A cluster must support both erasure codes and replicated data.
- Per-container setting
- Set on the container only if the container is empty (only at creation?)
- How do you avoid a race condition for metadata during a partition?
- Per-object setting
- A replicated indirection object (similar to a manifest) can point to an erasure coded location
- All requests into EC data will have an extra lookup
- How will the indirection object be safely created across the system?