Mel
Mel is an asychronous event-driven jobs processing engine designed to scale. Mel simplifies jobs management by abstracting away the nuances of scheduling and running jobs.
In Mel, a scheduled job is called a task. A single job may be scheduled in multiple ways, yielding multiple tasks from the same job.
Mel schedules all tasks in Redis, as a set of task id
s sorted by their times of next run. For recurring tasks, the next run is scheduled in Redis right before the current run runs.
This makes Redis the source of truth for schedules, allowing to easily scale out Mel to multiple instances (called workers), or replace or stop workers without losing schedules.
Mel supports bulk scheduling of jobs as a single atomic unit. There's also support for sequential scheduling to track a series of jobs and perform some action after they are all complete.
Types of tasks
-
Instant Tasks: These are tasks that run only once after they are scheduled, either immediately or at some specified time in the future.
-
Periodic Tasks: These are tasks that run regularly at a specified interval. They may run forever, or till some specified time in the future.
-
Cron Tasks: These are tasks that run according to a specified schedule in Unix Cron format. They may run forever, or till some specified time in the future.
Installation
-
Add the dependency to your
shard.yml
:dependencies: mel: github: GrottoPress/mel
-
Run
shards update
-
Require and configure Mel:
# ->>> src/app/config.cr # ... require "mel" require "../jobs/**" Mel.configure do |settings| settings.batch_size = 10 # <= Maximum tasks to retrieve per poll settings.poll_interval = 3.seconds settings.redis_pool_size = 25 settings.redis_url = "redis://localhost:6379/0" settings.timezone = Time::Location.load("Africa/Accra") end Log.setup(Mel.log.source, :info, Log::IOBackend.new) # ...
Usage
-
Define job:
# ->>> src/jobs/do_some_work.cr class DoSomeWork include Mel::Job # <= Required def initialize(@arg_1 : Int32, @arg_2 : String) end # <= Instance vars must be JSON-serializable # (Required) # # Main operation to be performed. # Called in a new fiber. def run # << Do work here >> end # Called in the main fiber, before spawning the fiber # that calls the `#run` method above. def before_run # ... end # Called in the same fiber that calls `#run`. # `success` is `true` only if the run succeeded. def after_run(success) if success # ... else # ... end end # Called in the main fiber before enqueueing the task in # Redis. def before_enqueue # ... end # Called in the main fiber after enqueueing the task in # Redis. `success` is `true` only if the enqueue succeeded. def after_enqueue(success) if success # ... else # ... end end end
-
Schedule job:
-
Run job now:
# ->>> src/app/some_file.cr DoSomeWork.run(arg_1: 5, arg_2: "value") # <= Alias: DoSomeWork.run_now(...)
-
Run job after given delay:
# ->>> src/app/some_file.cr DoSomeWork.run_in(5.minutes, arg_1: 5, arg_2: "value")
The given
Time::Span
can be negative. Eg:DoSomeWork.run_in(-5.minutes, ...)
. This may be useful for prioritizing certain tasks. -
Run job at specific time:
# ->>> src/app/some_file.cr DoSomeWork.run_at(10.minutes.from_now, arg_1: 5, arg_2: "value")
The specified
Time
can be in the past. Eg:DoSomeWork.run_at(-10.minutes.from_now, ...)
. This may be useful for prioritizing certain tasks. -
Run periodically:
# ->>> src/app/some_file.cr DoSomeWork.run_every(10.minutes, for: 1.hour, arg_1: 5, arg_2: "value")
Instead of
for:
, you may usetill:
and specify aTime
. Leave those out to run forever. -
Run on a Cron schedule:
# ->>> src/app/some_file.cr DoSomeWork.run_on("0 */2 * * *", for: 6.hours, arg_1: 5, arg_2: "value")
Instead of
for:
, you may usetill:
and specify aTime
. Leave those out to run forever.
The
DoSomeWork.run_*
methods accept the following additional arguments:retries
: Number of times to attempt a task after it fails, before giving up. Default:2
. Eg:DoSomeWork.run(... retries: 1, ...)
. A task fails when any exception is raised during run.
-
-
Start Mel:
-
As its own process (compiled separately):
# ->>> src/worker.cr require "./app/**" Mel.start # <= Blocks forever, polls for due tasks and runs them. # <= You may stop Mel by sending `Signal::INT` or `Signal::TERM`. # <= Mel will wait for all running tasks to complete before exiting.
-
As part of your app (useful for testing):
# ->>> spec/spec_helper.cr # ... Spec.before_each { Mel::Task::Query.truncate } Spec.after_suite do Mel.stop Mel::Task::Query.truncate end # <= `Mel.stop` waits for all running tasks to complete before exiting spawn { Mel.start } # ...
-
-
Configure compile targets:
# ->>> shard.yml # ... targets: app: main: src/app.cr worker: main: src/worker.cr # ...
Specifying task IDs
You may specify an ID whenever you schedule a new job, thus: DoSomeWork.run_*(... id: "1001", ...)
. If not specified, Mel automatically generates a unique dynamic ID for the task.
Dynamic task IDs may be OK for triggered jobs (jobs triggered by some kind of user interaction), such as a job that sends an email notification whenever a user logs in.
However, there may be jobs that are scheduled unconditionally when your app starts (global jobs). For example, sending invoices at the beginning of every month. You should specify unique static IDs for such tasks.
Otherwise, every time the app (re)starts, jobs are scheduled again, each time with a different set of IDs. Redis would accept the new schedules because the IDs are different, resulting in duplicated scheduling of the same jobs.
This is particularly important if you run multiple instances of your app. Hardcoding IDs for global jobs means that all instances hold the same IDs, so cannot reschedule a job that has already been scheduled by another instance.
A task ID may be a mixture of static and dynamic parts. For instance, you may include the current month and year for a global job that runs once a month, to ensure it is never scheduled twice within the same month.
Bulk scheduling
A common pattern is to break up long-running tasks into smaller tasks. For example:
class SendAllEmails
include Mel::Job
def initialize(@users : Array(User))
end
def run
@users.each { |user| send_email(user) }
end
private def send_email(user)
# Send email
end
end
# Schedule job
users = # ...
SendAllEmails.run(users: users)
The above job would run in a single fiber, managed by whichever worker pulls this task at run time. This could mean too much work for a single worker if the number of users is sufficiently large.
Moreover, some mails may be sent multiple times if the task is retried as a result of failure. Ideally, jobs should be idempotent, and as atomic as possible.
The preferred approach is to define a job that sends email to one user, and schedule that job for as many users as needed:
class SendAllEmails
include Mel::Job
def initialize(@users : Array(User))
end
def run
return if @users.empty?
# Pushes all jobs atomically, at the end of the block.
run do |redis|
@users.each { |user| SendEmail.run(redis: redis, user: user) }
end
end
class SendEmail
include Mel::Job
def initialize(@user : User)
end
def run
send_email(@user)
end
private def send_email(user)
# Send email
end
end
end
# Schedule job
users = # ...
SendAllEmails.run(users: users)
# <= Any `.run_*` method could be called here, as with any job.
Sequential scheduling
Bulk scheduling works OK as a fire-and-forget mechanism. However, you may need to keep track of a series of jobs as a single unit, and perform some action only after the last job is done.
This is where sequential scheduling comes in handy. Mel's event-driven design allows chaining jobs, by scheduling the next after the current one completes:
class SendAllEmails
include Mel::Job
def initialize(@users : Array(User))
end
def run
@users[0]?.try do |user|
send_email(user) # <= Send first email
end
end
def after_run(success)
return unless success
if @users[1]?
self.class.run(users: @users[1..]) # <= Schedule next email
else # <= All emails have been sent
# Do something
end
end
private def send_email(user)
# Send email
end
end
# Schedule job
users = # ...
SendAllEmails.run(users: users)
Although the example above involves a single job, sequential scheduling can be applied to multiple different jobs, each representing a step in a workflow, with each job scheduling the next job in its #after_run
callback:
struct SomeJob
include Mel::Job
def run
# Do something
end
def after_run(success)
SomeStep.run if success
end
struct SomeStep
include Mel::Job
def run
# Do something
end
def after_run(success)
SomeOtherStep.run if success
end
end
struct SomeOtherStep
include Mel::Job
def run
# Do something
end
def after_run(success)
# All done; do something
end
end
end
Query optimization
Mel's focus is on scaling out to multiple workers without hiccups. Each worker polls Redis every configurable period. Hard work has gone into reducing the number of queries made, which may be critical for performance.
There is only one queue in Mel on which all tasks are scheduled. When a worker polls redis, it makes only one query for due tasks. If any tasks are due, one more query is made to retrieve the actual task objects (JSON).
This means, each worker makes no more than two queries for every poll. When a worker runs a task, it may make a query to reschedule the task if it fails, or to schedule the next run if it is a recurring task.
Keep these in mind when configuring your poll interval, or deciding the number of workers to run.
Graceful shutdown
A Mel worker waits for all running tasks to complete before exiting, if it received a Signal::INT
or a Signal::TERM
, or if you called Mel.stop
somewhere in your code.
This means jobs are never lost mid-flight. However, because workers pull due tasks from Redis destructively, if there is a force shutdown (eg: a power cut), running tasks may be lost.
Development
Create a .env.sh
file:
#!/bin/bash
export REDIS_URL='redis://localhost:6379/0'
Update the file with your own details. Then run tests with source .env.sh && crystal spec -Dpreview_mt
.
Contributing
- Fork it
- Switch to the
master
branch:git checkout master
- Create your feature branch:
git checkout -b my-new-feature
- Make your changes, updating changelog and documentation as appropriate.
- Commit your changes:
git commit
- Push to the branch:
git push origin my-new-feature
- Submit a new Pull Request against the
GrottoPress:master
branch.