A Gentle Introduction

The CRUSH Algorithm as it is used in ceph is used to determine on which nodes data of interest may be stored without a central controller. This is done with ‘buckets’ that define the relationship between nodes (or in the case of ceph OSDs) and their failure/allocation domains and a set of rules which specify the placement of data on these nodes (and therefore their high availability characteristics) and is implemented as a simple VM that operates on the buckets.

To put it more simply, the buckets define the ‘topology’ of your setup (how things are linked and how they should fail) and rules define which nodes you select for data and operate on these buckets. By manipulating the rules in relationship to the failure domains in the buckets you can define arbitrary HA polices that prevent against any type of data loss in a generic fashion.

Why you may want to use it

This library or its concepts may be of sue for the following problems

• cache server selection

• locating services in a cluster without a central controller

• making placement decisions for data/jobs

• reference for building basic byte code interpreters

Diffrences to Ceph’s CRUSH maps

• Only linear buckets are implemented

• No weights for buckets

• Selection mechanism currently takes buckets in order they where added

The above differences are intended to be eliminated before a v1.0 release

Available Commands

As this implementation uses an interpreter, there are several instructions that you can use. these map almost 1:1 with those from the original paper with some slight modification.

Tools

The following command can be used to run a demo program:

$ python -m crushha.interpreter

While the output is not that interesting the src code backing it may be of use if you intend to use this module. Setting up and enabling logging will also give a highly detailed and annotated look into the decisions behind each nodes selection for auditing purposes.

Current State

This module is currently in a very alpha state and subject to change however the api defined by the implementation is about 80% of the way there. This will be designed as a coroutine due to the requirement to be able to generate an arbitrary long list of nodes from a large pool and the iteration model it provides has proved useful to work with

The current design makes use of ‘yield from’ from python 3.3, python versions before this will not be considered as a porting target due to the complexity of the required trampolines to make it work correctly. The code is simple and is intended to stay so.

The Future

• Better defined op codes

• slight optimization for shorter rules (eliminating buckets by name in select)

• Finalization of the API

• More examples

• hash ring bucket types

• Predictable selection mechanism

Repo: blitz.works Docs: docs@blitz.works Email: code at pocketnix.org

Changelog