The question
Can compaction policies reduce write amplification under changing workloads while respecting read, space and tail-latency constraints?
This is a proposed research direction. No novel algorithm, completed benchmark or performance gain is claimed.
Why it matters
Embedded databases and streaming state backends could reduce I/O and write stalls.
Search space
File selection, merge timing, level sizing and workload-aware policy switching.
What would count as progress?
- Correct reads and recovery
- write/read/space amplification
- write stalls
- p99 latency
- policy overhead.
A useful result improves a measured trade-off against strong baselines on held-out workloads, with reproducible code and documented limitations.
Baselines and boundaries
Compare tuned leveled and universal compaction. Include long steady-state runs, deletes, range scans and workload shifts; do not hide deferred work beyond the benchmark horizon.
Starting reference
Where you could contribute
Help define representative, legally shareable workloads; critique the correctness contract; implement baselines; propose candidates; or reproduce measurements. Negative results and failure cases are useful contributions.
A detailed research brief is being prepared. Access requests will become available when that brief and the request service are ready.