Mailing-List: contact user-help@couchdb.apache.org; run by ezmlm
Precedence: bulk
Reply-To: user@couchdb.apache.org
Received-SPF: neutral (athena.apache.org: local policy)
Message-Id: <05094CDC-C1F3-4028-9095-B061E263FD87@apache.org>
From: Jan Lehnardt <jan@apache.org>
To: user@couchdb.apache.org
In-Reply-To: 
 <5017258D295FBE41917880488689F7B80FA4F32E62@VCSSBS.visionarycs.local>
Content-Type: text/plain; charset=US-ASCII; format=flowed; delsp=yes
Content-Transfer-Encoding: 7bit
Mime-Version: 1.0 (Apple Message framework v930.3)
Subject: Re: A couple of couchdb questions.
Date: Tue, 24 Mar 2009 22:46:10 +0100
References: 
 <5017258D295FBE41917880488689F7B80FA4F32E62@VCSSBS.visionarycs.local>

Hi Gary,


On 24 Mar 2009, at 22:14, Gary Smith wrote:

> Hello,
>
> I'm working to implement a document warehouse for about 15m  
> documents.  These documents range between 10kb - 500kb (legacy  
> archived pdf's).  Currently we do this by maintaining a mysql  
> database with the document stored on a variety of servers  
> (consisting of about 6TB).  Most of the problems that we encounter  
> are a) backups and b) physical access to the documents as they are  
> on a private network.  Since not much changes backups aren't really  
> that much of a problem (but restoring is very slow).  We are now  
> looking to add documents to this regularly (about 10k per week).  So  
> we are looking to implement something new, or at least, more useful.
>
> So we thought about using Amazon S3 for storage but these documents  
> fall under HIPA constraints so we have decided to do this in house.
>
> Looking at couchdb, it pretty much does what we are looking to do.   
> We really only want to store a document and maybe some very basic  
> metadata (which we currently do by having both a PDF and a metadata  
> file).  Implementation doesnt look like a problem with the  
> documented API.

This sounds like a good application.


> So, the questions.
>
> I would like to break this down into multiple servers and  
> incorporate replication at the same time.  The document says that  
> pull is recommended over push but doesn't mention why.

Pull replication is faster, more reliable and is better at picking up  
failed replications again.


>  Does push replication require the slave (or other node) to accept  
> the put/post request as completed?

Not sure what you mean here.


> If we choose pull replication instead of push, I assume that this is  
> something we will need to crontab out to schedule it, or does it  
> have a background process that constantly syncs?  API looks like  
> just a single get request.

It's a POST request, but yeah, there's no automatism. See below.


> Either way, here is what we are looking to do at this time.  At two  
> seperate locations we will have multiple servers, setup in a master/ 
> master configuration.  We should not run into any conflicts are  
> updates are not allowed.  ID's are unique (MD5 checksum and some  
> other unique information).

Good.


> We wanted to use 4 servers at each location, partly because each  
> server has 4TB of space (actually 3TB of raid 5).  Each server will  
> hold files based upon the first digit of the MD5 checksum (0-3 on  
> server A, 4-7 on server B, 8-A on server C, and B-F on server D).   
> We were thinking of using Apache's URL rewriting to proxy the  
> request to the proper server.  This should work for both get/put/post.

This sounds sensible. Other proxies can handle that as well, but if  
your are comfortable with Apache httpd, no reason to not use it.


> We will also have the backup server at the second location (which  
> will be the active for their location) using the same ideology.
>
> What would be most useful is to be able to ensure that before a  
> commit is accepted on a server we could gaurentee that it has been  
> replicated to a second box.
>
> Any ideas or suggestions on that?

I think it is easiest when your application knows about the fact that  
there is no single server when doing a write. Reads can still be  
served through a single interface, but to make a proxy understand the  
replication protocol is probably not worth the effort.

Here's how I'd set it up: Your application sends a PUT request to your  
Apache proxy (or to the correct backend node directly), the document  
will end up on one of the four backend servers. Your application also  
knows about the backend nodes, the partition rules and can access them  
directly, in both locations. After the PUT request that creates the  
document completes in location 1 you send a POST request to the  
backend server in location 2 telling it to replicate from the node  
that the doc was written to in location 1 (issuing a pull  
replication). When this call returns successfully (if not, you just  
retry), you return to the original caller that wanted to save the  
document.

Downsides:

  - You need to deal with the fact that the communication between back  
ends might not work and decide what to do (save anyway, allowing  
reduced reliability and weak consistency) or deny the write because  
the consistency and availability requirements can no longer be  
guaranteed).

  - Your backend nodes must be public.


In the Future, CouchDB will be able to do the partitioning for you  
(solving the second downside). The first one can never be circumvented  
in a distributed system.

	
Cheers
Jan
--