Struct AttrPred 

struct AttrPred<'a> {
    op: Op,
    strs: Option<(&'a [u32], Vec<bool>)>,
    ints: Option<(&'a [i64], i64)>,
    doubles: Option<(&'a [f64], f64)>,
    bools: Option<(&'a BooleanArray, bool)>,
}

One attribute term with everything that does not depend on the row resolved once: the four value columns downcast, the query scalar canonicalized, and — for the string column — the predicate already evaluated against the dictionary.

This used to be attr_matches, which did all of it per row: two downcast_refs, a String allocation for canon, and a parse::<f64>() on every ordered comparison. That measured 30.9 ns per root row against 6.1 for the allocation-free integer arm, which made an attribute filter fifteen times dearer than a column one on the same block.

Fields

op: Op

strs: Option<(&'a [u32], Vec<bool>)>

The str column’s dictionary codes, and which dictionary entries satisfy the predicate.

The same trick as field_filter’s Dictionary arm: attribute values are where telemetry repeats (see crate::schema::ATTRS), so a few hundred string comparisons replace one per row and the row loop never touches string data.

ponytail: evaluated over the whole dictionary, which every key in the table shares — so filtering on a rare key in a block whose values are nearly all distinct pays one comparison per distinct value to reject a handful of rows. The ceiling is a value column with no repetition in it, which is the case its dictionary encoding is already the wrong layout for; the upgrade path is a lazily filled memo over the same array.

ints: Option<(&'a [i64], i64)>

doubles: Option<(&'a [f64], f64)>

bools: Option<(&'a BooleanArray, bool)>

Implementations

impl AttrPred<'_>

fn new<'a>(a: &'a RecordBatch, op: Op, v: &Value) -> AttrPred<'a>

fn test(&self, ty: u8, row: usize) -> bool

Compare one attribute row against the query scalar, dispatching on the stored type rather than on the query’s — the column decides what it is.

Empty, Bytes, Slice and Map are returned in results but not filterable in V0, and a column whose type does not match the schema reads the same way: no match.