User Tools

Site Tools


CEPH: Architecture Overview

Name: CEPH: Architecture Overview
Description: This document is the start line to understand this technology
Modification date :11/02/2019
Notify changes to:dodger
Tags: ceph, oss
Scalate to:The_fucking_bofh

Initial concepts

What is:

  • Ceph is a advanced Object Storage Service (OSS).
  • A easy way to store tons of objects of any kind and store it in multiple ways: S3/Swift, NFS, Block device.
  • It's fast, scale-out and fault tolerance.

What is NOT:

  • A database
  • A cache

You can read/watch a basic introduction here:

Basic architecture

VERY BASIC (2 minutes approach)

From a very simple point of view: Ceph acts as a disk.
With the correct OS library like can be the kernel module for “ext4”/“btrfs”, you'll be able to read/write directly.
This library is called librados.
And interact with RADOS the reliable autonomic distributed object store which is the Object storage itselv.
Continuing with this simple view, like ext4 for example, you'll have data blocks and journal blocks to maintain consistency; that are OSD (object store data) and MON (monitoring) nodes:

  • OSD: serve data
  • MON: keep track of OSD nodes, DOES NOT SERVE DATA

So you'll have something like this:

BASIC Architecture

Going deeper, you'll find that the data placement over OSD nodes is calculated by an algorithm called CRUSH: Controlled Replication Under Scalable Hashing which is:

  • Pseudo-random, very fast, repeatable and deterministic
  • Makes a uniform data distribution
  • Results in a stable mapping of data (with very limited data migration on changes)
  • It has a rule-based configuration: adjustable replication, weights …

So when a client want to write data in the CEPH cluster though RADOS, librados in the client side invokes CRUSH to perform the calculation on where of the available OSD's write the data.

That result in a very strong architecture with no single point of failure, cause you'll not have a/many node/s taking care of metadata.
Its also really fast: you'll have n*osd servers to perform read/writes.
Its robust, if any of the OSD node fails, the data is replicated N times (where N is a config option) through other OSD's and will be accessible with the CRUSH calculation.
Also if any OSD fail, the MONitors will re-map the cluster and OSD's will re-replicate the data to have copied N times in the cluster.

In a graph

Usage Cases

CEPH as REST Object Storage

The unique diference in this case is there's a new component involved, the gateway which translate HTTP/REST into librados.
That's all.
You'll have a lot of overhead/performance gap using the gateway instead of using librados directly…
So if you take the previous graph, simplified, you'll have:

CEPH as Filesystem Architecture

Official documentation:

Again, the diference in using CEPH as “filesystem” is that there's another component: the “Metadata server”.
Metadata a role similar to Monitor:

  • It has the metadata database, that is a “inode” in a common filesystem.
  • The Metadata nodes also replicate and vote between them.
  • The filesystem tree is split dynamically between all the metadata nodes.

Real Life

ePayments PRO/PRE

Node list:

  • Admin:
    • ACCLM-OSADM-001
  • OSD:
    • ACCLM-OSD-001
    • ACCLM-OSD-002
    • ACCLM-OSD-003
    • ACCLM-OSD-004
    • ACCLM-OSD-005
  • Gateways:
    • ACCLM-OSGW-001
    • ACCLM-OSGW-002
  • Monitors:
    • ACCLM-OSM-001
    • ACCLM-OSADM-001

Clover schematics (here comes the monster)


  • AVMLP-OSLB-001
  • AVMLP-OSLB-002

Object Gateways:

  • AVMLP-OSGW-001
  • AVMLP-OSGW-002
  • AVMLP-OSGW-003
  • AVMLP-OSGW-004


  • AVMLP-OSM-001
  • AVMLP-OSM-002
  • AVMLP-OSM-003
  • AVMLP-OSM-004
  • AVMLP-OSM-005
  • AVMLP-OSM-006

Metadata servers:

  • AVMLP-OSFS-001
  • AVMLP-OSFS-002
  • AVMLP-OSFS-003
  • AVMLP-OSFS-004

Data servers:

  • AVMLP-OSD-001
  • AVMLP-OSD-002
  • AVMLP-OSD-003
  • AVMLP-OSD-004
  • AVMLP-OSD-005
  • AVMLP-OSD-006
  • AVMLP-OSD-007
  • AVMLP-OSD-008
  • AVMLP-OSD-009
  • AVMLP-OSD-010
  • AVMLP-OSD-011
  • AVMLP-OSD-012
  • AVMLP-OSD-013
  • AVMLP-OSD-014
  • AVMLP-OSD-015
  • AVMLP-OSD-016
  • AVMLP-OSD-017
  • AVMLP-OSD-018
  • AVMLP-OSD-019
  • AVMLP-OSD-020


As object gateway

As Filesystem (cephfs)

Public ceph schema

Considerations for newcomers

When requesting access to any of our object storages or if you're a newcomer, you should know that:

  • The_fucking_bofh team will give you the “access_key” and “secret_key” to access the object storage, but also will inform you of the name of the user, this username is unique and will be something related to the name of the project.
  • The_fucking_bofh team will not create the bucket, you can easily do it :-)
  • bucket names are unique across ONE object storage instance: If you create the bucket “test” no one can create another bucket named “test” in the same object storage. Have this in mind as you're sharing ceph with other users and with yourself (for PRO/PRE/DEV/TEST/STAGING/DEMO/SERVERLESS)! So don't use “ciberterminal” as bucketname :-)
  • What you'll need to connect to the object storage is:
  • Please specify where do you want your user to be created, for example: PRO/PRE/DEV/TEST/STAGING/DEMO/SERVERLESS

Here you have a template to request a new user for the object storage:

Good morning #infrastructure,
We're facing a new project that involves store tons of objects and We want to use our incredible Ceph installation.
Please provide us a new User so we can store all the data from this project.

Name of the project: "This_template_sucks"
Environment: DEVELOPMENT
Expected number of buckets: 666

Thanks for your effort, best regards!
linux/ceph/start_here.txt · Last modified: 2022/02/11 11:36 by