hesela.devLazyBase
LazyBase is a research database that batches updates while letting read queries trade result freshness for query cost.
The EuroSys 2012 prototype applies self-consistent updates (SCUs) atomically and gives each query a consistent snapshot. An old snapshot need not mix parts of an atomic batch. Updates supply replacement values; the service does not support general read-modify-write transactions. [1]
Queries read the merged authority plus pipeline batches needed for their freshness requirement. Sorted batches have indexes; unsorted batches require scans. Relaxing freshness can avoid those scans, without promising a fixed speedup. [1]
The FAST 2014 production design uses independent filesystem-segment batches and disables client-requested freshness in the early product. FreshnessComplete marks the boundary before which every segment's events have been observed; FreshnessPartial marks the latest event processed anywhere. Coverage between them is incomplete, so the upper timestamp is not a completeness watermark for incremental backup. [2]
That design also removed auto-increment ID remapping because it could destroy pathname sort locality, require joins and scans, and serialize ingestion. [2]
Illustration: with completeness at 10:00 and the latest event at 10:05, another segment's 10:03 event may still be absent. Advancing a backup cursor to 10:05 can skip it. This example is constructed, not a measurement.
Relations
Citations
- Cipar et al. LazyBase: Trading Freshness for Performance in a Scalable Database. EuroSys 2012, 169-182, sections 3-4. DOI: 10.1145/2168836.2168854.
- Johnson et al. From Research to Practice: Engineering a Production Metadata Database for a Scale Out File System. FAST 2014, 191-198, sections 3.1-3.3.
Further reading
Sources & caveats
- Source synthesis of the research prototype and a separate historical production design, not a benchmark reproduction or a claim of current product availability.
- The FAST 2014 system is StoreAll Express Query. The EuroSys 2012 evaluation uses tweets and synthetic data on its stated hardware/Cassandra configuration; the reported comparisons do not establish performance against current releases. [1] [2]