SLOG (geo-replicated transactions)
SLOG is a research transaction system that combines per-region input logs, deterministic execution and data home regions to provide strict serializability.
A single-home transaction accesses granules assigned to one home region, possibly across several servers. SLOG-B acknowledges after the first region commits; SLOG-HA also waits for input-log replication to a configured number of nearby regions. Multi-home transactions need cross-region ordering. [1]
Remastering changes a granule's home and counter. Region logs can interleave differently, so a transaction's counter is checked before requesting each granule lock: an older counter aborts, a newer one waits, and a matching one permits the lock request. Checks repeat during execution. Waiting before locking leaves the granule available to the remaster operation. [1]
The prototype uses one ordering region for multi-home work. Cross-region Paxos is a more failure-tolerant alternative discussed in the paper, not the evaluated ordering deployment. The bounded counters include wrap-around detection; ordinary unbounded-counter reasoning is not a complete account of that representation. [1]
This transaction system is distinct from a ZFS separate intent-log device. The author repository contains an experimental rewrite of the paper implementation and warns against production use. [2]
Relations
- related
- Calvin, Strict serializability, Two-phase commit
Limitations
- Raw comparisons across separate codebases with different feature completeness are not like-for-like estimates of architectural performance. [1]
- The accessed granules must be identified before transaction processing, directly or through the techniques assumed by the design. [1]
Citations
- Ren, Li and Abadi. SLOG: Serializable, Low-latency, Geo-replicated Transactions. PVLDB 12(11), 1747-1761 (2019), sections 3-4.
- SLOG author repository: README, experimental rewrite and production-use warning; revision 2d71bd74f7db.
Sources & caveats
- Primary-source synthesis of the 2019 paper, DOI 10.14778/3342263.3342647, and the pinned author README. No SLOG execution, crash test or performance reproduction was performed.
- The repository's MIT license covers its code; the paper is linked rather than redistributed.
- The paper explicitly qualifies its Cloud Spanner comparison; the summary does not report its raw magnitudes as a controlled comparison. [1]