SmartStack vs etcd

Within a health check interval’s delay of a backend becoming healthy, it is made available in Zookeeper;this makes it instantly available to consumers via Synapse’s Zookeeper watches.;We detect problems within a health check interval, and take backends out of rotation. A mechanism which allows services to notify Nerve that they’re not healthy is planned, to reduce the interval further. In the meantime, deploys can stop Nerve when they start, and then re-start it at the end.;Synapse acts on information the moment it’s published in Zookeeper, and reconfiguring HAProxy is very very fast most of the time. Because we utilize HAProxy’s stats socket for many changes, we don’t even restart the process unless we have to add new backends.;Because our infrastructure is distributed, we cannot do centralized planning. But HAProxy provides very configurable queueing semantics. For our biggest clients, we set up intelligent queueing at the HAProxy layer;for others, we at least guarantee round-robin.;Doing debugging or maintenance on a backend is as simple as stopping the Nerve process on the machine;nothing else is affected!;You can see exactly which backends are available simply by looking at the HAProxy status page. Because of HAProxy’s excellent log output, you also get amazing aggregate and per-request information, including statistics on number of behavior of requests right in rsyslog.;The infrastructure is completely distributed. The most critical nodes are the Zookeeper nodes, and Zookeeper is specifically designed to be distributed and robust against failure.