OpenStack

[Jeff Darcy]

Mike McGrath wrote:
>Can anyone give a brief architectural overview?  (having not looked at it
>at all I was wondering stuff like:)

For the storage part (Swift) the authoritative source is
http://swift.openstack.org/overview_architecture.html or various files
in the docs/source subdir of the tarball, but I'll try to provide a
brief overview.

The storage model is very similar to Amazon's S3 - HTTP interface,
single-level hierarchy of containers (S3 buckets) and objects, mostly
whole-object get/put, some support for attributes and ACLs on objects,
etc.  The main difference that I've found is that Swift requires the API
user to obtain a token through a (semi-)separate auth service and then
use that same token for all subsequent requests, while S3 does its own
per-request auth.  Internally, the whole thing is based on consistent
hashing, but it's consistent hashing based on partitions rather than
hashing directly from item to server.  IOW, the item is hashed to a
partition rather than a server, and the assignment of partitions to
(N-way-replicated) servers is done offline via a "ring-builder" utility.
 There are separate rings and partition/server sets for objects,
containers, and accounts, with each server using its own internal
sqlite3 database to store metadata and plain files for data.  There are
also one or more HTTP proxy servers, based on WebOb and eventlet, which
provide API service.  Incoming objects are hashed to the appropriate
partition, the replica servers looked up in the global partition/server
map, and then the object contents directly streamed (I confirmed this
with the devs) to the N object servers which will hold copies.  There
are also some background processes to do re-replication, auditing, etc.
 It's all Python, and comes with a semi-decent set of unit and
functional tests which could also be used for other similar data stores.