[Sheepdog] Object node affinity

[Liu Yuan]

On 05/10/2012 08:39 PM, Matthew Law wrote:

> On Thu, May 10, 2012 11:44 am, Liu Yuan wrote:
>> Nope, at least by design, cached objects will always be the freshest
>> one. Though those cached objects are supposed to be flushed to the
>> cluster storage and would survive the host node crash, but you will
>> still be risking yourself to lose those 'updates' that haven't been
>> flushed yet in a disaster that the disk never comes back. You can
>> analogize object cache with kernel's page cache, and difference is that
>> data is cached into local disk instead of memory.
> 
> I wonder what the most common read/write ratio is? -if most users have a
> read heavy workload (I suspect this is true) then wouldn't ensuring all
> writes are flushed to disk as soon as possible (and are only allowed to be
> read from cache when this is completed) still offer some advantage?
>


Yes, this is called writethrough cache, but this term is overloaded with
many things, probably should be called 'read cache', which write is
always through the cache to the lower storage. This is very useful for
APP which has a heavy read workload and doesn't risk data consistency.

It is reasonable to implement this feature for object cache and
considered to be implemented after object reclaim is implemented.

Thanks,
Yuan

 
> I can see that would impact heavily on write performance at the cost of
> data security, but then I'm not a developer :-)
> 
>