[Gluster-devel] FUSE nodes TTL and FORGET

[Emmanuel Dreyfus]

On Mon, Jun 18, 2012 at 01:09:59AM -0400, Raghavendra Gowdappa wrote:
> Wouldn't tinkering with TTL value achieve similar functionality? Or is 
> it that a fixed TTL value is too static to predict next lookup?

Rising TTL means there are fewer lookups, and therefore a client may not
notice immediatly that another client removed a file.

There is no FORGET TTL in FUSE protocol, therefore the filesystem cannot
hint the kernel about when it should forget nodes. It is hence the duty of
the kernel to guess if a node should be forgotten early or not. Obseving
lookup activity may give a hint: if the file has been lookued up 100 times
in a row without being forgotten, I guess the kernel can bet that it is
wise to prostpone forget a little bit.

-- 
Emmanuel Dreyfus
manu at netbsd.org