On Tue, Aug 30, 2016 at 1:49 AM, Maciej Katafiasz < mkatafi...@purestorage.com> wrote:
> On 29 August 2016 at 16:39, Igor Cicimov <ig...@encompasscorporation.com> > wrote: > > On Tue, Aug 30, 2016 at 6:18 AM, Maciej Katafiasz > > <mkatafi...@purestorage.com> wrote: > >> Be aware though that DNS round-robin reduces the availability of the > >> entire setup, since there are no provisions in the protocol for the > >> eviction of dead nodes. So unless you're very sure there will never be > >> any in your DNS and also have the TTL set to some very low value, > >> multiple DNS records will defeat some of the care HAProxy takes to > >> ensure it only sends requests to backends that can service them. > > > > Hmmm, one would think though the backend health check and fail over > should > > take care of this ... or maybe not??? > > > > Anyway, in case you use something like Consul which I mentioned before to > > provide the DNS records, then Consul itself will remove the failed node > from > > the DNS record. > > Right, I missed the "independent healthchecks" in the original > description, in which case it'd work well enough (albeit a low enough > TTL value is still a concern). > > Cheers, > Maciej > > The way we designed the feature is more like a "server template" line which may be used to pre-configure in memory X servers sharing the same DNS resolution. In your case X=2. If you intend to have up to 10 servers for this service, simply set X to 10. HAProxy will use A records to create the servers and the health checks will ensure that the servers are available before sending them traffic. If a A record disappear from the response, the corresponding server will get down. If a new server is added and we provisioned less than X, then a new server is provisioned. This X "upper" limit is to ensure compatibility with all HAProxy features (such as hash LBing algorithms). Could you let me know if that meets your requirements? (we can still change this description). Baptiste
