← Back to team overview

openstack team mailing list archive

Re: Fwd: Chef Deployment System for Swift - a proposed design - feedback?

 

Andi,

That's great!  Taking a look right now

-jwb

On Mon, May 23, 2011 at 1:41 PM, andi abes <andi.abes@xxxxxxxxx> wrote:

>
>
> It took a while, but finally:
> https://github.com/dellcloudedge/openstack-swift
>
> Jay, I've added a swift-proxy-acct and swift-proxy (without account
> management).
>
> This cookbook is an advanced "leak" of for swift, soon to be followed with
> a leak of a nova cookbook. The full crowbar that was mentioned is on its
> way...
>
> To use these recipes (with default settings) you just need to pick your
> storage nodes and 1 or more proxies. Then assign the appropriate roles
> (swift-storage swift-proxy or swift-proxy-acct) using the chef ui or a knife
> command. Choose one of the nodes and assign it the swift-node-compute. and
> the swift cluster is built (because of async nature of multi-node
> deployments, it might require a few chef-client runs while the ring files
> are generated and pushed around.
>
> have a spin. eager to hear comments.
>
>
>
>
>
>
>
>
> On Mon, May 2, 2011 at 11:36 AM, andi abes <andi.abes@xxxxxxxxx> wrote:
>
>> Jay,
>>
>> hmmm, interesting point about account management in the proxy. Guess
>> you're suggesting that you have 2 flavors of a proxy server - one with
>> account management enabled and one without?
>>  Is the main concern here security - you'd have more controls on the
>> account management servers? Or is this about something else?
>>
>> About ring-compute:
>> so there are 2 concerns I was thinking about with rings - a) make sure the
>> ring information is consistent across all the nodes in the cluster, and b)
>> try not to lose the ring info.
>>
>> The main driver to have only 1 ring compute node was a). the main concern
>> being guaranteeing consistency of the ring data among all nodes without
>> causing too strong coupling to the underlying mechanisms used to build the
>> ring.
>> For example - if 2 new rings are created independently, then the order in
>> which disks are added to the ring should be consistent (assuming that the
>> disk/partition allocation algorithm is sensitive to ordering). Which implies
>> that the query to chef should always return data in exactly the same order.
>> If also would require that the ring building (and mandate that it will
>> never be changed) does not use any heuristics that are time or machine
>> dependent (I _think_ that right now that is the case, but I would rather not
>> depend on it).
>>
>> I was thinking that these restrictions can be avoided easily by making
>> sure that only 1 node computes the ring. To make sure that b) (don't lose
>> the ring) is addressed - the ring is copied around.
>> If the ring compute node fails, then any other node can be used to seed a
>> new compute ring without any loss.
>> Does that make sense?
>>
>>
>> Right now I'm using a snapshot deb package built from bzr266. Changing the
>> source of the bits is pretty esay... (and installing the deb includes the
>> utilities you mentioned)
>>
>>
>> Re: load balancers:
>> What you're proposing makes perfect sense. Chef is pretty modular. So the
>> swift configuration recipe focuses on setting up swift - not the whole
>> environment. It would make sense to deploy some load balancer, firewall
>> appliance etc in an environment. However, these would be add-ons to the
>> basic swift configuration.
>> A simple way to achieve this would be to have a recipe that would query
>> the chef server for all nodes which have the swift-proxy role, and add them
>> as internal addresses for the load balancer of your choice.
>> (e.g. :
>> http://wiki.opscode.com/display/chef/Search#Search-FindNodeswithaRoleintheExpandedRunList
>> )
>>
>>
>> a.
>>
>>
>>
>> On Sun, May 1, 2011 at 10:14 AM, Jay Payne <letterj@xxxxxxxxx> wrote:
>>
>>> Andi,
>>>
>>> This looks great.   I do have some thoughts/questions.
>>>
>>> If you are using 1.3, do you have a separate role for the management
>>> functionality in the proxy?    It's not a good idea to have all your
>>> proxy servers running in management mode (unless you only have one
>>> proxy).
>>>
>>> Why only 1 ring-compute node?  If that node is lost or unavailable do
>>> you loose your ring-builder files?
>>>
>>> When I create an environment I always setup utilities like st,
>>> get-nodes, stats-report, and a simple functional test script on a
>>> server to help troubleshoot and manage the cluster(s).
>>>
>>> Are you using packages or eggs to deploy the swift code?   If your
>>> using packages, are you building them yourself or using the ones from
>>> launchpad?
>>>
>>> If you have more than three proxy servers, do you plan on using load
>>> balancers?
>>>
>>>
>>> Thanks
>>> --J
>>>
>>>
>>>
>>>
>>> On Sun, May 1, 2011 at 8:37 AM, andi abes <andi.abes@xxxxxxxxx> wrote:
>>> > Judd,
>>> >   Sorry, today I won't be around. I'd love to hear feedback and
>>> suggestions
>>> > on what I have so far ( I'm not 100% sure when I can make the fully
>>> > available, but I'm hoping this is very soon). I'm running with swift
>>> 1.3 on
>>> > ubuntu 10.10.
>>> > I'm using  the environment pattern in chef - when nodes search for
>>> their
>>> > peers a predicate comparing the node[:swift][:config][:environment] to
>>> the
>>> > corresponding value on the prospective peer. A "default" value is
>>> assigned
>>> > to this by the default recipe's attributes, so if only 1 cluster is
>>> present,
>>> > all nodes are eligible. For example, when proxy recipe creates the
>>> memcached
>>> > list of addresses, it searches for all the other nodes with the
>>> swift-proxy
>>> > role assigned to it anded with the above.
>>> >  Is that what you meant about having a classifier?
>>> > To the basic roles you've described (base, storage and proxy) I've
>>> added 1:
>>> > ring-compute. This role should be assigned to only 1 node on top of
>>> either
>>> > the storage or proxy. This server will recompute the rings whenever the
>>> set
>>> > of storage servers change (new disks or machines added or existing ones
>>> > removed). It uses information on the affected machines (ip, disk and
>>> zone
>>> > assignment) to update the ring. Once the ring is updated, it is rsynced
>>> to
>>> > all the other nodes in the cluster.
>>> > [ this is achieved with a LWRP as I mentioned earlier, which first
>>> parses
>>> > the current ring info, then compares it to the current set of disks. It
>>> > notifies chef if any changes were made, so that the rsync actions can
>>> be
>>> > triggered]
>>> > At this point, all disks are added to all rings. Though it should be
>>> easy to
>>> > make this conditional on an attribute on the node/disk or some
>>> heuristic.
>>> > The 'base role' is also a bit more extensive than your description, but
>>> > needs a bit more work. The recipe uses a swift_disk LWRP to ensure the
>>> > partition table matches the configuration. The LWRP accepts an array
>>> > describing the desired partition table, and allows using a :remaining
>>> token
>>> > to indicate using what's left (should only be used on the last
>>> partition).
>>> > At this point it, the recipe is pretty hard coded. It assumes that
>>> /dev/sdb
>>> > is dedicated to storage. it just requires a hard coded single
>>> partition,
>>> > that uses the whole disk. If it's not present, or different, the LWRP
>>> > creates a BSD label, a single partition and xfs filesystem.  Once this
>>> is
>>> > done, the available disk is 'published' as node attributes. If the
>>> current
>>> > state of the system matches the desired state, nothing is modified.
>>> > The proxy and storage roles, do as you'd expect. install the relevant
>>> > packages (including memcached on the proxy using the opscode recipe)
>>> and
>>> > plunk in the server/cluster specific info into the relevant config file
>>> > templates.
>>> > What's totally not yet addressed:
>>> > - starting services
>>> > - prep-ing  the authentication
>>> > - injecting disk configuration.
>>> > This cookbook can be used by it self, but it is more powerful when used
>>> in
>>> > conjunction with the crowbar proposal mechanism. crowbar basically
>>> allows
>>> > you to define the qualifications for a system to be assigned a given
>>> role,
>>> > edit the automatic role assignment and configuration of the cluster and
>>> then
>>> > materialize the cluster based on these values by driving chef. Part of
>>> the
>>> > planned crowbar capability is performing automatic disk/zone
>>> allocation,
>>> > based on network topology and connectivity. In the mean time, the
>>> allocation
>>> > is done when the storage node is created.
>>> > Currently,  a proposal can provide values for the following:
>>> > "cluster_hash",  "cluster_admin_pw", "replicas", and user/group to run
>>> swift
>>> > as. I'm really curious to hear what other configuration parameters
>>> might be
>>> > useful
>>> > I'm really curious to hear some feedback about this.
>>> > and sorry about the timing, I'm in the boston area and it's finally
>>> nice
>>> > around here... hence other weekend plans.
>>> >
>>> >
>>> > On Thu, Apr 28, 2011 at 11:25 PM, Judd Maltin <judd@xxxxxxxxxxxxxx>
>>> wrote:
>>> >>
>>> >> Hey andi,
>>> >> I'm psyched to collaborate on this.  I'm a busy guy, but I'm
>>> dedicating
>>> >> Sunday to this. So if you have time Sunday, that would be best to
>>> catch up
>>> >> via IRC, IM or voice.
>>> >> Having a node classifier of some sort is critical.
>>> >> -Judd
>>> >>
>>> >> Judd Maltin
>>> >> +1 917 882 1270
>>> >> Happiness is a straight line and a goal. -fn
>>> >> On Apr 28, 2011, at 11:59, andi abes <andi.abes@xxxxxxxxx> wrote:
>>> >>
>>> >> Judd,
>>> >>  Ok. Here are some of the thoughts I've had (and have  mostly working,
>>> but
>>> >> hitting some swift snags..) Maybe we can collaborate on this?
>>> >> Since Chef promotes idempotent operations and cookbooks, I put some
>>> effort
>>> >> in making sure that changes are only made if they're
>>> required, particularly
>>> >> around destructive or expensive operations.
>>> >> The 2 main cases are:
>>> >> - partitioning disks, which is obviously destructive
>>> >> - building / rebuilding the ring files - rebalancing the ring is
>>> >> relatively expensive.
>>> >> for both cases I've built a LWRP which reads the current state of
>>> affairs,
>>> >> and decides if and what are the required changes. For disks, I'm using
>>> '
>>> >> 'parted', which produces machine friendly output. For ring files I'm
>>> using
>>> >> the output from ring-builder.
>>> >> the LWRP are driven by recipes which inspect databag - very similar to
>>> >> your approach. However, they also utilize inventory information about
>>> >> available resources created by crowbar in chef during the initial
>>> bring up
>>> >> of the deployment.
>>> >> (As a side note - Dell has announced plans to opensource most of
>>> crowbar,
>>> >> but there's legalese involved)
>>> >>
>>> >> I'd be more than happy to elaborate and collaborate on this !
>>> >>
>>> >> a
>>> >>
>>> >>
>>> >>
>>> >>
>>> >>
>>> >>
>>> >>
>>> >> On Thu, Apr 28, 2011 at 11:35 AM, Judd Maltin <judd@xxxxxxxxxxxxxx>
>>> wrote:
>>> >>>
>>> >>> Hi Andi,
>>> >>>
>>> >>> Indeed, the swift recipes hadn't been updated since mid 2010, so I
>>> pushed
>>> >>> forward with my own.
>>> >>>
>>> >>> Thanks!
>>> >>> -judd
>>> >>>
>>> >>> On Thu, Apr 28, 2011 at 10:03 AM, andi abes <andi.abes@xxxxxxxxx>
>>> wrote:
>>> >>>>
>>> >>>> Judd,
>>> >>>>  this is a great idea... actually so great, that some folks @Dell
>>> and
>>> >>>> OpsCode, me included, have been working on it.
>>> >>>> Have a peek on
>>> >>>> :
>>> https://github.com/opscode/openstack-cookbooks/tree/master/cookbooks
>>> >>>> This effort is also being included into Crowbar (take a peek
>>> >>>> here: http://robhirschfeld.com/tag/crowbar/) which adds the steps
>>> needed to
>>> >>>> start with bare metal (rather than installed OS), then using chef to
>>> get to
>>> >>>> a working Open stack deployment.
>>> >>>> (if you're at the design meeting, there are demos scheduled).
>>> >>>> That said - I'm updating the swift cookbook, and hope to update
>>> github
>>> >>>> soon.
>>> >>>> a.
>>> >>>>
>>> >>>>
>>> >>>>
>>> >>>>
>>> >>>> On Wed, Apr 27, 2011 at 9:55 PM, Jay Payne <letterj@xxxxxxxxx>
>>> wrote:
>>> >>>>>
>>> >>>>> Judd,
>>> >>>>>
>>> >>>>> I'm not that familiar with Chef (I'll do some research) but I have
>>> a
>>> >>>>> couple of questions and some thoughts:
>>> >>>>>
>>> >>>>> 1.  Is this for a multi-server environment?
>>> >>>>> 2.  Are all your proxy nodes going to have
>>> "allow_account_management =
>>> >>>>> true" in the configs?   It might be a good idea to have a second
>>> proxy
>>> >>>>> config for account management only
>>> >>>>> 3.  Have you looked at using swauth instead of auth?
>>> >>>>> 4.  Have you thought about an admin or client node that has
>>> utilities
>>> >>>>> on it like st and stats-report?
>>> >>>>> 5.  How where will you do on-going ring management or changes?
>>> >>>>> 6.  I would think about including some type of functional test at
>>> the
>>> >>>>> end of the deployment process to verify everything was created
>>> >>>>> properly and that all nodes can communicate.
>>> >>>>>
>>> >>>>>
>>> >>>>>
>>> >>>>> --J
>>> >>>>>
>>> >>>>> On Wed, Apr 27, 2011 at 6:18 PM, Judd Maltin <judd@xxxxxxxxxxxxxx>
>>> >>>>> wrote:
>>> >>>>> > Hi Folks,
>>> >>>>> >
>>> >>>>> > I've been hacking away at creating an automated deployment system
>>> for
>>> >>>>> > Swift
>>> >>>>> > using Chef.  I'd like to drop a design idea on you folks (most of
>>> >>>>> > which I've
>>> >>>>> > already implemented) and get feedback from this esteemed group.
>>> >>>>> >
>>> >>>>> > My end goal is to have a "manifest" (apologies to Puppet) which
>>> will
>>> >>>>> > define
>>> >>>>> > an entire swift cluster, deploy it automatically, and allow edits
>>> to
>>> >>>>> > the
>>> >>>>> > ingredients to manage the cluster.  In this case, a "manifest" is
>>> a
>>> >>>>> > combination of a chef databag describing the swift settings, and
>>> a
>>> >>>>> > spiceweasel infrastructure.yaml file describing the OS
>>> configuration.
>>> >>>>> >
>>> >>>>> > Ingredients:
>>> >>>>> > - swift cookbook with base, proxy and server recipes.  proxy
>>> nodes
>>> >>>>> > also
>>> >>>>> > (provisionally) contain auth services. storage nodes handle
>>> object,
>>> >>>>> > container and account services.
>>> >>>>> > -- Base recipe handles common package install, OS user creation.
>>> >>>>> > Sets up
>>> >>>>> > keys.
>>> >>>>> > -- Proxy recipe handles proxy nodes: network config, package
>>> install,
>>> >>>>> > memcache config, proxy and auth package config, user creation,
>>> ring
>>> >>>>> > management (including builder file backup), user management
>>> >>>>> > -- Storage recipe handles storage nodes: network config, storage
>>> >>>>> > device
>>> >>>>> > config, package install, ring management.
>>> >>>>> >
>>> >>>>> > - chef databag that describes a swift cluster (eg:
>>> >>>>> > mycluster_databag.json)
>>> >>>>> > -- proxy config settings
>>> >>>>> > -- memcached settings
>>> >>>>> > -- settings for all rings and devices
>>> >>>>> > -- basic user settings
>>> >>>>> > -- account management
>>> >>>>> >
>>> >>>>> > - chef "spiceweasel" file that auto-vivifies the infrastructure:
>>> (eg:
>>> >>>>> > mycluster_infra.yaml)
>>> >>>>> > -- uploads cookbooks
>>> >>>>> > -- uploads roles
>>> >>>>> > -- uploads the cluster's databag
>>> >>>>> > -- kicks off node provisioning by requesting from infrastructure
>>> API
>>> >>>>> > (ec2 or
>>> >>>>> > what have you) the following:
>>> >>>>> > --- chef roles applied (role[swift:proxy] or role[swift:storage])
>>> >>>>> > --- server flavor
>>> >>>>> > --- storage device configs
>>> >>>>> > --- hostname
>>> >>>>> > --- proxy and storage network details
>>> >>>>> >
>>> >>>>> > By calling this spiceweasel file, the infrastructure can leap
>>> into
>>> >>>>> > existence.
>>> >>>>> >
>>> >>>>> > I'm more or less done with all this stuff - and I'd really
>>> appreciate
>>> >>>>> > conceptual feedback before I take out all the non-sense code I
>>> have
>>> >>>>> > in the
>>> >>>>> > files and publish.
>>> >>>>> >
>>> >>>>> > Many thanks!  Happy spring, northern hemispherians!
>>> >>>>> > -judd
>>> >>>>> >
>>> >>>>> > Judd Maltin
>>> >>>>> > T: 917-882-1270
>>> >>>>> > F: 501-694-7809
>>> >>>>> > A loving heart is never wrong.
>>> >>>>> >
>>> >>>>> >
>>> >>>>> >
>>> >>>>> >
>>> >>>>> > _______________________________________________
>>> >>>>> > Mailing list: https://launchpad.net/~openstack
>>> >>>>> > Post to     : openstack@xxxxxxxxxxxxxxxxxxx
>>> >>>>> > Unsubscribe : https://launchpad.net/~openstack
>>> >>>>> > More help   : https://help.launchpad.net/ListHelp
>>> >>>>> >
>>> >>>>> >
>>> >>>>>
>>> >>>>> _______________________________________________
>>> >>>>> Mailing list: https://launchpad.net/~openstack
>>> >>>>> Post to     : openstack@xxxxxxxxxxxxxxxxxxx
>>> >>>>> Unsubscribe : https://launchpad.net/~openstack
>>> >>>>> More help   : https://help.launchpad.net/ListHelp
>>> >>>>
>>> >>>
>>> >>>
>>> >>>
>>> >>> --
>>> >>> Judd Maltin
>>> >>> T: 917-882-1270
>>> >>> F: 501-694-7809
>>> >>> A loving heart is never wrong.
>>> >>>
>>> >>>
>>> >>>
>>> >>
>>> >
>>> >
>>>
>>
>>
>
>
> _______________________________________________
> Mailing list: https://launchpad.net/~openstack
> Post to     : openstack@xxxxxxxxxxxxxxxxxxx
> Unsubscribe : https://launchpad.net/~openstack
> More help   : https://help.launchpad.net/ListHelp
>
>

References