test: add test for float encoding rules

read_buffer/src/column/float.rs

@@ -1,5 +1,4 @@
-use std::cmp::Ordering;
-use std::mem::size_of;
+use std::{cmp::Ordering, mem::size_of};
 
 use arrow::{self, array::Array};
 
@@ -13,8 +12,13 @@ use crate::column::{RowIDs, Scalar, Value, Values};
 #[allow(clippy::upper_case_acronyms)] // TODO(edd): these will be OK in 1.52
 #[derive(Debug)]
 pub enum FloatEncoding {
+    // A fixed-width "no compression" vector of non-nullable values
     Fixed64(Fixed<f64>),
+
+    // A fixed-width "no compression" vector of nullable values (as Arrow array)
     FixedNull64(FixedNull<arrow::datatypes::Float64Type>),
+
+    // A RLE compressed encoding of nullable values.
     RLE64(RLE<f64>),
 }
 
@@ -253,31 +257,45 @@ impl std::fmt::Display for FloatEncoding {
     }
 }
 
-fn check_run_lengths_above(arr: &[f64], min_rl: usize) -> usize {
-    if min_rl < 1 || arr.len() < min_rl {
-        return 0;
-    }
-
-    let (mut rl, mut v) = (1, arr[0]);
-    let mut total_matching_rl = 0;
+fn rle_rows(arr: &[f64]) -> usize {
+    let mut v = arr[0];
+    let mut total_rows = 0;
     for next in arr.iter().skip(1) {
         if let Some(Ordering::Equal) = v.partial_cmp(next) {
-            rl += 1;
             continue;
         }
 
-        // run length was big enough to be considered
-        if rl > min_rl {
-            total_matching_rl += 1;
-        }
-
-        rl = 1;
+        total_rows += 1;
         v = *next;
     }
 
-    total_matching_rl
+    total_rows + 1 // account for original run
 }
 
+fn rle_rows_opt(mut itr: impl Iterator<Item = Option<f64>>) -> usize {
+    let mut v = match itr.next() {
+        Some(v) => v,
+        None => return 0,
+    };
+
+    let mut total_rows = 0;
+    for next in itr {
+        if let Some(Ordering::Equal) = v.partial_cmp(&next) {
+            continue;
+        }
+
+        total_rows += 1;
+        v = next;
+    }
+
+    total_rows + 1 // account for original run
+}
+
+/// A lever to decide the minimum size in bytes that RLE the column needs to
+/// reduce the overall footprint by. 0.1 means that the size of the column must
+/// be reduced by 10%
+pub const MIN_RLE_SIZE_REDUCTION: f64 = 0.3; // 30%
+
 /// Converts a slice of `f64` values into a `FloatEncoding`.
 ///
 /// TODO(edd): figure out what sensible heuristics look like.
@@ -290,12 +308,11 @@ fn check_run_lengths_above(arr: &[f64], min_rl: usize) -> usize {
 /// The encoding is chosen based on the heuristics in the `From` implementation
 impl From<&[f64]> for FloatEncoding {
     fn from(arr: &[f64]) -> Self {
-        // The total number of run-lengths to find in order to decide to RLE
-        // this column is in the range `[10, 1/10th column size]`
-        // For example, if the columns is 1000 rows then we need to find 100
-        // run lengths to RLE encode it.
-        let total_rl_required = 10.max(arr.len() / 10);
-        if check_run_lengths_above(arr, 3) >= total_rl_required {
+        // The number of rows we would reduce the column by if we encoded it
+        // as RLE.
+        let base_size = arr.len() * size_of::<f64>();
+        let rle_size = rle_rows(arr) * size_of::<(u32, Option<f64>)>(); // size of a run length
+        if (base_size as f64 - rle_size as f64) / base_size as f64 >= MIN_RLE_SIZE_REDUCTION {
             return Self::RLE64(RLE::from(arr));
         }
 
@@ -320,9 +337,11 @@ impl From<arrow::array::Float64Array> for FloatEncoding {
             return Self::from(arr.values());
         }
 
-        // TODO(edd) Right now let's just RLE encode the column if it is 50% NULL.
-        // and has at least 100 values in it.
-        if arr.len() >= 100 && arr.null_count() >= arr.len() / 2 {
+        // The number of rows we would reduce the column by if we encoded it
+        // as RLE.
+        let base_size = arr.len() * size_of::<f64>();
+        let rle_size = rle_rows_opt(arr.iter()) * size_of::<(u32, Option<f64>)>(); // size of a run length
+        if (base_size as f64 - rle_size as f64) / base_size as f64 >= MIN_RLE_SIZE_REDUCTION {
             return Self::RLE64(RLE::from(arr));
         }
 
@@ -332,7 +351,10 @@ impl From<arrow::array::Float64Array> for FloatEncoding {
 
 #[cfg(test)]
 mod test {
+    use std::iter;
+
     use super::*;
+    use arrow::array::Float64Array;
     use cmp::Operator;
 
     #[test]
@@ -357,6 +379,60 @@ mod test {
         assert_eq!(enc.size_raw(false), 56);
     }
 
+    fn rle_rows() {
+        let cases = vec![
+            (vec![0.0, 0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0], 9),
+            (vec![0.0, 0.0], 1),
+            (vec![1.0, 2.0, 1.0], 3),
+            (vec![1.0, 2.0, 1.0, 1.0], 3),
+            (vec![1.0], 1),
+        ];
+
+        for (input, exp) in cases {
+            assert_eq!(super::rle_rows(input.as_slice()), exp);
+        }
+    }
+
+    #[test]
+    fn rle_rows_opt() {
+        let cases = vec![
+            (vec![Some(0.0), Some(2.0), Some(1.0)], 3),
+            (vec![Some(0.0), Some(0.0)], 1),
+        ];
+
+        for (input, exp) in cases {
+            assert_eq!(super::rle_rows_opt(input.into_iter()), exp);
+        }
+    }
+
+    #[test]
+    fn from_arrow_array() {
+        // Rows not reduced
+        let input: Vec<Option<f64>> = vec![Some(33.2), Some(1.2), Some(2.2), None, Some(3.2)];
+        let arr = Float64Array::from(input);
+        let enc = FloatEncoding::from(arr);
+        assert!(matches!(enc, FloatEncoding::FixedNull64(_)));
+
+        // Rows not reduced and no nulls so can go in `Fixed64`.
+        let input: Vec<Option<f64>> = vec![Some(33.2), Some(1.2), Some(2.2), Some(3.2)];
+        let arr = Float64Array::from(input);
+        let enc = FloatEncoding::from(arr);
+        assert!(matches!(enc, FloatEncoding::Fixed64(_)));
+
+        // Goldilocks - encode as RLE
+        let input: Vec<Option<f64>> = vec![Some(33.2); 10];
+        let arr = Float64Array::from(input);
+        let enc = FloatEncoding::from(arr);
+        assert!(matches!(enc, FloatEncoding::RLE64(_)));
+
+        // Goldilocks - encode as RLE
+        let mut input: Vec<Option<f64>> = vec![Some(33.2); 10];
+        input.extend(iter::repeat(None).take(10));
+        let arr = Float64Array::from(input);
+        let enc = FloatEncoding::from(arr);
+        assert!(matches!(enc, FloatEncoding::RLE64(_)));
+    }
+
     #[test]
     // Test NaN behaviour when `FloatEncoder`s are used.
     //