Multiple Apache webdav servers sharing repositories via NFS

[Cheyenne Wills <ch...@gmail.com>]

I've done several searches, and have found old and conflicting responses to
the question of sharing a repository via NFS. So what is the current set of
concerns with sharing repositories using NFS among several web servers?

Here is the scenario, several hundred repositories shared via NFS to a
couple of webdav (Apache mod-svn) servers.  The Apache servers set up via a
round-robin DNS server (thus they are all sharing a common virtualhost
name).  All user access is via webdav (authentication and access controlled
by an Apache authentication handler).  The users will see a common hostname
for all repositories.  All the servers are "network close" to each other.

Assuming that the NFS is current (NFS 4) with subtree checking disabled (as
per the FAQ), are there any gotcha's or other concerns.

I currently have a solution using some custom Apache proxying that is
working, but I'm looking at trying to simplify my solution.

Re: Multiple Apache webdav servers sharing repositories via NFS

[Ryan Schmidt <su...@ryandesign.com>]

On Sep 7, 2013, at 19:24, Nico Kadel-Garcia wrote:

> > http://www.pbs.org/wgbh/evolution/library/01/6/l_016_08.html
> 
> *Do not do this* with write access repositories, is my advice from knowledge of NFS filesystems. Writes, and partial writes, can occur from one host for the next "commit" but not actually be seen on the next "round robin" server until a fairly arbitrary amount of time later, due to propagation delays inherent in NFS. There's a lot tht NFS can do well, but this is not one of those tasks.
> 
> I'd really encourage you, for commercial graide high availability or load distribution, to look at WanDisco's "Multi-Site" or to use a failover load balancer setup. If server 1 goes toes up, serer 2 is available, and handles all the traffic. But mixing and matching back and forth randomly as can happen with round-robin? That sounds like a dangerous idea likely to bite you at really, really bad moments due to simultaneous commits through two different NFS enabled front ends.

It sounds like a great idea, *if* you are using a SAN with a cluster filesystem, such as Apple's Xsan. NFS is not a cluster filesystem so that would not be a good idea.

Re: Multiple Apache webdav servers sharing repositories via NFS

[Nico Kadel-Garcia <nk...@gmail.com>]

> http://www.pbs.org/wgbh/evolution/library/01/6/l_016_08.html

*Do not do this* with write access repositories, is my advice from
knowledge of NFS filesystems. Writes, and partial writes, can occur from
one host for the next "commit" but not actually be seen on the next "round
robin" server until a fairly arbitrary amount of time later, due to
propagation delays inherent in NFS. There's a lot tht NFS can do well, but
this is not one of those tasks.

I'd really encourage you, for commercial graide high availability or load
distribution, to look at WanDisco's "Multi-Site" or to use a failover load
balancer setup. If server 1 goes toes up, serer 2 is available, and handles
all the traffic. But mixing and matching back and forth randomly as can
happen with round-robin? That sounds like a dangerous idea likely to bite
you at really, really bad moments due to simultaneous commits through two
different NFS enabled front ends.


On Wed, Sep 4, 2013 at 10:49 PM, Cheyenne Wills <ch...@gmail.com>wrote:

> I've done several searches, and have found old and conflicting responses
> to the question of sharing a repository via NFS. So what is the current set
> of concerns with sharing repositories using NFS among several web servers?
>
> Here is the scenario, several hundred repositories shared via NFS to a
> couple of webdav (Apache mod-svn) servers.  The Apache servers set up via a
> round-robin DNS server (thus they are all sharing a common virtualhost
> name).  All user access is via webdav (authentication and access controlled
> by an Apache authentication handler).  The users will see a common hostname
> for all repositories.  All the servers are "network close" to each other.
>
> Assuming that the NFS is current (NFS 4) with subtree checking disabled
> (as per the FAQ), are there any gotcha's or other concerns.
>
> I currently have a solution using some custom Apache proxying that is
> working, but I'm looking at trying to simplify my solution.
>