Synnefo » snf-ganeti

=====================================

Unit tests for cmdlib / LogicalUnit's

=====================================

.. contents:: :depth: 4

This is a design document describing unit tests for the cmdlib module.

Other modules are deliberately omitted, as LU's contain the most complex

logic and are only sparingly tested.

Current state and shortcomings

==============================

The current test coverage of the cmdlib module is at only ~14%. Given

the complexity of the code this is clearly too little.

The reasons for this low coverage are numerous. There are organisational

reasons, like no strict requirements for unit tests for each feature.

But there are also design and technical reasons, which this design

document wants to address. First, it's not clear which parts of LU's

should be tested by unit tests, i.e. the test boundaries are not clearly

defined. And secondly, it's too hard to actually write unit tests for

LU's. There exists no good framework or set of tools to write easy to

understand and concise tests.

Proposed changes

================

This design document consists of two parts. Initially, the test

boundaries for cmdlib are laid out, and considerations about writing

unit tests are given. Then the test framework is described, together

with a rough overview of the individual parts and how they are meant

to be used.

Test boundaries

---------------

For the cmdlib module, every LogicalUnit is seen as a unit for testing.

Unit tests for LU's may only execute the LU but make sure that no side

effect (like filesystem access, network access or the like) takes

place. Smaller test units (like individual methods) are sensible and

will be supported by the test framework. However, they are not the main

scope of this document.

LU's require the following environment to be provided by the test code

in order to be executed:

An input opcode

  LU's get all the user provided input and parameters from the opcode.

The command processor

  Used to get the execution context id and to output logging  messages.

  It also drives the execution of LU's by calling the appropriate

  methods in the right order.

The Ganeti context

  Provides node-management methods and contains

   * The configuration. This gives access to the cluster configuration.

   * The Ganeti Lock Manager. Manages locks during the execution.

The RPC runner

  Used to communicate with node daemons on other nodes and to perform

  operations on them.

The IAllocator runner

  Calls the IAllocator with a given request.

All of those components have to be replaced/adapted by the test

framework.

The goal of unit tests at the LU level is to exercise every possible

code path in the LU at least once. Shared methods which are used by

multiple LU's should be made testable by themselves and explicit unit

tests should be written for them.

Ultimately, the code coverage for the cmdlib module should be higher

than 90%. As Python is a dynamic language, a portion of those tests

only exists to exercise the code without actually asserting for

anything in the test. They merely make sure that no type errors exist

and that potential typos etc. are caught at unit test time.

Test framework

--------------

The test framework will it make possible to write short and concise

tests for LU's. In the simplest case, only an opcode has to be provided

by the test. The framework will then use default values, like an almost

empty configuration with only the master node and no instances.

All aspects of the test environment will be configurable by individual

tests.

MCPU mocking

************

The MCPU drives the execution of LU's. It has to perform its usual

sequence of actions, but additionally it has to provide easy access to

the log output of LU's. It will contain utility assertion methods on the

output.

The mock will be a sub-class of ``mcpu.Processor`` which overrides

portions of it in order to support the additional functionality. The

advantage of being a sub-class of the original processor is the

automatic compatibility with the code running in real clusters.

Configuration mocking

*********************

Per default, the mocked configuration will contain only the master node,

no instances and default parameters. However, convenience methods for

the following use cases will be provided:

 - "Shortcut" methods to add objects to the configuration.

 - Helper methods to quickly create standard nodes/instances/etc.

 - Pre-populated default configurations for standard use-cases (i.e.

   cluster with three nodes, five instances, etc.).

 - Convenience assertion methods for checking the configuration.

Lock mocking

************

Initially, the mocked lock manager always grants all locks. It performs

the following tasks:

 - It keeps track of requested/released locks.

 - Provides utility assertion methods for checking locks (current and

   already released ones).

In the future, this component might be extended to prevent locks from

being granted. This could eventually be used to test optimistic locking.

RPC mocking

***********

No actual RPC can be made during unit tests. Therefore, those calls have

to be replaced and their results mocked. As this will entail a large

portion of work when writing tests, mocking RPC's will be made as easy as

possible. This entails:

 - Easy construction of RPC results.

 - Easy mocking of RPC calls (also multiple ones of the same type during

   one LU execution).

 - Asserting for RPC calls (including arguments, affected nodes, etc.).

IAllocator mocking

******************

Calls (also multiple ones during the execution of a LU) to the

IAllocator interface have to be mocked. The framework will provide,

similarly to the RPC mocking, provide means to specify the mocked result

and to assert on the IAllocator requests.

Future work

===========

With unit tests for cmdlib in place, further unit testing for other

modules can and should be added. The test boundaries therefore should be

aligned with the boundaries from cmdlib.

The mocked locking module can be extended to allow testing of optimistic

locking in LU's. In this case, on all requested locks are actually

granted to the LU, so it has to adapt for this situation correctly.

A higher test coverage for LU's will increase confidence in our code and

tests. Refactorings will be easier to make as more problems are caught

during tests.

After a baseline of unit tests is established for cmdlib, efficient

testing guidelines could be put in place. For example, new code could be

required to not lower the test coverage in cmdlib. Additionally, every

bug fix could be required to include a test which triggered the bug

before the fix is created.

.. vim: set textwidth=72 :

.. Local Variables:

.. mode: rst

.. fill-column: 72

.. End: