parallax · research · also Parallax key-value store

Parallax

Parallax is a research LSM key-value store that combines in-place records, a transient medium-record log and a garbage-collected large-record log.

Placement: Small records stay in the index. Medium records begin in L0 and its recovery log, enter a sorted transient log during compaction, then return in place at a lower level so whole transient segments can be reclaimed. Large records still require log garbage collection. [1]

Classification: Revision d4aae58 uses min(key bytes, value bytes) / max(key bytes, value bytes), plus an in-place size cap. With the cap inactive, 8-byte keys and 120-byte values are medium, while 32-byte keys and 96-byte values are small: equal 128-byte totals need not share a category. Both 0.02 and 0.2 belong to the medium interval. [2]

Configuration: The implementation moves medium-log records back in place at the configured level_medium_inplace, rather than assuming that every deployment uses the last level. The checked options file sets this level to 3. [3] [4]

Relations

related
Key-value separation, Value log, Log-structured merge-tree, Kreon

Limitations

  • The paper's recovery contract allows loss of the latest acknowledged writes. Recovery to a consistent point is not a promise that every returned write is durable. [1]

Citations

  1. Xanthakis et al. Balancing Garbage Collection vs I/O Amplification using hybrid Key-Value Placement. arXiv:2106.03840v2, sections 2-5. (accessed 2026-11-15)
  2. Parallax author source d4aae58: calculate_KV_category in lib/btree/btree.c. (accessed 2026-11-15)
  3. Parallax author source d4aae58: medium-record transitions in compaction_worker.c. (accessed 2026-11-15)
  4. Parallax author source d4aae58: options.yml. (accessed 2026-11-15)
  5. Parallax: Hybrid Key-Value Placement in LSM-based Key-Value Stores. SoCC 2021, pp. 305-318. DOI:10.1145/3472883.3487012. (accessed 2026-11-15)
Sources & caveats
  • Design and evaluation scope use the June 2021 preprint, not an assumed byte-identical copy of the final SoCC paper. [1] [5]
  • The evaluation used one server, a 375-GB Optane P4800X and RocksDB/BlobDB 6.11.4 with Bloom filters disabled. Results are not current-engine rankings. [1]
  • Source inspection of revision d4aae5890630bb33405f3a9e03cc4e4978133ecf; no engine benchmark, power-loss test or recovery reproduction was performed.

JSON: /data/terms/parallax.json

Published by Hesela team · Updated

Report a correction · Editorial standards · Reuse guidance

Cite this page
Hesela team. Parallax. Hesela. 2026-11-15. https://hesela.dev/terms/parallax/