instead of having time-based naming, block manager will perform occasional compaction at startup time and delete unwanted blocks
The protocol is safe from concurrency standpoint and provides eventual consistency.
We only delete blocks after creating compacted blocks.
We retry loading index until (any) consistent index is fetched (possibly from cache) and all underlying blocks are also available, not necessarily the latest ones.
TODO - we need to periodically snapshot block index contents, so that if we have a bug somewhere in compaction code, we have a way of restoring working indexes.
