some qs about the thin client

[Justin Mason]

hi all --

1.6.0 looks good -- we'll try it out soon ;)  I'm curious about the new thin client implementation -- how does consistency resolution work in this release?

What is the typical deployment model for the thin-client coordinator service -- should it live on the same hosts as the client fleet, the server fleet, or a separate fleet entirely?

Finally, could you be more specific about the critical client-side bugfixes in 1.6 which Vinoth noted in that mail earlier?

cheers!

--j.

[Vinoth C]

Hey Justin,

Replies inline

On Monday, February 3, 2014 1:57:25 AM UTC-8, Justin Mason wrote:

hi all --

1.6.0 looks good -- we'll try it out soon ;)  I'm curious about the new thin client implementation -- how does consistency resolution work in this release?

If you are referring to Conflict resolution,  it happens on the coordinator and it defaults to Vector Clock and timestamp based chained resolver, as before. if you want your own resolving on the thin client, there is a  X-VOLD-Inconsistency-Resolver header you can set on your REST requests.. It will bypass conflict resolution on the coordinator.

What is the typical deployment model for the thin-client coordinator service -- should it live on the same hosts as the client fleet, the server fleet, or a separate fleet entirely?

thin client would replace the current DefaultStoreClient/ZenStoreClient. There would be a separate tier or Coordinator services (We use D2, a Linkedin open source project for load balancing here. But you can choose whatever works for you), then the servers..
The idea here is to shrink 100s of clients to < 10 coordinator boxes (just a ballpark example). Hope that helps..
 

Finally, could you be more specific about the critical client-side bugfixes in 1.6 which Vinoth noted in that mail earlier?

Well, if you are referring to my response on upgrading from 0.8 to 1.6, some things we have fixed are  : more faithfully registering slops, Failure detector is now more accurate.
ZenStoreClient should help keep the servers and clients in sync ..

cheers!

--j.

[Justin Mason]

On Mon, Feb 3, 2014 at 4:02 PM, Vinoth C <mail.vino...@gmail.com> wrote:

if you want your own resolving on the thin client, there is a  X-VOLD-Inconsistency-Resolver header you can set on your REST requests.. It will bypass conflict resolution on the coordinator.

and the thin-client-calling code can then perform resolution in the client, as usual?

thanks,

--j.

[Vinoth C]

The "as usual resolving" happens on the coordinator.. Thin client only sees the resolved value (unless you set the header I was mentioning)

Here is some sample config, if you want to start playing with it.

https://github.com/vinothchandar/voldemort/tree/master/config/coordinator_sample_config