Re: Watchman support for git

From: David Turner <dturner@twopensource.com>
Date: May 9, 2014, 18:47 UTC
Message-ID: <1399661273.11843.133.camel@stross>
In-Reply-To: <alpine.DEB.2.02.1405091124150.5876@nftneq.ynat.uz>

On Fri, 2014-05-09 at 11:27 -0700, David Lang wrote:

Show 16 quoted lines

> >
> > That's not my understanding from Durham Goode's talk in January.  Yes,
> > operations involving history go to the server.  But the client also
> > maintains a copy of the working tree, and it is for this that watchman
> > is used.  Otherwise, why bother with watchman at all?  The server knows
> > when it changes files and could simply maintain its own index of what's
> > changed.  Watchman is built on inotify/fsevents -- it doesn't have
> > anything to do with any sort of storage device beyond a vanilla hard
> > drive.
> 
> When you have such a massive repo, your clients aren't storing the data on their 
> local drives, they are accessing the data on a network attached storage (via NFS 
> or through a fuse mount). So they can have their watchman send queries to the 
> storage server to find out what has changed in this massive repo rather than 
> having to walk the directory tree (or try to monitor it for changes on the 
> client machine)

Yes, you could do that. But I repeat: that is not what Facebook is actually doing. If it were, they would have no need for inotify or FSEvents, since neither even works with NFS (and a FUSE-based solution would need special support at which point it might as well just implement watchman itself).

Please see this section of Durham Goode's talk for evidence that Facebook is not doing that: http://www.youtube.com/watch?v=Dlguc63cRXg&t=10m40s

(it's at 10:40 in case that link doesn't work)

Maybe things have changed since at Facebook since Durham gave that talk in January and you have knowledge of that. If so, say so. But from what Facebook has said publicly, this is simply not what's going on.