... protocols.¹

We do have a checkpoint mechanism (discussed later) that permits us to recover in the case that any of these cluster properties fail, however all state changes that happen after the last checkpoint will be lost should this occur.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... maps²

It is important to use large enough of a hash to make the probability of collision negligible; we currently use 32 bits.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... disk.³

Write throughput is less than read throughput because a hash bucket must be read before it can written, in case there is already data stored in that bucket that must be preserved. There is therefore an additional read for every write, nearly halving the effective throughput for DDS writes.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... project⁴

https://ninja.cs.berkeley.edu/

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.