Re: [MooseFS-Users] High availability for master servers

[Piotr R. K. <pio...@mo...>]

> Hi,
> 
> I have a question about high availability for master servers in MooseFS Pro.
> 
> If two masters replicate changes in metadata, is this replication 
> synchronous or asynchronous?

Asynchronous. Synchronous replication is on our roadmap. Of course synchronous replications will mean that operations on metadata will be much slower, so we don't want to change it globally. Instead we plan to add flag for directory which will mean that all changes inside this directory should be replicate synchronously.

> I mean the situation that a user created, deleted or renamed a file 
> or appended data to a file just before the primary master for this 
> operation goes down. The operation was reported to the user as finished.
> 
> Does this operation always survive master failure and is visible 
> after another master takes over the responsibility of the failed master?
> Can latest filesystem operations be lost?

Theoretically yes. We write changes to socket connected to other masters first and then send positive answer to client, but kernel can of course change the order and send packet to client first and then computer can go down without sending changes to other masters.

> How it is in case of chunk server failure? Can latest write 
> operations be lost in this type of failure?

Write operations are repeated in this case (also in case of chunkserver malfunction). The only problem you may encounter in this scenario is size of the file. At the end of write session client sends new size of the file similarly to any other metadata change and when master goes down just after sending proper answer to the client, but before sending this change to other masters, then your file may be smaller than expected. Of course this is theoretic scenario - we never witnessed it practically.

> Thanks


Best regards,

-- 
Piotr Robert Konopelko
MooseFS Technical Support Engineer | moosefs.com <https://moosefs.com/>
> On 17 Sep 2015, at 2:27 pm, Arek Wojna <are...@in...> wrote:
> 
> Hi,
> 
> I have a question about high availability for master servers in MooseFS Pro.
> 
> If two masters replicate changes in metadata, is this replication 
> synchronous or asynchronous?
> I mean the situation that a user created, deleted or renamed a file or 
> appended data to a file just before the primary master for this 
> operation goes down. The operation was reported to the user as finished.
> 
> Does this operation always survive master failure and is visible after 
> another master takes over the responsibility of the failed master?
> Can latest filesystem operations be lost?
> 
> How it is in case of chunk server failure? Can latest write operations 
> be lost in this type of failure?
> 
> Thanks
> Arek