Chimp compression
Chimp is a lossless floating-point time-series compression method that encodes XOR residuals using compact leading-zero categories and selective trailing-zero removal.
Mechanism: The basic method uses two-bit cases and eight leading-zero categories. Its binary64 variant removes trailing zeros when their count exceeds six; otherwise those bits remain in the payload. [1] [2]
History: Chimp128 adds a bounded history and a low-bit lookup to find an earlier reference. It does not exhaustively compare all 128 predecessors on every input. Reference identifiers and lookup memory have costs. [1]
Counterexample: In the synthetic ledger, a residual with 16 leading and 7 trailing zeros costs 52 bits in the tail branch, versus 50 with an already reusable leading category. A threshold rule need not minimize each individual encoding decision. [4]
Data domain: The pinned Java decoder reserves the raw word 0x7ff8000000000000 as an end marker. That NaN payload cannot be treated as an ordinary sample by this decoder; the XOR operation itself does not impose that restriction. [3]
Relations
Citations
- Liakos, Papakonstantinopoulou and Kotidis, Chimp, PVLDB 15(11), 3058-3070 (2022)
- Chimp Java encoder: leading buckets and state updates, pinned 320d397
- Chimp Java decoder: NAN_LONG end-of-stream sentinel, pinned 320d397
- Hesela: exact XOR-window and restart accounting
Sources & caveats
- Paper DOI: 10.14778/3551793.3551852. Its evaluation uses 14 time-series and five other datasets with 1000-value blocks, not every telemetry population.
- Source inspection, not native execution, establishes the implementation-specific sentinel boundary. It is not a claim that every Chimp format rejects that word.
- The synthetic ledger excludes terminators and byte padding and does not implement Chimp128 or decode a stream. No paper benchmark or speedup is reproduced.