django-orchestra-test/orchestra/apps/orchestration
Marc Aymerich b29c554878 Added orchestrate management command 2015-04-01 15:49:21 +00:00
..
management Added orchestrate management command 2015-04-01 15:49:21 +00:00
tests Improved variable names consistency, cleaned up *settings.py and refactored resource aggregation 2015-03-31 12:39:08 +00:00
README.md Improvements on resource monitoring and asynchronous orchestration 2014-11-18 17:47:26 +00:00
__init__.py Improved resource monitoring 2014-07-09 16:17:43 +00:00
admin.py Improved variable names consistency, cleaned up *settings.py and refactored resource aggregation 2015-03-31 12:39:08 +00:00
backends.py Improved variable names consistency, cleaned up *settings.py and refactored resource aggregation 2015-03-31 12:39:08 +00:00
helpers.py Split webapps types into separate files 2015-03-11 16:32:33 +00:00
manager.py Added orchestrate management command 2015-04-01 15:49:21 +00:00
methods.py Added orchestrate management command 2015-04-01 15:49:21 +00:00
middlewares.py Added orchestrate management command 2015-04-01 15:49:21 +00:00
models.py Added orchestrate management command 2015-04-01 15:49:21 +00:00
settings.py Added saas icons 2014-10-10 17:17:20 +00:00
signals.py Merge all php webapps into one 2015-03-12 14:05:23 +00:00
widgets.py Webapps massive refactoring 2015-03-04 21:06:16 +00:00

README.md

Orchestration

This module handles the management of the services controlled by Orchestra.

Orchestration module has the following pieces:

  • Operation encapsulates an operation, storing the related object, the action and the backend
  • OperationsMiddleware collects and executes all save and delete operations, more on next section
  • manager it manage the execution of the operations
  • backends defines the logic that will be executed on the servers in order to control a particular service
  • router determines in which server an operation should be executed
  • Server defines a server hosting services
  • methods script execution methods, e.g. SSH
  • ScriptLog it logs the script execution

Routes

This application provides support for mapping services to server machines accross the network.

It supports routing based on Python expression, which means that you can efectively control services that are distributed accross several machines. For example, different websites that are distributed accross n web servers on a shared hosting environment.

OperationsMiddleware

When enabled, middlewares.OperationsMiddleware automatically executes the service backends when a change on the data model occurs. The main steps that performs are:

  1. Collect all save and delete model signals triggered on each HTTP request
  2. Find related backends using the routing backend
  3. Generate a single script per server (unit of work)
  4. Execute the scripts on the servers

Service Management Properties

We can identify three different characteristics regarding service management:

  • Authority: Whether or not Orchestra is the only source of the service configuration. When Orchestra is the authority then service configuration is completely generated from the Orchestra database (or services are configured to read their configuration directly from Orchestra database). Otherwise Orchestra will execute small tasks translating model changes into configuration changes, allowing manual configurations to be preserved.
  • Flow: push, when Orchestra drives the execution or pull, when external services connects to Orchestra.
  • Execution: synchronous, when the execution blocks the HTTP request, or asynchronous when it doesn't. Asynchronous execution means concurrency, and concurrency scalability.

Sorry for the bad terminology, I was not able to find more appropriate terms on the literature.

Registry vs Synchronization vs Task

From the above management properties we can extract three main service management strategies: (a) registry based management, (b) synchronization based management and (c) task based management. Orchestra provides support for all of them, it is left to you to decide which one suits your requirements better.

Following a brief description and evaluation of the tradeoffs to help on your decision making.

a. Registry Based Management

When Orchestra acts as a pure configuration registry (authority), doing nothing more than store service's configuration on the database. The configuration is pulled from Orchestra by the servers themselves, so it is asynchronous by nature.

This strategy considers two different implementations:

  • The service is configured to read the configuration directly from Orchestra database (or REST API). This approach simplifies configuration management but also can make Orchestra a single point of failure on your architecture.
  • A client-side application periodically fetches the service configuration from the Orchestra database and regenerates the service configuration files. This approach is very tolerant to failures, since the services will keep on working, and the new configuration will be applied after recovering. A delay may occur until the changes are applied to the services (eventual consistency), but it can be mitigated by notifying the application when a relevant change occur.

b. Synchronization Based Management

When Orchestra is the configuration authority and also the responsible of applying the changes on the servers (push flow). The configuration files are regenerated every time by Orchestra, deleting any existing manual configuration. This model is very consistent since it only depends on the current state of the system (memoryless). Therefore, it makes sense to execute the synchronization operation in asynchronous fashion.

In contrast to registry based management, synchronization management is fully centralized, all the management operations are driven by Orchestra so you don't need to install nor configure anything on your servers.

c. Task Based Management

This model refers when Orchestra is not the only source of configuration. Therefore, Orchestra translates isolated data model changes directly into localized changes on the service configuration, and executing them using a push strategy. For example save() or delete() object-level operations may have sibling configuration management operations. In contrast to synchronization, tasks are able to preserve configuration not performed by Orchestra.

This model is intuitive, efficient and also very consistent when tasks are execute synchronously with the request/response cycle. However, asynchronous task execution can have consistency issues; tasks have state, and this state can be lost when:

  • A failure occur while applying some changes, e.g. network error or worker crash while deleting a service
  • Scripts are executed out of order, e.g. create and delete a service is applied in inverse order

In general, synchornous execution of tasks is preferred over asynchornous, unless response delays are not tolerable.

What state does actually mean?

Lets assume you have deleted a mailbox, and Orchestra has created an script that deletes that mailbox on the mail server. However a failure has occurred and the mailbox deletion task has been lost. Since the state has also been lost it is not easy to tell what to do now in order to maintain consistency.

Additional Notes

  • The script that manage the service needs to be idempotent, i.e. the outcome of running the script is always the same, no matter how many times it is executed.

  • Renaming of attributes may lead to undesirable effects, e.g. changing a database name will create a new database rather than just changing its name.

  • The system does not magically perform data migrations between servers when its route has changed