...
- Need to take a serious look at the distribution, however. Will need to force bimodality of shardlet distribution, which will bank on some locality (temporal will be just fine)
- We tie our memory management of the server to the disk layout
- Specifically, we can't afford to have the memory be under utilized, but we won't benefit much if shardlets are dense
Fast (fine-grained, shard interleaved) logging + shardlets
- May be worth looking at some more, but back of the envelope numbers suggest that on top of the 10 s or so to restore a shard from the shardlet log another 2 minutes or so would be needed to replay the log on top of it
- A possible approach is to do shardlets are multiple granularities, but exploring this is probably only worth it if we feel we aren't inside our time budget for TTD against power failures