FuzzyLog
FuzzyLog is a shared-log abstraction that represents updates as a partially ordered graph partitioned by data shard and geographical region.
Colors identify shards; each color has one ordered chain per region. Cross-region playback within a color respects causality, not a global total order. Causality across different colors needs additional application machinery. [1]
Within one region, multi-color appends are serializable, not necessarily strictly serializable. TXMap adds conflict validation and decision records; appending an intention alone is insufficient. [1]
Dapple implements the abstraction using chain replication and assumes battery-backed DRAM for durability. Client failure can block appends to affected colors. Recovery uses leases, fencing and a write-ahead log; competing recovery clients can livelock. [1]
The ZooKeeper comparison used DRAM and emulated renames with successive delete/create operations, unlike DappleZK's atomic renames. It is not an equivalent-operations comparison or a current product ranking. [1]
Relations
- related
- Tango, Causal consistency, Serializability
Citations
- Lockerman et al. The FuzzyLog: A Partially Ordered Shared Log. OSDI 2018, pp. 357-372; sections 3, 4.1, 4.4, 5 and 6.5.
- JLockerman/FuzzyLog README at 400a458c28524040b790e33f8e58809c60a948c5: build instructions and test resource limits.
Sources & caveats
- Full-paper synthesis, not a benchmark or power-loss reproduction. No performance ranking is inferred.
- The pinned README warns that concurrent tests can exhaust socket/file resources and suggests RUST_TEST_THREADS=1 or a test subset. The repository was inspected, not executed. [2]
- Code snapshot 400a458c28524040b790e33f8e58809c60a948c5 is a reproducible inspection reference, not a verified identification of the paper's experimental revision.