couchdb-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Ryan Ramage <ryan.ram...@gmail.com>
Subject Re: Update conflicts?
Date Wed, 06 Apr 2011 03:12:56 GMT
Luis,

Thats a lot to take in, but a quick suggestion.

Have a parent doc that looks like this:
{
    id: node1,
	type: node,
    location: blah,
}

and some 'children' docs that look like this

{
    id: 3232323323223-32323232322-3232,
	timestamp: 1299794532000,
    type: cpu,
    node: node1,
    cpu: 0.94,
    ccores: 4,
    acores: 4,
    cmemory: 4096,
    amemory: 1024
}

and
{
    id: 3232323323223-32323232322-3232,
	timestamp: 1299794532000,
    type: disk,
    node: node1,
    disk: 100000
}
and
{
    id: 433432323323223-3232323322332,
	timestamp: 1299794532000,
    type: netio,
    node: node1,
    in: 100,
    out: 200
}
and
{
    id: 323432423432534534-534534-543534534
	timestamp: 1299794532000,
    type: generic,
    node: node1,
    name: "foo",
    value: "bar"
}

create a status view
"node_status" : function (doc) {
	if (doc.type != 'node') {
		emit([doc.node, doc.type, doc.timestamp],null);
	}
}

This allows you to not have to ever update a doc. Just keep inserting.
Couchdb is good at that.




On Tue, Apr 5, 2011 at 7:20 PM, Luis Miguel Silva
<luismiguelferreirasilva@gmail.com> wrote:
> Thanks for your email Ryan.
>
> Let me give you some more information on what i'm trying to do...
> Essentially, i have to create a "sort of CMDB" system that stores, not only configuration
data, but also operational data (so...i guess you could call it a OMDB instead).
>
> Either way, my company develops a meta-scheduler that can be used for HPC or Cloud environments.
It will guarantee that your resources are used the best way possible, maximizing their usage,
based on the policies you set up in it.
>
> To do that, our software needs to be aware of how the environment looks and this is why
an OMDB piece is very important for us (as it allows us to store information on the environment).
>
> Also, our software talks with external resource managers by a protocol we developed more
than a dozen years ago called "WIKI" (not as in "wikipedia" but, WIKI as in the hawayan word
for fast). That protocol is heavily based around key/value pairs so this is one of the reasons
i was EXTREMELY excited to find out that, with CouchDB's "view" functionality, i would be
able to map document attributes to more meaningful attributes that our software understands
(i.e. map the document's "available_cores" attribute to "ccores" [the "consumable cores" parameter
our software understands]).
>
> Another important thing to notice is that resources can be off different types: node
(for bare metal nodes), vm (for vms running on nodes) and storage (we can actually have more
data types but those are enough to exemplify what i'm talking about).
>
> This is why i created those "big documents" instead of smaller ones!
> For instance, each document would represent an entire node (i.e. procs, memory, etc).
>
> So my idea was to have an external process initially populate the database with documents
representing ALL the nodes we are managing (hence why i started my benchmarks with 100K increments)
and OTHER external processes (i.e. other types of resource managers) would update individual
attributes in each document.
>
> Let's imagine a document with id "node01":
> These fields would be updated by an agent that collected some of the hardware specs:\
>        ccores: 4 // total cores on machine
>        acores: 4 // available cores on machine
>        cmemory: 4096 // total memory on machine
>        amemory: 1024 // available memory
>        cpuload: 94%
> This field would be updated by our storage resource manager:
>        GMETRIC["disk"]: 1000000
> And, for instance, these fields would be updated by a network resource manager:
>        GMETRIC["NETIO"]: { "in":100, "out":200 }
>
> So, as you can see, different processes would manage the same document (just different
attributes in it).
>
> And the REALLY cool thing about the Views is the fact that our customers could VERY easily
adapt the database so that it would store THEIR extra data and shove it in a generic parameter
that our software woulder understand [i.e. the GMETRIC parameters are generic metrics...).
>
> So, based on these requirements, do you have any suggestions on how we should store our
data (keeping its structure easy enough for external consumers to maintain it without having
to bust their heads figuring out the logic behind the document attributes)?? :o)
>
> Thank you!
> Luis Miguel Silva
>
> On Apr 5, 2011, at 6:45 PM, Ryan Ramage <ryan.ramage@gmail.com> wrote:
>
>> Luis,
>>
>> Having the rev is very important when you update a doc. It lets you
>> know that your piece of information is out of date. This is a good
>> thing....
>>
>> I am wondering if the way you are modeling your data is not leading
>> you to do this update with less chance of conflict. See if you can
>> break your docs into even smaller docs. For example, I noticed from a
>> prior post you had a lot of Arrays in your docs. If multiple processes
>> are changing that array, you might be better served by making each
>> element in the array a separate doc.
>>
>> Ryan
>>
>> On Tue, Apr 5, 2011 at 4:41 PM, Luis Miguel Silva
>> <luismiguelferreirasilva@gmail.com> wrote:
>>> More or less!
>>>
>>> The most common scenario will be:
>>> - two or more processes writing to the same document, but only to a
>>> specific attribute (not overwriting the whole document)
>>>
>>> If, by any chance, two processes overwrite the same field, i'm ok with
>>> the last one always winning.
>>>
>>> Thanks,
>>> Luis
>>>
>>> On Tue, Apr 5, 2011 at 4:26 PM, Robert Newson <robert.newson@gmail.com>
wrote:
>>>> "Ideally, we would be able to update without specifying the _rev, just
>>>> posting (or, in this case PUTting) to the document..."
>>>>
>>>> So you want to blindly overwrite some unknown data?
>>>>
>>>> B.
>>>>
>>>> On 5 April 2011 22:57, Zachary Zolton <zachary.zolton@gmail.com> wrote:
>>>>> Luis,
>>>>>
>>>>> Checkout _update handlers:
>>>>>
>>>>> http://wiki.apache.org/couchdb/Document_Update_Handlers
>>>>>
>>>>>
>>>>> Cheers,
>>>>>
>>>>> Zach
>>>>>
>>>>> On Tue, Apr 5, 2011 at 4:46 PM, Luis Miguel Silva
>>>>> <luismiguelferreirasilva@gmail.com> wrote:
>>>>>> Dear all,
>>>>>>
>>>>>> I'm trying to play around with updates and i'm bumping into some
problems.
>>>>>>
>>>>>> Let's image we have to clients that poll a document from the server
at
>>>>>> the same time and get the same _rev.
>>>>>> Then one of them updates the doc based on the _rev it got:
>>>>>> [root@xkitten ~]# curl -X PUT -d
>>>>>> '{"_rev":"3-0d519bcf08130bf784f3c35d79760740","hello2":"fred2"}'
>>>>>> http://localhost:5984/benchmark/test?conflicts=true
>>>>>> {"ok":true,"id":"test","rev":"4-03640ebafbb4fcaf127844671f8e2de7"}
>>>>>> Then another one tries to update the doc based on the same exact
_rev:
>>>>>> [root@xkitten ~]# curl -X PUT -d
>>>>>> '{"_rev":"3-0d519bcf08130bf784f3c35d79760740","hello3":"fred3"}'
>>>>>> http://localhost:5984/benchmark/test?conflicts=true
>>>>>> {"error":"conflict","reason":"Document update conflict."}
>>>>>> [root@xkitten ~]#
>>>>>>
>>>>>> Is there a way to avoid this?! (like...make the update just create
a
>>>>>> new _rev or something)??
>>>>>>
>>>>>> Ideally, we would be able to update without specifying the _rev,
just
>>>>>> posting (or, in this case PUTting) to the document...
>>>>>>
>>>>>> Thoughts??
>>>>>>
>>>>>> Thank you,
>>>>>> Luis
>>>>>>
>>>>>
>>>>
>>>
>

Mime
View raw message