I have not analysed cost so far.

I did some analysis for https://mesosphere.com/blog/2015/11/30/arangodb-benchmark-dcos/ but was essentially only interested in maximizing throughput per vCPU of single document operations. I found that for that the sweet spot was using instances with 8 vCPUs and fast local SSDs, all running a primary DBServer, a secondary DBServer (asynchronous replica) and a coordinator. This was for AWS, but I would expect similar results for DO.

The best way to cut costs is almost certainly by not using local SSDs, since they make the instances expensive with all providers. However, this will almost immediately cost throughput, simply because the combined I/O performance of the instances cannot keep up. Note that we try to avoid write amplification as much as possible, but there is always some overhead, in particular since we have to write every document once to the write ahead log of the primary, once to the actual data file, and then the same for the asynchronous replica. If the combined I/O performance is the bottleneck, then one can often buy instances with less CPU power but the same I/O bandwidth and cut costs in this way.

As to reliability, Version 3.0 will greatly increase reliability because of the synchronous replication we are currently putting in.

For a real world application, I would first specify the replication and reliability needs, then specify the needed throughput with some reserves, which essentially will tell you whether or not you need SSDs. Then I would simply compare clusters with different sizes and numbers of droplets with the intended load. This should give you the sweet spot w.r.t. costs for your particular needs.

Thanks