feat: CountSketch + KLL + DDSketch + HLL decoders for modified-OTLP hot path (PR C)

[zzylol]

PR C — completes the per-sketch-type decoder roll-out for the Phase 1 modified-OTLP sketch hot path (PR #5 + #6 + #7). With this PR, all five modified-OTLP Metric.data sketch variants — CountMin, CountSketch, KLL, DDSketch, HLL — reach concrete accumulators end-to-end instead of falling through to the §5.2 fallback.

Scope

Variant	Accumulator	sketch-core type	Decoder	E2E test
CountMin	CountMinSketchAccumulator (PR #6)	existing count_min::CountMinSketch	existing	existing
CountSketch	new CountSketchAccumulator	new count_sketch::CountSketch	new	new
KLL	existing DatasketchesKLLAccumulator	existing kll::KllSketch	new (lossy statistical reconstruction)	new
DDSketch	new DDSketchAccumulator	new dd_sketch::DdSketch	new	new
HLL	new HllSketchAccumulator	new hll_sketch::HllSketch	new	new

What's new

sketch-core types

• sketch_core::count_sketch::CountSketch — element-wise-mergeable signed-counter matrix.
• sketch_core::dd_sketch::DdSketch — log-bucketed quantile sketch; merge aligns bucket arrays on absolute store_offset and sums counts element-wise, combining {count, sum, min, max} aggregates.
• sketch_core::hll_sketch::HllSketch — register array + variant + precision + HIP accumulators; merge is register-wise max with HIP-component addition when variant == Hip.

QE accumulators (precompute_operators/)

• CountSketchAccumulator, DDSketchAccumulator, HllSketchAccumulator — each implements the full AggregateCore + SerializableToSink surface and a from_sketchlib_proto_bytes constructor that decodes the corresponding asap_sketchlib::proto::sketchlib::*State proto emitted by DataCollector's per-sketch processors.
• DatasketchesKLLAccumulator::from_sketchlib_proto_bytes — decodes KllState via lossy statistical reconstruction (replays items[] through update() on a fresh sketch with the same k, since sketch-core's KLL backend types keep their level structure private). Quantile estimates remain approximately equivalent to the source's — within KLL's own rank-error bound, which the source already inherited — so queries hitting the reconstructed sketch return answers the user would already have accepted from the source. Bit-identical reconstruction is tracked as a sketchlib upstream follow-up.

Dispatcher (drivers/ingest/otel.rs)

• decode_modified_otlp_sketch_bytes now dispatches all five SketchKind variants to their concrete decoders. Previously only CountMin was wired (PR feat: route modified-OTLP sketch metrics through precompute engine (PR B) #6); CountSketch, Kll, DdSketch, and Hll used to return a not yet implemented error and fall through to §5.2.

Enums

• AggregationType::CountSketch (was already added in the earlier CountSketch commit).
• AggregationType::DDSketch — new variant, wired through Display / FromStr.

Deferred

Query semantics for the new accumulators — cardinality estimation for HLL, quantile estimation on the log-bucket representation for DDSketch, median-of-estimators heavy-hitter tracking for CountSketch — are intentionally deferred to a follow-up. query_statistic on these accumulators returns a placeholder not yet implemented error that the caller falls through to the §5.2 fallback with, so queries still return correct answers while the matrix/register/bucket merge + store round-trip works end-to-end today.

Tests

Unit tests

• sketch-core::count_sketch — 6 passed
• sketch-core::dd_sketch — 6 passed (aligned/overlapping/disjoint merge, alpha mismatch, msgpack round-trip)
• sketch-core::hll_sketch — 6 passed (register-wise max, variant/precision mismatch, merge_refs, msgpack round-trip)
• CountSketchAccumulator — 6 passed (int64 / float64 / dim mismatch / zero dims / merge / wrong-type rejection)
• DatasketchesKLLAccumulator::from_sketchlib_proto_bytes — 3 passed (reconstructed quantile, small-k rejection, inconsistent levels rejection)
• DDSketchAccumulator — 4 passed (round-trip, invalid-alpha rejection, bucket alignment merge, wrong-type rejection)
• HllSketchAccumulator — 6 passed (regular/HIP round-trip, register-length mismatch, zero-precision rejection, register max merge, wrong-type rejection)

E2E tests (tests/e2e_modified_otlp_sketch_path.rs)

All five tests POST a real ExportMetricsServiceRequest through the OTLP HTTP receiver, route it through a running PrecomputeEngine, close a window, and assert the captured accumulator's inner state matches the payload:

• ✅ e2e_count_min_sketch_modified_otlp_path (from PR test: e2e test for modified-OTLP CountMin sketch hot path (PR D) #7)
• ✅ e2e_count_sketch_modified_otlp_path
• ✅ e2e_kll_sketch_modified_otlp_path — verifies reconstructed median of [1..=100] is within KLL's rank-error bound of 50
• ✅ e2e_dd_sketch_modified_otlp_path — verifies bucket counts, offset, count, sum, alpha all round-trip
• ✅ e2e_hll_sketch_modified_otlp_path — verifies register bytes + precision round-trip

Validation

• ✅ cargo check --all-targets: clean
• ✅ cargo clippy --all-targets -- -D warnings: clean
• ✅ cargo fmt --check: clean
• ✅ cargo test -p sketch-core --lib: 54 passed
• ✅ cargo test -p query_engine_rust --lib: 458 passed
• ✅ cargo test -p query_engine_rust --test e2e_modified_otlp_sketch_path: 5 passed (all variants hot end-to-end, ~6s serial)

🤖 Generated with Claude Code