HI Alex, Thanks for taking a look. We have simplified the design since the conference. The Allocation and Anonymous modules where only helping us to control the offers sent to the frameworks. Now we think that Roles and Quota in Moses elegantly solve this problem and we could take advantage of it.
The current design does not propose Mesos Modules, the POC we demonstrated @ the mesoscon is slightly out of date in that respect. The current design only enforces that any Policy Engine implementation should honour certain REST apis. This also removes Consul out of the picture, but at Huawei our implementation would pretty much consider Consul or something similar. 1) Failure semantics I do agree it is not straight forward to declare that a DC is lost just because framework lost the connection intermittently. Probing the 'Gossiper' we would know that the DC is still active but not just reachable to us, In that case its worth the wait. If the DC in question is not reachable from everyother DC, only then we could come to such conclusion. 2) Can you share more details about the allocator modules. As mentioned earlier these modules are no longer relevant we have much simpler way to achieve this. 3) High Availability I think you are talking about the below section? "Sequence Diagram for High Availability (Incase of local datacenter failure) Very Similar to cloud bursting use-case scenario. " The sequence diagram only represents flow of events in case if the current datacenter fails and the framework needs to connect to a new one. It is not talking about the approach you mentioned. I will update doc couple more diagrams soon to make it more understandable. We would certainly like to have a federated K/V storage layer across the DCs which is why Consul was considered in the first place. 4) Metrics / Monitoring - probably down the line The experimental version of gossiper already queries the maser at a frequent interval and exchange it amongst them. Ultimately DC federation is a hard problem to solve. We have plenty of use cases which is why we wanted to reach out to the community, share our experience and build something that is useful for all of us. Regards, Dhilip On Wed, Jul 13, 2016 at 7:58 PM, Alexander Gallego <agall...@concord.io> wrote: > This is very cool work, i had a chat w/ another company thinking about > doing the exact same thing. > > I think the proposal is missing several details that make it hard to > evaluate on paper (also saw your presentation). > > > 1) Failure semantics, seem to be the same from the proposed design. > > > As a framework author, how do you suggest you deal w/ tasks on multiple > clusters, i.e.: i feel like there have to be richer semantics about the > task at least on the mesos.proto level where the state is > STATUS_FAILED_DC_OUTAGE or smth along those lines. > > We respawn operators and having this information may allow me as a > framework author to wait a little longer before trying to declare that task > as dead (KILLED/FAILED/LOST) if I spawn it on a different data center. > > Would love to get details on how you were thinking of extending the > failure semantics for multi datacenters. > > > 2) Can you share more details about the allocator modules. > > > After reading the proposal, I anderstand it as follows. > > > [ gossiper ] -> [ allocator module ] -> [mesos master] > > > Is this correct ? if so, you are saying that you can tell the mesos master > to run a task that was fulfilled by a framework on a different data > center? > > Is the constraint that you are forced to run a scheduler per framework on > each data center? > > > > 3) High availability > > > High availability on a multi dc layout means something entirely different. > So are all frameworks now on standby on every other cluster? the problem i > see with this is that the metadata stored by each framework to support HA > now has to spans multiple DC's. It would be nice to perhaps at the mesos > level extend/expose an API for setting state. > > a) On the normal mesos layout, this key=value data store would be > zookeeper. > > b) On the multi dc layout it could be zookeeper per data center but then > one can piggy back on the gossiper to replicate that state in the other > data centers. > > > 4) Metrics / Monitoring - probably down the line, but would be good to > also piggy back some of the mesos master endpoints > through the gossip architecture. > > > > Again very cool work, would love to get some more details on the actual > implementation that you built plus some of the points above. > > - Alex > > > > > > > > On Wed, Jul 13, 2016 at 6:11 PM, DhilipKumar Sankaranarayanan < > s.dhilipku...@gmail.com> wrote: > >> Hi All, >> >> Please find the initial version of the Design Document >> <https://docs.google.com/document/d/1U4IY_ObAXUPhtTa-0Rw_5zQxHDRnJFe5uFNOQ0VUcLg/edit?usp=sharing> >> for Federating Mesos Clusters. >> >> >> https://docs.google.com/document/d/1U4IY_ObAXUPhtTa-0Rw_5zQxHDRnJFe5uFNOQ0VUcLg/edit?usp=sharing >> >> We at Huawei had been working on this federation project for the past few >> months. We also got an opportunity to present this in recent MesosCon >> 2016. From further discussions and feedback we have received so far, we >> have greatly simplified the design. >> >> Also I see that no one assigned to this JIRA now could i get that >> assigned to myself ? It would be great to know if there is anyone willing >> to shepherd this too. >> >> I would also like to bring this up in the community Sync that happens >> tomorrow. >> >> We would love to hear your thoughts. We will be glad to see collaborate >> with you in the implementation. >> >> Regards, >> Dhilip >> >> >> Reference: >> JIRA: https://issues.apache.org/jira/browse/MESOS-3548 >> Slides: >> http://www.slideshare.net/mKrishnaKumar1/federated-mesos-clusters-for-global-data-center-designs >> Video : >> https://www.youtube.com/watch?v=kqyVQzwwD5E&index=17&list=PLGeM09tlguZQVL7ZsfNMffX9h1rGNVqnC >> >> > > > -- > > > > > > Alexander Gallego > Co-Founder & CTO >
