This project has retired. For details please refer to its Attic page.
Apache Aurora

Aurora Configuration Tutorial

How to write Aurora configuration files, including feature descriptions and best practices. When writing a configuration file, make use of aurora job inspect. It takes the same job key and configuration file arguments as aurora job create or aurora job update. It first ensures the configuration parses, then outputs it in human-readable form.

You should read this after going through the general Aurora Tutorial.

The Basics

To run a job on Aurora, you must specify a configuration file that tells Aurora what it needs to know to schedule the job, what Mesos needs to run the tasks the job is made up of, and what Thermos needs to run the processes that make up the tasks. This file must have a.aurora suffix.

A configuration file defines a collection of objects, along with parameter values for their attributes. An Aurora configuration file contains the following three types of objects:

  • Job
  • Task
  • Process

A configuration also specifies a list of Job objects assigned to the variable jobs.

  • jobs (list of defined Jobs to run)

The .aurora file format is just Python. However, Job, Task, Process, and other classes are defined by a type-checked dictionary templating library called Pystachio, a powerful tool for configuration specification and reuse. Pystachio objects are tailored via {{}} surrounded templates.

When writing your .aurora file, you may use any Pystachio datatypes, as well as any objects shown in the Aurora+Thermos Configuration Reference, without import statements - the Aurora config loader injects them automatically. Other than that, an .aurora file works like any other Python script.

Aurora+Thermos Configuration Reference has a full reference of all Aurora/Thermos defined Pystachio objects.

Use Bottom-To-Top Object Ordering

A well-structured configuration starts with structural templates (if any). Structural templates encapsulate in their attributes all the differences between Jobs in the configuration that are not directly manipulated at the Job level, but typically at the Process or Task level. For example, if certain processes are invoked with slightly different settings or input.

After structural templates, define, in order, Processes, Tasks, and Jobs.

Structural template names should be UpperCamelCased and their instantiations are typically UPPER_SNAKE_CASED. Process, Task, and Job names are typically lower_snake_cased. Indentation is typically 2 spaces.

An Example Configuration File

The following is a typical configuration file. Don’t worry if there are parts you don’t understand yet, but you may want to refer back to this as you read about its individual parts. Note that names surrounded by curly braces {{}} are template variables, which the system replaces with bound values for the variables.

# --- templates here ---
class Profile(Struct):
  package_version = Default(String, 'live')
  java_binary = Default(String, '/usr/lib/jvm/java-1.7.0-openjdk/bin/java')
  extra_jvm_options = Default(String, '')
  parent_environment = Default(String, 'prod')
  parent_serverset = Default(String,
                             '/foocorp/service/bird/{{parent_environment}}/bird')

# --- processes here ---
main = Process(
  name = 'application',
  cmdline = '{{profile.java_binary}} -server -Xmx1792m '
            '{{profile.extra_jvm_options}} '
            '-jar application.jar '
            '-upstreamService {{profile.parent_serverset}}'
)

# --- tasks ---
base_task = SequentialTask(
  name = 'application',
  processes = [
    Process(
      name = 'fetch',
      cmdline = 'curl -O
              https://packages.foocorp.com/{{profile.package_version}}/application.jar'),
  ]
)

    # not always necessary but often useful to have separate task
    # resource classes
    staging_task = base_task(resources =
                     Resources(cpu = 1.0,
                               ram = 2048*MB,
                               disk = 1*GB))
production_task = base_task(resources =
                        Resources(cpu = 4.0,
                                  ram = 2560*MB,
                                  disk = 10*GB))

# --- job template ---
job_template = Job(
  name = 'application',
  role = 'myteam',
  contact = 'myteam-team@foocorp.com',
  instances = 20,
  service = True,
  task = production_task
)

# -- profile instantiations (if any) ---
PRODUCTION = Profile()
STAGING = Profile(
  extra_jvm_options = '-Xloggc:gc.log',
  parent_environment = 'staging'
)

# -- job instantiations --
jobs = [
      job_template(cluster = 'cluster1', environment = 'prod')
               .bind(profile = PRODUCTION),

      job_template(cluster = 'cluster2', environment = 'prod')
                .bind(profile = PRODUCTION),

      job_template(cluster = 'cluster1',
                    environment = 'staging',
        service = False,
        task = staging_task,
        instances = 2)
        .bind(profile = STAGING),
]

Defining Process Objects

Processes are handled by the Thermos system. A process is a single executable step run as a part of an Aurora task, which consists of a bash-executable statement.

The key (and required) Process attributes are:

  • name: Any string which is a valid Unix filename (no slashes, NULLs, or leading periods). The name value must be unique relative to other Processes in a Task.
  • cmdline: A command line run in a bash subshell, so you can use bash scripts. Nothing is supplied for command-line arguments, so $* is unspecified.

Many tiny processes make managing configurations more difficult. For example, the following is a bad way to define processes.

copy = Process(
  name = 'copy',
  cmdline = 'curl -O https://packages.foocorp.com/app.zip'
)
unpack = Process(
  name = 'unpack',
  cmdline = 'unzip app.zip'
)
remove = Process(
  name = 'remove',
  cmdline = 'rm -f app.zip'
)
run = Process(
  name = 'app',
  cmdline = 'java -jar app.jar'
)
run_task = Task(
  processes = [copy, unpack, remove, run],
  constraints = order(copy, unpack, remove, run)
)

Since cmdline runs in a bash subshell, you can chain commands with && or ||.

When defining a Task that is just a list of Processes run in a particular order, use SequentialTask, as described in the Defining Task Objects section. The following simplifies and combines the above multiple Process definitions into just two.

stage = Process(
  name = 'stage',
  cmdline = 'curl -O https://packages.foocorp.com/app.zip && '
            'unzip app.zip && rm -f app.zip')

run = Process(name = 'app', cmdline = 'java -jar app.jar')

run_task = SequentialTask(processes = [stage, run])

Process also has five optional attributes, each with a default value if one isn’t specified in the configuration:

  • max_failures: Defaulting to 1, the maximum number of failures (non-zero exit statuses) before this Process is marked permanently failed and not retried. If a Process permanently fails, Thermos checks the Process object’s containing Task for the task’s failure limit (usually 1) to determine whether or not the Task should be failed. Setting max_failuresto 0 means that this process will keep retrying until a successful (zero) exit status is achieved. Retries happen at most once every min_duration seconds to prevent effectively mounting a denial of service attack against the coordinating scheduler.

  • daemon: Defaulting to False, if daemon is set to True, a successful (zero) exit status does not prevent future process runs. Instead, the Process reinvokes after min_duration seconds. However, the maximum failure limit (max_failures) still applies. A combination of daemon=True and max_failures=0 retries a Process indefinitely regardless of exit status. This should generally be avoided for very short-lived processes because of the accumulation of checkpointed state for each process run. When running in Aurora, max_failures is capped at 100.

  • ephemeral: Defaulting to False, if ephemeral is True, the Process’ status is not used to determine if its bound Task has completed. For example, consider a Task with a non-ephemeral webserver process and an ephemeral logsaver process that periodically checkpoints its log files to a centralized data store. The Task is considered finished once the webserver process finishes, regardless of the logsaver’s current status.

  • min_duration: Defaults to 15. Processes may succeed or fail multiple times during a single Task. Each result is called a process run and this value is the minimum number of seconds the scheduler waits before re-running the same process.

  • final: Defaulting to False, this is a finalizing Process that should run last. Processes can be grouped into two classes: ordinary and finalizing. By default, Thermos Processes are ordinary. They run as long as the Task is considered healthy (i.e. hasn’t reached a failure limit). But once all regular Thermos Processes have either finished or the Task has reached a certain failure threshold, Thermos moves into a finalization stage and runs all finalizing Processes. These are typically necessary for cleaning up after the Task, such as log checkpointers, or perhaps e-mail notifications of a completed Task. Finalizing processes may not depend upon ordinary processes or vice-versa, however finalizing processes may depend upon other finalizing processes and will otherwise run as a typical process schedule.

Getting Your Code Into The Sandbox

When using Aurora, you need to get your executable code into its “sandbox”, specifically the Task sandbox where the code executes for the Processes that make up that Task.

Each Task has a sandbox created when the Task starts and garbage collected when it finishes. All of a Task’s processes run in its sandbox, so processes can share state by using a shared current working directory.

Typically, you save this code somewhere. You then need to define a Process in your .aurora configuration file that fetches the code from that somewhere to where the slave can see it. For a public cloud, that can be anywhere public on the Internet, such as S3. For a private cloud internal storage, you need to put in on an accessible HDFS cluster or similar storage.

The template for this Process is:

<name> = Process(
  name = '<name>'
  cmdline = '<command to copy and extract code archive into current working directory>'
)

Note: Be sure the extracted code archive has an executable.

Defining Task Objects

Tasks are handled by Mesos. A task is a collection of processes that runs in a shared sandbox. It’s the fundamental unit Aurora uses to schedule the datacenter; essentially what Aurora does is find places in the cluster to run tasks.

The key (and required) parts of a Task are:

  • name: A string giving the Task’s name. By default, if a Task is not given a name, it inherits the first name in its Process list.

  • processes: An unordered list of Process objects bound to the Task. The value of the optional constraints attribute affects the contents as a whole. Currently, the only constraint, order, determines if the processes run in parallel or sequentially.

  • resources: A Resource object defining the Task’s resource footprint. A Resource object has three attributes: - cpu: A Float, the fractional number of cores the Task requires. - ram: An Integer, RAM bytes the Task requires. - disk: An integer, disk bytes the Task requires.

A basic Task definition looks like:

Task(
    name="hello_world",
    processes=[Process(name = "hello_world", cmdline = "echo hello world")],
    resources=Resources(cpu = 1.0,
                        ram = 1*GB,
                        disk = 1*GB))

There are four optional Task attributes:

  • constraints: A list of Constraint objects that constrain the Task’s processes. Currently there is only one type, the order constraint. For example the following requires that the processes run in the order foo, then bar.

    constraints = [Constraint(order=['foo', 'bar'])]
    

    There is an order() function that takes order('foo', 'bar', 'baz') and converts it into [Constraint(order=['foo', 'bar', 'baz'])]. order() accepts Process name strings ('foo', 'bar') or the processes themselves, e.g. foo=Process(name='foo', ...), bar=Process(name='bar', ...), constraints=order(foo, bar)

    Note that Thermos rejects tasks with process cycles.

  • max_failures: Defaulting to 1, the number of failed processes needed for the Task to be marked as failed. Note how this interacts with individual Processes’ max_failures values. Assume a Task has two Processes and a max_failures value of 2. So both Processes must fail for the Task to fail. Now, assume each of the Task’s Processes has its own max_failures value of 10. If Process “A” fails 5 times before succeeding, and Process “B” fails 10 times and is then marked as failing, their parent Task succeeds. Even though there were 15 individual failures by its Processes, only 1 of its Processes was finally marked as failing. Since 1 is less than the 2 that is the Task’s max_failures value, the Task does not fail.

  • max_concurrency: Defaulting to 0, the maximum number of concurrent processes in the Task. 0 specifies unlimited concurrency. For Tasks with many expensive but otherwise independent processes, you can limit the amount of concurrency Thermos schedules instead of artificially constraining them through order constraints. For example, a test framework may generate a Task with 100 test run processes, but runs it in a Task with resources.cpus=4. Limit the amount of parallelism to 4 by setting max_concurrency=4.

  • finalization_wait: Defaulting to 30, the number of seconds allocated for finalizing the Task’s processes. A Task starts in ACTIVE state when Processes run and stays there as long as the Task is healthy and Processes run. When all Processes finish successfully or the Task reaches its maximum process failure limit, it goes into CLEANING state. In CLEANING, it sends SIGTERMS to any still running Processes. When all Processes terminate, the Task goes into FINALIZING state and invokes the schedule of all processes whose final attribute has a True value. Everything from the end of ACTIVE to the end of FINALIZING must happen within finalization_wait number of seconds. If not, all still running Processes are sent SIGKILLs (or if dependent on yet to be completed Processes, are never invoked).

SequentialTask: Running Processes in Parallel or Sequentially

By default, a Task with several Processes runs them in parallel. There are two ways to run Processes sequentially:

  • Include an order constraint in the Task definition’s constraints attribute whose arguments specify the processes’ run order:

    Task( ... processes=[process1, process2, process3],
          constraints = order(process1, process2, process3), ...)
    
  • Use SequentialTask instead of Task; it automatically runs processes in the order specified in the processes attribute. No constraint parameter is needed:

    SequentialTask( ... processes=[process1, process2, process3] ...)
    

SimpleTask

For quickly creating simple tasks, use the SimpleTask helper. It creates a basic task from a provided name and command line using a default set of resources. For example, in a .aurora configuration file:

SimpleTask(name="hello_world", command="echo hello world")

is equivalent to

Task(name="hello_world",
     processes=[Process(name = "hello_world", cmdline = "echo hello world")],
     resources=Resources(cpu = 1.0,
                         ram = 1*GB,
                         disk = 1*GB))

The simplest idiomatic Job configuration thus becomes:

import os
hello_world_job = Job(
  task=SimpleTask(name="hello_world", command="echo hello world"),
  role=os.getenv('USER'),
  cluster="cluster1")

When written to hello_world.aurora, you invoke it with a simple aurora job create cluster1/$USER/test/hello_world hello_world.aurora.

Combining tasks

Tasks.concat(synonym,concat_tasks) and Tasks.combine(synonym,combine_tasks) merge multiple Task definitions into a single Task. It may be easier to define complex Jobs as smaller constituent Tasks. But since a Job only includes a single Task, the subtasks must be combined before using them in a Job. Smaller Tasks can also be reused between Jobs, instead of having to repeat their definition for multiple Jobs.

With both methods, the merged Task takes the first Task’s name. The difference between the two is the result Task’s process ordering.

  • Tasks.combine runs its subtasks’ processes in no particular order. The new Task’s resource consumption is the sum of all its subtasks’ consumption.

  • Tasks.concat runs its subtasks in the order supplied, with each subtask’s processes run serially between tasks. It is analogous to the order constraint helper, except at the Task level instead of the Process level. The new Task’s resource consumption is the maximum value specified by any subtask for each Resource attribute (cpu, ram and disk).

For example, given the following:

setup_task = Task(
  ...
  processes=[download_interpreter, update_zookeeper],
  # It is important to note that {{Tasks.concat}} has
  # no effect on the ordering of the processes within a task;
  # hence the necessity of the {{order}} statement below
  # (otherwise, the order in which {{download_interpreter}}
  # and {{update_zookeeper}} run will be non-deterministic)
  constraints=order(download_interpreter, update_zookeeper),
  ...
)

run_task = SequentialTask(
  ...
  processes=[download_application, start_application],
  ...
)

combined_task = Tasks.concat(setup_task, run_task)

The Tasks.concat command merges the two Tasks into a single Task and ensures all processes in setup_task run before the processes in run_task. Conceptually, the task is reduced to:

task = Task(
  ...
  processes=[download_interpreter, update_zookeeper,
             download_application, start_application],
  constraints=order(download_interpreter, update_zookeeper,
                    download_application, start_application),
  ...
)

In the case of Tasks.combine, the two schedules run in parallel:

task = Task(
  ...
  processes=[download_interpreter, update_zookeeper,
             download_application, start_application],
  constraints=order(download_interpreter, update_zookeeper) +
                    order(download_application, start_application),
  ...
)

In the latter case, each of the two sequences may operate in parallel. Of course, this may not be the intended behavior (for example, if the start_application Process implicitly relies upon download_interpreter). Make sure you understand the difference between using one or the other.

Defining Job Objects

A job is a group of identical tasks that Aurora can run in a Mesos cluster.

A Job object is defined by the values of several attributes, some required and some optional. The required attributes are:

  • task: Task object to bind to this job. Note that a Job can only take a single Task.

  • role: Job’s role account; in other words, the user account to run the job as on a Mesos cluster machine. A common value is os.getenv('USER'); using a Python command to get the user who submits the job request. The other common value is the service account that runs the job, e.g. www-data.

  • environment: Job’s environment, typical values are devel, test, or prod.

  • cluster: Aurora cluster to schedule the job in, defined in /etc/aurora/clusters.json or ~/.clusters.json. You can specify jobs where the only difference is the cluster, then at run time only run the Job whose job key includes your desired cluster’s name.

You usually see a name parameter. By default, name inherits its value from the Job’s associated Task object, but you can override this default. For these four parameters, a Job definition might look like:

foo_job = Job( name = 'foo', cluster = 'cluster1',
          role = os.getenv('USER'), environment = 'prod',
          task = foo_task)

In addition to the required attributes, there are several optional attributes. The first (strongly recommended) optional attribute is:

  • contact: An email address for the Job’s owner. For production jobs, it is usually a team mailing list.

Two more attributes deal with how to handle failure of the Job’s Task:

  • max_task_failures: An integer, defaulting to 1, of the maximum number of Task failures after which the Job is considered failed. -1 allows for infinite failures.

  • service: A boolean, defaulting to False, which if True restarts tasks regardless of whether they succeeded or failed. In other words, if True, after the Job’s Task completes, it automatically starts again. This is for Jobs you want to run continuously, rather than doing a single run.

Three attributes deal with configuring the Job’s Task:

  • instances: Defaulting to 1, the number of instances/replicas/shards of the Job’s Task to create.

  • priority: Defaulting to 0, the Job’s Task’s preemption priority, for which higher values may preempt Tasks from Jobs with lower values.

  • production: a Boolean, defaulting to False, specifying that this is a production job backed by quota. Tasks from production Jobs may preempt tasks from any non-production job, and may only be preempted by tasks from production jobs in the same role with higher priority. WARNING: To run Jobs at this level, the Job role must have the appropriate quota. To grant quota to a particular role in production, operators use the aurora_admin set_quota command.

The final three Job attributes each take an object as their value.

  • update_config: An UpdateConfig object provides parameters for controlling the rate and policy of rolling updates. The UpdateConfig parameters are:
    • batch_size: An integer, defaulting to 1, specifying the maximum number of shards to update in one iteration.
    • restart_threshold: An integer, defaulting to 60, specifying the maximum number of seconds before a shard must move into the RUNNING state before considered a failure.
    • watch_secs: An integer, defaulting to 45, specifying the minimum number of seconds a shard must remain in the RUNNING state before considered a success.
    • max_per_shard_failures: An integer, defaulting to 0, specifying the maximum number of restarts per shard during an update. When the limit is exceeded, it increments the total failure count.
    • max_total_failures: An integer, defaulting to 0, specifying the maximum number of shard failures tolerated during an update. Cannot be equal to or greater than the job’s total number of tasks.
  • health_check_config: A HealthCheckConfig object that provides parameters for controlling a Task’s health checks via HTTP. Only used if a health port was assigned with a command line wildcard. The HealthCheckConfig parameters are:
    • initial_interval_secs: An integer, defaulting to 15, specifying the initial delay for doing an HTTP health check.
    • interval_secs: An integer, defaulting to 10, specifying the number of seconds in the interval between checking the Task’s health.
    • timeout_secs: An integer, defaulting to 1, specifying the number of seconds the application must respond to an HTTP health check with OK before it is considered a failure.
    • max_consecutive_failures: An integer, defaulting to 0, specifying the maximum number of consecutive failures before a task is unhealthy.
  • constraints: A dict Python object, specifying Task scheduling constraints. Most users will not need to specify constraints, as the scheduler automatically inserts reasonable defaults. Please do not set this field unless you are sure of what you are doing. See the section in the Aurora + Thermos Reference manual on Specifying Scheduling Constraints for more information.

The jobs List

At the end of your .aurora file, you need to specify a list of the file’s defined Jobs to run in the order listed. For example, the following runs first job1, then job2, then job3.

jobs = [job1, job2, job3]

Templating

The .aurora file format is just Python. However, Job, Task, Process, and other classes are defined by a templating library called Pystachio, a powerful tool for configuration specification and reuse.

Aurora+Thermos Configuration Reference has a full reference of all Aurora/Thermos defined Pystachio objects.

When writing your .aurora file, you may use any Pystachio datatypes, as well as any objects shown in the Aurora+Thermos Configuration Reference without import statements - the Aurora config loader injects them automatically. Other than that the .aurora format works like any other Python script.

Templating 1: Binding in Pystachio

Pystachio uses the visually distinctive {{}} to indicate template variables. These are often called “mustache variables” after the similarly appearing variables in the Mustache templating system and because the curly braces resemble mustaches.

If you are familiar with the Mustache system, templates in Pystachio have significant differences. They have no nesting, joining, or inheritance semantics. On the other hand, when evaluated, templates are evaluated iteratively, so this affords some level of indirection.

Let’s start with the simplest template; text with one variable, in this case name;

Hello {{name}}

If we evaluate this as is, we’d get back:

Hello

If a template variable doesn’t have a value, when evaluated it’s replaced with nothing. If we add a binding to give it a value:

{ "name" : "Tom" }

We’d get back:

Hello Tom

Every Pystachio object has an associated .bind method that can bind values to {{}} variables. Bindings are not immediately evaluated. Instead, they are evaluated only when the interpolated value of the object is necessary, e.g. for performing equality or serializing a message over the wire.

Objects with and without mustache templated variables behave differently:

>>> Float(1.5)
Float(1.5)

>>> Float('{{x}}.5')
Float({{x}}.5)

>>> Float('{{x}}.5').bind(x = 1)
Float(1.5)

>>> Float('{{x}}.5').bind(x = 1) == Float(1.5)
True

>>> contextual_object = String('{{metavar{{number}}}}').bind(
... metavar1 = "first", metavar2 = "second")

>>> contextual_object
String({{metavar{{number}}}})

>>> contextual_object.bind(number = 1)
String(first)

>>> contextual_object.bind(number = 2)
String(second)

You usually bind simple key to value pairs, but you can also bind three other objects: lists, dictionaries, and structurals. These will be described in detail later.

Structurals in Pystachio / Aurora

Most Aurora/Thermos users don’t ever (knowingly) interact with String, Float, or Integer Pystashio objects directly. Instead they interact with derived structural (Struct) objects that are collections of fundamental and structural objects. The structural object components are called attributes. Aurora’s most used structural objects are Job, Task, and Process:

class Process(Struct):
  cmdline = Required(String)
  name = Required(String)
  max_failures = Default(Integer, 1)
  daemon = Default(Boolean, False)
  ephemeral = Default(Boolean, False)
  min_duration = Default(Integer, 5)
  final = Default(Boolean, False)

Construct default objects by following the object’s type with (). If you want an attribute to have a value different from its default, include the attribute name and value inside the parentheses.

>>> Process()
Process(daemon=False, max_failures=1, ephemeral=False,
  min_duration=5, final=False)

Attribute values can be template variables, which then receive specific values when creating the object.

>>> Process(cmdline = 'echo {{message}}')
Process(daemon=False, max_failures=1, ephemeral=False, min_duration=5,
        cmdline=echo {{message}}, final=False)

>>> Process(cmdline = 'echo {{message}}').bind(message = 'hello world')
Process(daemon=False, max_failures=1, ephemeral=False, min_duration=5,
        cmdline=echo hello world, final=False)

A powerful binding property is that all of an object’s children inherit its bindings:

>>> List(Process)([
... Process(name = '{{prefix}}_one'),
... Process(name = '{{prefix}}_two')
... ]).bind(prefix = 'hello')
ProcessList(
  Process(daemon=False, name=hello_one, max_failures=1, ephemeral=False, min_duration=5, final=False),
  Process(daemon=False, name=hello_two, max_failures=1, ephemeral=False, min_duration=5, final=False)
  )

Remember that an Aurora Job contains Tasks which contain Processes. A Job level binding is inherited by its Tasks and all their Processes. Similarly a Task level binding is available to that Task and its Processes but is not visible at the Job level (inheritance is a one-way street.)

Mustaches Within Structurals

When you define a Struct schema, one powerful, but confusing, feature is that all of that structure’s attributes are Mustache variables within the enclosing scope once they have been populated.

For example, when Process is defined above, all its attributes such as {{name}}, {{cmdline}}, {{max_failures}} etc., are all immediately defined as Mustache variables, implicitly bound into the Process, and inherit all child objects once they are defined.

Thus, you can do the following:

>>> Process(name = "installer", cmdline = "echo {{name}} is running")
Process(daemon=False, name=installer, max_failures=1, ephemeral=False, min_duration=5,
        cmdline=echo installer is running, final=False)

WARNING: This binding only takes place in one direction. For example, the following does NOT work and does not set the Process name attribute’s value.

>>> Process().bind(name = "installer")
Process(daemon=False, max_failures=1, ephemeral=False, min_duration=5, final=False)

The following is also not possible and results in an infinite loop that attempts to resolve Process.name.

>>> Process(name = '{{name}}').bind(name = 'installer')

Do not confuse Structural attributes with bound Mustache variables. Attributes are implicitly converted to Mustache variables but not vice versa.

Templating 2: Structurals Are Factories

A Second Way of Templating

A second templating method is both as powerful as the aforementioned and often confused with it. This method is due to automatic conversion of Struct attributes to Mustache variables as described above.

Suppose you create a Process object:

>>> p = Process(name = "process_one", cmdline = "echo hello world")

>>> p
Process(daemon=False, name=process_one, max_failures=1, ephemeral=False, min_duration=5,
        cmdline=echo hello world, final=False)

This Process object, “p”, can be used wherever a Process object is needed. It can also be reused by changing the value(s) of its attribute(s). Here we change its name attribute from process_one to process_two.

>>> p(name = "process_two")
Process(daemon=False, name=process_two, max_failures=1, ephemeral=False, min_duration=5,
        cmdline=echo hello world, final=False)

Template creation is a common use for this technique:

>>> Daemon = Process(daemon = True)
>>> logrotate = Daemon(name = 'logrotate', cmdline = './logrotate conf/logrotate.conf')
>>> mysql = Daemon(name = 'mysql', cmdline = 'bin/mysqld --safe-mode')

Advanced Binding

As described above, .bind() binds simple strings or numbers to Mustache variables. In addition to Structural types formed by combining atomic types, Pystachio has two container types; List and Map which can also be bound via .bind().

Bind Syntax

The bind() function can take Python dictionaries or kwargs interchangeably (when “kwargs” is in a function definition, kwargs receives a Python dictionary containing all keyword arguments after the formal parameter list).

>>> String('{{foo}}').bind(foo = 'bar') == String('{{foo}}').bind({'foo': 'bar'})
True

Bindings done “closer” to the object in question take precedence:

>>> p = Process(name = '{{context}}_process')
>>> t = Task().bind(context = 'global')
>>> t(processes = [p, p.bind(context = 'local')])
Task(processes=ProcessList(
  Process(daemon=False, name=global_process, max_failures=1, ephemeral=False, final=False,
          min_duration=5),
  Process(daemon=False, name=local_process, max_failures=1, ephemeral=False, final=False,
          min_duration=5)
))

Binding Complex Objects

Lists
>>> fibonacci = List(Integer)([1, 1, 2, 3, 5, 8, 13])
>>> String('{{fib[4]}}').bind(fib = fibonacci)
String(5)
Maps
>>> first_names = Map(String, String)({'Kent': 'Clark', 'Wayne': 'Bruce', 'Prince': 'Diana'})
>>> String('{{first[Kent]}}').bind(first = first_names)
String(Clark)
Structurals
>>> String('{{p.cmdline}}').bind(p = Process(cmdline = "echo hello world"))
String(echo hello world)

Structural Binding

Use structural templates when binding more than two or three individual values at the Job or Task level. For fewer than two or three, standard key to string binding is sufficient.

Structural binding is a very powerful pattern and is most useful in Aurora/Thermos for doing Structural configuration. For example, you can define a job profile. The following profile uses HDFS, the Hadoop Distributed File System, to designate a file’s location. HDFS does not come with Aurora, so you’ll need to either install it separately or change the way the dataset is designated.

class Profile(Struct):
  version = Required(String)
  environment = Required(String)
  dataset = Default(String, hdfs://home/aurora/data/{{environment}}')

PRODUCTION = Profile(version = 'live', environment = 'prod')
DEVEL = Profile(version = 'latest',
                environment = 'devel',
                dataset = 'hdfs://home/aurora/data/test')
TEST = Profile(version = 'latest', environment = 'test')

JOB_TEMPLATE = Job(
  name = 'application',
  role = 'myteam',
  cluster = 'cluster1',
  environment = '{{profile.environment}}',
  task = SequentialTask(
    name = 'task',
    resources = Resources(cpu = 2, ram = 4*GB, disk = 8*GB),
    processes = [
  Process(name = 'main', cmdline = 'java -jar application.jar -hdfsPath
             {{profile.dataset}}')
    ]
   )
 )

jobs = [
  JOB_TEMPLATE(instances = 100).bind(profile = PRODUCTION),
  JOB_TEMPLATE.bind(profile = DEVEL),
  JOB_TEMPLATE.bind(profile = TEST),
 ]

In this case, a custom structural “Profile” is created to self-document the configuration to some degree. This also allows some schema “type-checking”, and for default self-substitution, e.g. in Profile.dataset above.

So rather than a .bind() with a half-dozen substituted variables, you can bind a single object that has sensible defaults stored in a single place.

Configuration File Writing Tips And Best Practices

Use As Few .aurora Files As Possible

When creating your .aurora configuration, try to keep all versions of a particular job within the same .aurora file. For example, if you have separate jobs for cluster1, cluster1 staging, cluster1 testing, andcluster2, keep them as close together as possible.

Constructs shared across multiple jobs owned by your team (e.g. team-level defaults or structural templates) can be split into separate .aurorafiles and included via the include directive.

Avoid Boilerplate

If you see repetition or find yourself copy and pasting any parts of your configuration, it’s likely an opportunity for templating. Take the example below:

redundant.aurora contains:

download = Process(
  name = 'download',
  cmdline = 'wget http://www.python.org/ftp/python/2.7.3/Python-2.7.3.tar.bz2',
  max_failures = 5,
  min_duration = 1)

unpack = Process(
  name = 'unpack',
  cmdline = 'rm -rf Python-2.7.3 && tar xzf Python-2.7.3.tar.bz2',
  max_failures = 5,
  min_duration = 1)

build = Process(
  name = 'build',
  cmdline = 'pushd Python-2.7.3 && ./configure && make && popd',
  max_failures = 1)

email = Process(
  name = 'email',
  cmdline = 'echo Success | mail feynman@tmc.com',
  max_failures = 5,
  min_duration = 1)

build_python = Task(
  name = 'build_python',
  processes = [download, unpack, build, email],
  constraints = [Constraint(order = ['download', 'unpack', 'build', 'email'])])

As you’ll notice, there’s a lot of repetition in the Process definitions. For example, almost every process sets a max_failures limit to 5 and a min_duration to 1. This is an opportunity for factoring into a common process template.

Furthermore, the Python version is repeated everywhere. This can be bound via structural templating as described in the Advanced Binding section.

less_redundant.aurora contains:

class Python(Struct):
  version = Required(String)
  base = Default(String, 'Python-{{version}}')
  package = Default(String, '{{base}}.tar.bz2')

ReliableProcess = Process(
  max_failures = 5,
  min_duration = 1)

download = ReliableProcess(
  name = 'download',
  cmdline = 'wget http://www.python.org/ftp/python/{{python.version}}/{{python.package}}')

unpack = ReliableProcess(
  name = 'unpack',
  cmdline = 'rm -rf {{python.base}} && tar xzf {{python.package}}')

build = ReliableProcess(
  name = 'build',
  cmdline = 'pushd {{python.base}} && ./configure && make && popd',
  max_failures = 1)

email = ReliableProcess(
  name = 'email',
  cmdline = 'echo Success | mail {{role}}@foocorp.com')

build_python = SequentialTask(
  name = 'build_python',
  processes = [download, unpack, build, email]).bind(python = Python(version = "2.7.3"))

Thermos Uses bash, But Thermos Is Not bash

Bad

Many tiny Processes makes for harder to manage configurations.

copy = Process(
  name = 'copy',
  cmdline = 'rcp user@my_machine:my_application .'
 )

 unpack = Process(
   name = 'unpack',
   cmdline = 'unzip app.zip'
 )

 remove = Process(
   name = 'remove',
   cmdline = 'rm -f app.zip'
 )

 run = Process(
   name = 'app',
   cmdline = 'java -jar app.jar'
 )

 run_task = Task(
   processes = [copy, unpack, remove, run],
   constraints = order(copy, unpack, remove, run)
 )

Good

Each cmdline runs in a bash subshell, so you have the full power of bash. Chaining commands with && or || is almost always the right thing to do.

Also for Tasks that are simply a list of processes that run one after another, consider using the SequentialTask helper which applies a linear ordering constraint for you.

stage = Process(
  name = 'stage',
  cmdline = 'rcp user@my_machine:my_application . && unzip app.zip && rm -f app.zip')

run = Process(name = 'app', cmdline = 'java -jar app.jar')

run_task = SequentialTask(processes = [stage, run])

Rarely Use Functions In Your Configurations

90% of the time you define a function in a .aurora file, you’re probably Doing It Wrong™.

Bad

def get_my_task(name, user, cpu, ram, disk):
  return Task(
    name = name,
    user = user,
    processes = [STAGE_PROCESS, RUN_PROCESS],
    constraints = order(STAGE_PROCESS, RUN_PROCESS),
    resources = Resources(cpu = cpu, ram = ram, disk = disk)
 )

 task_one = get_my_task('task_one', 'feynman', 1.0, 32*MB, 1*GB)
 task_two = get_my_task('task_two', 'feynman', 2.0, 64*MB, 1*GB)

Good

This one is more idiomatic. Forced keyword arguments prevents accidents, e.g. constructing a task with “32*MB” when you mean 32MB of ram and not disk. Less proliferation of task-construction techniques means easier-to-read, quicker-to-understand, and a more composable configuration.

TASK_TEMPLATE = SequentialTask(
  user = 'wickman',
  processes = [STAGE_PROCESS, RUN_PROCESS],
)

task_one = TASK_TEMPLATE(
  name = 'task_one',
  resources = Resources(cpu = 1.0, ram = 32*MB, disk = 1*GB) )

task_two = TASK_TEMPLATE(
  name = 'task_two',
  resources = Resources(cpu = 2.0, ram = 64*MB, disk = 1*GB)
)