Hyder
Hyder is a research transaction engine that validates updates against a shared, ordered log containing a multiversion database.
Storage: Servers cache parts of a tree-shaped database whose persistent representation is the log itself. Transactions run against snapshots and append intention records containing their proposed changes. [1]
Commit: Appending an intention does not decide whether it commits. Each server runs deterministic meld in log order to validate conflicts and incorporate accepted changes. A rejected intention has no effect on the committed database state. [2]
Isolation: Snapshot-based execution does not itself select snapshot isolation. Conflict checking depends on the requested isolation level; serializable execution must also protect read dependencies and key ranges. [1]
Scaling: Every server must process the update stream. Log bandwidth, networking, validation work and conflicting transactions bound scale-out. The later pipeline uses premeld and group meld to reduce work before a still-sequential final meld stage. [1] [2]
Relations
Limitations
- The 2015 evaluation focuses on throughput and leaves failed or unanswered log operations outside its scope; it is not a fault-tolerance validation. [2]
Citations
- Bernstein, Reid and Das. Hyder: A Transactional Record Manager for Shared Flash. CIDR 2011, sections 1, 4 and 5.
- Bernstein, Das, Ding and Pilman. Optimizing Optimistic Concurrency Control for Tree-Structured, Log-Structured Databases. SIGMOD 2015.
Sources & caveats
- Source synthesis, not a native Hyder execution or benchmark.
- CIDR 2011 combines component measurements and simulations; section 5 explicitly excludes an integrated end-to-end flash prototype. [1]
- The 2015 experiments use a CORFU log on SSD-backed servers and a modified YCSB workload. Their results concern that setup, not arbitrary storage hardware. [2]