Comparison — 0.15 vs baselines & Lane-B ladder

v0.15.0 milestone snapshot (condensed) · flight v0.15.0@sha256:30c2b10c (multi-arch INDEX) · connector 0.15.0 · Trino 481 · Cassandra 5.0 RF=3 · 3× i4i.xlarge · ~1.93M partitions/node, 2 SSTable gens · 2026-07-17

vs R11b / R12 baselines

MetricR11bR120.15 (snap)
warm throughput~34 qps @8-thr~33 qps @8-thr~39 qps @32-thrparity+
warm p50 / p99227 / 366ms242 / 363ms798 / 1366ms @32-thrhigher concurrency
count(*) rows1,939,2861,947,7751,927,467dataset-dependent
count(*) wall66.2s61.1s~60sparity
idle RSS3–4 Mi3 Mi4–5 Miparity
peak RSS270–391Mi~603MB @80-thr~310Mi @80-thrlower
do_get error rate2.3%1.2%0.89%improving
snapshot grace-sweepquery-triggered (738 held)background (660→6, no query)FIXED
OOMKills / restarts0 / 00 / 00 / 0clean

Note: p50/p99 are not directly comparable — R11b/R12 measured 8-thread load, this run 32-thread. At equal concurrency the numbers track; the higher latency here is queuing at 4× the threads against the same throughput ceiling.

Lane-B goals ladder — where 0.15 lands

GoalTarget0.15
B1 warm point-read floor≤300ms~800ms p50 @32-thr (incl. ~2.2s JDBC floor amortized)below goal at 32-thr
B1 warm scan-lite floor≤500msLIMIT-class within steady p50report
B2 throughput≥100 qps @32-thr~39 qpsNOT met — fan-out skew is the bottleneck
B3 full-scan trajectory66s→10s~60son old trajectory (this run didn't target scan tuning)
B4 cold start≤3s / idle ≤16Mi2.2s / 4–5MiMET

The Lane-B ladder is progress-tracking, not pass/fail. The one clear gap — B2 throughput — traces directly to the fan-out skew finding: a single pod is the ceiling. Cold-start (B4) is comfortably met.