jobs - a Job scheduler for load regulation

Copyright (c) 2014-2018 Ulf Wiger

Version: 0.9.0

JOBS

Jobs is a job scheduler for load regulation of Erlang applications. It provides a queueing framework where each queue can be configured for throughput rate, credit pool and feedback compensation. Queues can be added and modified at runtime, and customizable "samplers" propagate load status across all nodes in the system.

Specifically, jobs provides three features:

• Job scheduling: A job is scheduled according to certain constraints. For instance, you may want to define that no more than 9 jobs of a certain type can execute simultaneously and the maximal rate at which you can start such jobs are 300 per second.
• Job queueing: When load is higher than the scheduling limits additional jobs are queued by the system to be run later when load clears. Certain rules govern queues: are they dequeued in FIFO or LIFO order? How many jobs can the queue take before it is full? Is there a deadline after which jobs should be rejected. When we hit the queue limits we reject the job. This provides a feedback mechanism on the client of the queue so you can take action.
• Sampling and dampening: Periodic samples of the Erlang VM can provide information about the health of the system in general. If we have high CPU load or high memory usage, we apply dampening to the scheduling rules: we may lower the concurrency count or the rate at which we execute jobs. When the health problem clears, we remove the dampener and run at full speed again.

Error recovery

The Jobs server is designed to not crash. However, in the unlikely event that it should occur (and it has!) Jobs does not automatically restore changes that have been effected through the API. This can be enabled, setting the Jobs environment variable auto_restore to true, or calling the function jobs_server:auto_restore(true). This will tell the jobs_server to remember every configuration change and replay them, in order, after a process restart.

Examples

The following examples are fetched from the EUC 2013 presentation on Jobs.

Regulate incoming HTTP requests (e.g. JSON-RPC)

%% @doc Handle a JSON-RPC request.
handler_session(Arg) ->
    jobs:run(
        rpc_from_web,
        fun() ->
               try
                  yaws_rpc:handler_session(
                    maybe_multipart(Arg),{?MODULE, web_rpc})
               catch
                   error:E ->
                       ...
               end
    end).

From Riak prototype, using explicit ask/done

case jobs:ask(riak_kv_fsm) of
  {ok, JobId} ->
    try
      {ok, Pid} = riak_kv_get_fsm_sup:start_get_fsm(...),
      Timeout = recv_timeout(Options),
      wait_for_reqid(ReqId, Timeout)
    after
      jobs:done(JobId)  %% Only needed if process stays alive
    end;
  {error, rejected} ->  %% Overload!
    {error, timeout}
end

Shell demo - simple rate-limited queue

2> jobs:add_queue(q, [{standard_rate,1}]).
ok
3> jobs:run(q, fun() -> io:fwrite("job: ~p~n", [time()]) end).
job: {14,37,7}
ok
4> jobs:run(q, fun() -> io:fwrite("job: ~p~n", [time()]) end).
job: {14,37,8}
ok
...
5> jobs:run(q, fun() -> io:fwrite("job: ~p~n", [time()]) end).
job: {14,37,10}
ok
6> jobs:run(q, fun() -> io:fwrite("job: ~p~n", [time()]) end).
job: {14,37,11}
ok

Shell demo - "stateful" queues

Eshell V5.9.2 (abort with ^G)
1> application:start(jobs).
ok
2> jobs:add_queue(q,
     [{standard_rate,1},
      {stateful,fun(init,_) -> {0,5};
                   ({call,{size,Sz},_,_},_) -> {reply,ok,{0,Sz}};
                   ({N,Sz},_) -> {N, {(N+1) rem Sz,Sz}}
                end}]).
ok
3> jobs:run(q,fun(Opaque) -> jobs:job_info(Opaque) end).
0
4> jobs:run(q,fun(Opaque) -> jobs:job_info(Opaque) end).
1
5> jobs:run(q,fun(Opaque) -> jobs:job_info(Opaque) end).
2
6> jobs:run(q,fun(Opaque) -> jobs:job_info(Opaque) end).
3
7> jobs:run(q,fun(Opaque) -> jobs:job_info(Opaque) end).
4
8> jobs:run(q,fun(Opaque) -> jobs:job_info(Opaque) end).
0
9> jobs:run(q,fun(Opaque) -> jobs:job_info(Opaque) end).
1
%% Resize the 'pool'
10> jobs:ask_queue(q, {size,3}).
ok
11> jobs:run(q,fun(Opaque) -> jobs:job_info(Opaque) end).
0
12> jobs:run(q,fun(Opaque) -> jobs:job_info(Opaque) end).
1
13> jobs:run(q,fun(Opaque) -> jobs:job_info(Opaque) end).
2
14> jobs:run(q,fun(Opaque) -> jobs:job_info(Opaque) end).
0
...

Demo - producers

Eshell V5.9.2 (abort with ^G)
1> application:start(jobs).
ok
2> jobs:add_queue(p,
  [{producer, fun() -> io:fwrite("job: ~p~n",[time()]) end},
   {standard_rate,1}]).
job: {14,33,51}
ok
3> job: {14,33,52}
job: {14,33,53}
job: {14,33,54}
job: {14,33,55}
...

Demo - passive queues

2> jobs:add_queue(q,[passive]).
ok
3> Fun = fun() -> io:fwrite("~p starting...~n",[self()]),
3>                Res = jobs:dequeue(q, 3),
3>                io:fwrite("Res = ~p~n", [Res])
3>       end.
#Fun<erl_eval.20.82930912>
4> jobs:add_queue(p, [{standard_counter,3},{producer,Fun}]).
<0.47.0> starting...
<0.48.0> starting...
<0.49.0> starting...
ok
5> jobs:enqueue(q, job1).
Res = [{113214444910647,job1}]
ok
<0.54.0> starting...

Demo - linked queues

3> Pid = spawn(fun() -> receive stop -> ok end end).
<0.131.0>
4> jobs:add_queue(q, [{standard_rate,1}, {link, Pid}]).
ok
5> jobs:run(q, fun() -> io:fwrite("job: ~p~n", [time()]) end).
job: {19,33,37}
ok
6> exit(Pid, kill).

=INFO REPORT==== 29-May-2020::19:33:45 ===
    jobs: removing_queue
    name: q
    reason: linked
true
7> jobs:run(q, fun() -> io:fwrite("job: ~p~n", [time()]) end).
** exception error: bad argument
     in function  jobs_server:call/3 (/home/uwiger/uw/jobs/src/jobs_server.erl, line 236)
     in call from jobs_server:run/2 (/home/uwiger/uw/jobs/src/jobs_server.erl, line 117)
8> jobs:queue_info(q).
undefined

Demo - queue status

(a@uwair)1> jobs:queue_info(q).
{queue,[{name,q},
        {mod,jobs_queue},
        {type,fifo},
        {group,undefined},
        {regulators,[{rr,[{name,{rate,q,1}},
                          {rate,{rate,[{limit,1},
                          {preset_limit,1},
                          {interval,1.0e3},
                          {modifiers,
                           [{cpu,10},{memory,10}]},
                          {active_modifiers,[]}
                         ]}}]}]},
        {max_time,undefined},
        {max_size,undefined},
        {latest_dispatch,113216378663298},
        {approved,4},
        {queued,0},
        ...,
        {stateful,undefined},
        {st,{st,45079}}]}

---

##Scenarios and Corresponding Configuration Examples

####EXAMPLE 1:

• Add counter regulated queue called heavy_crunches to limit your cpu intensive code executions to no more than 7 at a time

Configuration:

{ jobs, [
    { queues, [
        { heavy_crunches, [ { regulators, [{ counter, [{ limit, 7 }] } ] }] }
      ]
    }
  ]
}

Anywhere in your code wrap cpu-intensive work in a call to jobs server and-- voilà! --it is counter-regulated:

jobs:run( heavy_crunches,fun()->my_cpu_intensive_calculation() end)

####EXAMPLE 2:

• Add rate regulated queue called http_requests to ensure that your http server gets no more than 1000 requests per second.
• Additionally, set the queue size to 10,000 (i.e. to control queue memory consumption)

Configuration:

{ jobs, [
      { queues, [
            { http_requests, [ { max_size, 10000},  {regulators, [{ rate, [{limit, 1000}]}]}]}
        ]
      }
  ]
}

Wrap your request entry point in a call to jobs server and it will end up being rate-regulated.

jobs:run(http_requests,fun()->handle_http_request() end)

NOTE: with the config above, once 10,000 requests accumulates in the queue any incoming requests are dropped on the floor.

####EXAMPLE 3:

• HTTP requests will always have a reasonable execution time. No point in keeping them in the queue past the timeout.

• Let's create patient_user_requests queue that will keep requests in the queue for up to 10 seconds

{ patient_user_requests, [
    { max_time, 10000},
    { regulators, [{rate, [ { limit, 1000 } ] }
  ]
}

• Let's create impatient_user_requests queue that will keep requests in the queue for up to 200 milliseconds. Additionally, we'll make it a LIFO queue. Unfair, but if we assume that happy/unhappy is a boolean we're likely to maximize the happy users!

{ impatient_user_requests, [
    { max_time, 200},
    { type, lifo},
    { regulators, [{rate, [ { limit, 1000 } ] }
  ]
}

NOTE: In order to pace requests from both queues at 1000 per second, use group_rate regulation (EXAMPLE 4)

####EXAMPLE 4:

• Rate regulate http requests from multiple queues

Create group_rates regulator called http_request_rate and assign it to both impatient_user_requests and patient_user_requests

{ jobs, [
    { group_rates,[{ http_request_rate, [{limit,1000}] }] },
    { queues, [
        { impatient_user_requests,
            [ {max_time, 200},
              {type, lifo},
              {regulators,[{ group_rate, http_request_rate}]}
            ]
        },
        { patient_user_requests,
            [ {max_time, 10000},
              {regulators,[{ group_rate, http_request_rate}
            ]
        }
      ]
    }
  ]
}

####EXAMPLE 5:

• Can't afford to drop http requests on the floor once max_size is reached?
• Implement and use your own queue to persist those unfortunate http requests and serve them eventually

 -module(my_persistent_queue).
 -behaviour(jobs_queue).
 -export([  new/2,
            delete/1,
            in/3,
            out/2,
            peek/1,
            info/2,
            all/1]).

 ## implementation
 ...

Configuration:

{ jobs, [
    { queues, [
        { http_requests, [
            { mod, my_persistent_queue},
            { max_size, 10000 },
            { regulators, [ { rate, [ { limit, 1000 } ] } ] }
          ]
        }
      ]
    }
  ]
}

###The use of sampler framework

1. Get a sampler running and sending feedback to the jobs server.
2. Apply its feedback to a regulator limit.

####EXAMPLE 6:

• Adjust rate regulator limit on the fly based on the feedback from jobs_sampler_cpu named cpu_feedback

{ jobs, [
    { samplers, [{ cpu_feedback, jobs_sampler_cpu, [] } ] },
    { queues, [
        { http_requests, [
            { regulators,   [ { rate, [ { limit,1000 } ]  },
            { modifiers,    [ { cpu_feedback,  10} ] } %% 10 = % increment by which to modify the limit
          ]
        }
      ]
    }
  ]
}

Prerequisites

This application requires 'exprecs'. The 'exprecs' module is part of http://github.com/uwiger/parse_trans

Contribute

For issues, comments or feedback please [create an issue!] 1

Modules