Struct WorkloadAnalyzer 

pub struct WorkloadAnalyzer { /* private fields */ }

Tracks historical query logs and exposes heuristics for deciding which predicates warrant a dedicated SIEVE subindex.

Implementations

impl WorkloadAnalyzer

pub fn new(stable_fraction_threshold: f32) -> Self

Create a new analyzer. stable_fraction_threshold is the minimum fraction of total queries a predicate must represent to be considered stable.

pub fn record(&mut self, signature: PredicateSignature, timestamp_secs: u64)

Record a query that used the given predicate signature.

pub fn stable_predicates(&self) -> Vec<(PredicateSignature, f32)>

Returns (signature, frequency) pairs sorted by frequency descending, filtered to those whose frequency exceeds stable_fraction_threshold.

Frequency is the fraction of total log entries contributed by that predicate. Returns an empty Vec if the log is empty.

pub fn estimate_subindex_cost( &self, vectors_in_subindex: usize, dim: usize, ) -> (usize, f32, f32)

Rough 3D cost model for a candidate subindex.

Returns (memory_bytes, latency_ms, recall) estimates:

• memory_bytes ≈ n * (dim * 4 + sub_m * 4 * avg_layers) where avg_layers ≈ 2.
• latency_ms ≈ log2(n) * 0.01.
• recall = 0.95 (assumed for a correctly built HNSW index).

These are intentionally coarse: they guide the build/no-build decision for a planner, not a precision benchmark.

pub fn compact(&mut self, now_secs: u64, retention_secs: u64)

Drop all log entries older than retention_secs seconds before now.

Entries with timestamp_secs < now - retention_secs are removed. Entries at or after that boundary are kept.