There isn't a whole lot of need for buffering since we've got all the tweaks in place that make our system as fast as it is, but we do have a couple of levels of buffering incase of network issues, downtime before a replica set member takes over as primary, etc. so that we can try to avoid any isolated failure from slowing down the processing that comes before that point in the work flow.
Working from the end backwards, the first (or last) buffer is a Go buffered channel running within the process on the router which feeds into the database. These channels work as a sort of queue between concurrent Go-routines, or "workers", which have a set amount of allocated space in memory. These are empty most of the time, but if there is a failure with pushing to the database they can start to fill up in order to not block the process before it in the workflow until the system recovers. Before those Go workers on each of the "Router" servers that push to the database is a Redis queue which basically serves as a holding pin in between the CDN servers and the database cluster, which is the data's first stop after leaving it's originating servers.
On those originating servers is the Go process which reads the logs and pushes the data to Redis on the routers, these also utilize buffered channels. So all these layers of buffers work to prevent a block in the workflow during momentary downtimes while the system recovers. However, each of these layers of buffering do have a limit and if they fill up will begin to block on the process before them. In the event of a major failure (such as all members of a replica set being down or the network between the CDN server and the Router being down) the Go process running on the CDN server will stop reading the logs at a position where it lands when it is unable to feed more logs into the channel and hold that position until the channel starts to be drained from the other side and space in the channel becomes available.
We did prototype for a while with fluentd as well as Cassandra and a few other technologies which we decided not to use in the end. The answer to why we chose a lot of what we did was to satisfy our requirements. The "Minimal Hardware with Redundancy And Scale" and "Minimal Software Layers" requirements came into play in regards to your question. We found that most off the shelf technologies are great at scaling, but they have a cost that would have required us to employ a lot of hardware. We were able to reach the speed we needed by writing our own Go client to ship off the logs to our cluster, and that also gave us the customizability and control to be able to make sure we weren't consuming all of the resources our servers needed during peak times to deliver content.
As for why we chose the database we did, while TokuMX is basically a drop in replacement for MongoDB, it has features that make it much more usable for our situation. Specifically, a compression rate of over 80% without negatively impacting our insert speed and document level locking. And because TokuMX works as a drop in replacement for MongoDB, we were able to use the MongoDB driver (mgo http://labix.org/v2/mgo) for Go which is just a really great tool. Additionally, we are now able to leverage the MongoDB aggregation framework which allows us to use the data in a way that builds very helpful aggregations of the data very quickly.
Working from the end backwards, the first (or last) buffer is a Go buffered channel running within the process on the router which feeds into the database. These channels work as a sort of queue between concurrent Go-routines, or "workers", which have a set amount of allocated space in memory. These are empty most of the time, but if there is a failure with pushing to the database they can start to fill up in order to not block the process before it in the workflow until the system recovers. Before those Go workers on each of the "Router" servers that push to the database is a Redis queue which basically serves as a holding pin in between the CDN servers and the database cluster, which is the data's first stop after leaving it's originating servers.
On those originating servers is the Go process which reads the logs and pushes the data to Redis on the routers, these also utilize buffered channels. So all these layers of buffers work to prevent a block in the workflow during momentary downtimes while the system recovers. However, each of these layers of buffering do have a limit and if they fill up will begin to block on the process before them. In the event of a major failure (such as all members of a replica set being down or the network between the CDN server and the Router being down) the Go process running on the CDN server will stop reading the logs at a position where it lands when it is unable to feed more logs into the channel and hold that position until the channel starts to be drained from the other side and space in the channel becomes available.