CORFU
CORFU is a distributed shared-log design that separates allocation of log positions from replicated writes to a flash cluster.
Clients map log positions to replica sets using a projection. A sequencer allocates positions without carrying the payload; obtaining a token does not complete an append. [1]
Client-driven chain replication writes replicas in order. A reader that does not know whether an append completed must consult the chain tail. A stalled writer can leave a hole: fill completes an existing prefix value or writes a junk marker if the first replica is unwritten. [1]
Hole filling is not a perfect failure detector; a slow token holder can still compete with a filler. Write-once semantics at the first replica prevent both values from winning. An unwritten tail does not prove that the earlier replicas contain no data; filling may complete their existing value. [1]
Reconfiguration seals the old projection before installing its successor, rejecting stale operations on affected units. Sequencer allocation alone is not the safety mechanism. [1]
The sequencer reduces contention but is not required for safety or progress in the paper's design. Reconfiguration still needs a durable, write-once sequence of projections, the auxiliary, for which the paper lists a disk volume, a Paxos state machine or a static-projection CORFU instance. Removing the sequencer from the payload path does not remove that coordination dependency. [1]
Benchmark scope: section 5.2 uses 32 server-attached SSDs, but emulates writes to avoid wear. Reads access prewritten SSD data. The roughly 180,000 appends/s ceiling for 4 KB entries in Figure 7 is not measured durable-flash write throughput. [1]
Relations
Citations
Sources & caveats
- Historical design synthesis from the full NSDI 2012 paper; no hardware reproduction or current-product performance claim.