Long-term storage

[fabxc]

Much like Prometheus, oklog does not intend to be a long-term storage for logs. Thanos is a aggregation and long-term storage solution for Prometheus based on arbitrary object storage. Experiences with Thanos have shown that it is well possible to use object storage as a highly cost-efficient real-time database (at the order of "real-time" for human interaction).

This should be even more valid for the logging use case, where the majority of reads are linear scans. Thus, object storage seems like a suitable and very portable solution to solve long-term storage for oklog as well. Depending on evaluated performance, it could even be feasible to shrink oklog's own retention to a few days to hours.

In general it would probably work based on a client to oklog's store nodes, that reads, deduplicates, and uploads data to the object storage. At query time, data from object storage and store nodes will be merged transparently.

This obviously will need a full design discussion.