-devel)
• libcurl(-devel)
• boost(-devel) 1.48 or later
• gcc 4.8 or later

Build

$ cmake .
$ make

Install

$ sudo make install

Configuration

driveshaft takes two arguments:

$ driveshaft
Allowed options:
  --help                    produce help message
  --user arg                username to run as (OPTIONAL)
  --pid_file arg            file to write process ID out to (OPTIONAL)
  --daemonize               Daemon, detach and run in the background (OPTIONAL)
  --jobsconfig arg          jobs config file path
  --logconfig arg           log config file path
  --max_running_time arg    how long can a job run before it is considered failed
                            (in seconds)
  --loop_timeout arg        how long to wait for a response from gearmand before
                            restarting event-loop (in seconds)
  --status_port arg         port to listen on to return status

jobsconfig

A simple jobsconfig file looks like this:

{
    "gearman_servers_list":
    [
        "localhost"
    ],
    "pools_list":
    {
        "ShopStats":
        {
            "job_processing_uri": "http://localhost/job.php",
            "worker_count": 0,
            "jobs_list":
            [
                "ShopStats"
            ]
        },
        "Newsfeed":
        {
            "job_processing_uri": "http://localhost/job.php",
            "worker_count": 0,
            "jobs_list":
            [
                "Newsfeed"
            ]
        },
        "Regular":
        {
            "job_processing_uri": "http://localhost/job.php",
            "worker_count": 0,
            "jobs_list":
            [
                "Sum3",
                "Sum",
                "Sum2"
            ]
        }
    }
}

Jobs Config Options

• gearman_servers_list - addresses of gearmand servers
• pools list - a list of named pools and corresponding configuration for every pool:
  • worker_count - Number of workers to reserve for jobs in this pool
  • jobs_list - Names of jobs that should be ran on the workers in this pool
  • job_processing_uri - the uri to send the job payload to for execution

logconfig

An example log config is included in the repository. For more information, see the log4cxx documentation.

loop timeout

Expressed in seconds, this is how long to wait for a job from gearmand before restarting the event loop. It is passed in to gearman_worker_set_timeout. This also influences the shutdown wait durations (hard shutdown is 2x, and graceful is 4x this value).

status port

The server listens on status_port and currently supports the command threads. For every running thread, the server returns back text formatted as <Thread-ID>\t<Pool-Name>\t<Shutdown-Flag>\tjob_handle=<Job-Handle> job_unique=<Job-Unique>\r\n

Design

1. Jobs are grouped into pools and every pool has a worker_count setting in order to define the maximum concurrency.
2. For every pool, it registers the jobs in jobs_list with gearmand and maintains worker_count threads with persistent connections to fetch jobs and submit back results.
3. When the config changes, it signals the appropriate pool threads to die. Any currently running job on that thread has a max_running_time seconds window to finish, otherwise the job is considered failed, gearmand is updated and the thread is closed. Note that the job may keep on running on the HTTP endpoint. New pool threads are created as needed to match the configuration.
4. Jobs are run via the HTTP endpoint job_processing_uri defined in the config. The endpoint will receive the class name and all the args and will have to do the right thing and return SUCCESS/FAILURE along with any response text. The thread that is processing the job blocks waiting for a response.

By reusing connections and not re-registering with gearmand on every job completion, Driveshaft saves gearmand a lot of work that impacts enqueue latency.

And by using an HTTP endpoint to actually do the heavy lifting, we get the benefits of a clean-sandbox and Opcache (and can even use HHVM!).

Endpoint Request Format

Driveshaft will send the job name and arguments via a POST request. Example:

{
    "function_name": "Sum",
    "job_handle": "H:localhost:6",
    "unique": "57a7b604-659a-11e5-9442-04013e647701",
    "workload": "[1,2]"
}

Endpoint Response Format

The endpoint should respond with a JSON payload in the body of the document. Success is indicated by returning zero for gearman_ret. The response string must be a string. Non-strings will not be accepted. Example:

{
    "gearman_ret": 0,
    "response_string" => "3"
}

Contribute

See the Contributing Guide