Module alp 

ALP (Adaptive Lossless floating-Point) codec for f64 columns.

Most real-world float metrics originate as fixed-point decimals (e.g., 23.5, 99.99). ALP finds the optimal power-of-10 multiplier that converts them to integers losslessly: 23.5 × 10 = 235 round-trips exactly via 235 / 10 = 23.5. The resulting integers have tiny bit-widths and compress 3-6x better than Gorilla with SIMD-friendly bit-packing.

Values that don’t round-trip exactly (true arbitrary doubles, NaN, Inf) are stored as exceptions — their original f64 bits are preserved separately.

Wire format:

[4 bytes] total value count (LE u32)
[1 byte]  exponent e (power of 10: factor = 10^e, range 0-18)
[1 byte]  factor index f (combined encode/decode factor)
[4 bytes] exception count (LE u32)
[exception_count × 12 bytes] exceptions: (index: u32, value_bits: u64)
[N bytes] FastLanes-packed encoded integers (FOR + bit-pack)

Reference: “ALP: Adaptive Lossless floating-Point Compression” (Afroozeh et al., SIGMOD 2023)

Functions

alp_encodability

Check what fraction of values can be ALP-encoded at the best exponent.

decode

Decode ALP-compressed bytes back to f64 values.

encode

Encode a slice of f64 values using ALP compression.