The data that we had (disclaimer: I started the project and left a couple of years ago) tolerated occasional loss of a small part of it, so there were nothing f

The data that we had (disclaimer: I started the project and left a couple of years ago) tolerated occasional loss of a small part of it, so there were nothing fancy there: "dirty" NIF-based circular buffers (courtesy of ecirca library) are dumped to binary and wrote on disk using basho's bitcask from time to time. Dumb and pretty efficient solution given that constraints.