Boki
Boki is a research serverless runtime that exposes shared logs for durable application state and separates log ordering from read-side caching.
LogBooks support tagged reads over a total order; sequence numbers need not be consecutive. A metalog orders fully replicated shard prefixes using componentwise minima of storage progress. The engine returns an append's sequence number after observing that ordering decision. [1]
Reads track metalog progress to preserve monotonic reads and read-your-writes. A lagging index must catch up; a child invocation inherits its parent's progress for the same LogBook. Reconfiguration seals the current metalogs before starting a new term. [1]
Auxiliary data caches reconstructed state beside log records, but has no durability guarantee. Applications must supply consistent auxiliary values. A cache miss can require replaying earlier records; the auxiliary cache is not a recovery checkpoint. [1]
In the pinned implementation, TxnCommit appends a speculative commit record, then calls checkTxnCommitResult and returns a separate committed boolean. Conflict detection compares object write sets within the transaction's log interval. An overlapping earlier transaction only rejects the current one if that earlier transaction actually committed. Failed transactions are skipped during object replay. [2] [3]
Relations
- related
- Scalog, Delos, Tango, Snapshot isolation
Limitations
- The 2021 BokiStore transaction design provides snapshot isolation. A totally ordered log does not by itself make those transactions serializable. [1]
- The prototype replicates metalogs in DRAM and assumes a sequencer quorum remains alive. Losing all sequencer memory is outside that failure model. [1]
Citations
- Jia and Witchel. Boki: Stateful Serverless Computing with Shared Logs. SOSP 2021, pp. 691-707, sections 3-6. DOI: 10.1145/3477132.3483541.
- Boki sosp-release: slib/statestore/txn.go, TxnCommit; revision 8e8c9ca8586a120a2bf18b3a6ae5041da628fac9.
- Boki sosp-release: slib/statestore/log.go, checkTxnCommitResult and syncToBackward; revision 8e8c9ca8586a120a2bf18b3a6ae5041da628fac9.
Sources & caveats
- Source-based synthesis of the 2021 paper and the pinned sosp-release implementation; not a benchmark replication or a deployment recommendation.
- The auxiliary-data contract is from paper sections 3 and 4.4; source inspection covers transaction outcome and object replay, not an end-to-end fault-tolerance proof.