Stability — full-run time series

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

Flat lines are the story. RSS held 301→357Mi across the whole run; fds/threads bounded; egress & admission returned to 0 after every load phase. 0 OOMKills, 0 unplanned restarts.

Memory — bounded, no leak (claim 8 proxy)

PointRSS (busiest flight pod)
idle / boot4–5 Mi
warm steady-state (32-thr, 90 min)~297 Mi (flat, no upward drift)
peak (80-thr burst)~310 Mi
OOMKills / unplanned restarts0 / 0
proc_rss_bytes per flight pod — flat across the run; no leak
proc_rss_bytes per flight pod — flat across the run; no leak
proc_threads + proc_fds — bounded, no drift
proc_threads + proc_fds — bounded, no drift

Saturation gauges (#2419) — rise under load, return to 0

Sampled at 10s (report) / 15s (OTLP). Both undersample the ~2s in-process gauge, so plotted peaks are lower bounds — the true peaks may be higher. Peaks here are lower than R12's because this run's query mix is point-read-heavy (bounded lookups) vs R12's scan-heavy load.

GaugePeak (this run)R12 peakPost-drain
cqlite_flight_admission_in_use_ratio (limit 64)4120
cqlite_merge_egress_channel_depth124435050
cqlite_flight_blocking_tasks_in_use480
cqlite_proc_threads / fds26 / 3493 / 135baseline
merge_egress_channel_depth — spikes under load then returns to 0 (drift-free)
merge_egress_channel_depth — spikes under load then returns to 0 (drift-free)
admission_in_use_ratio — peak 4, far under the 64 limit
admission_in_use_ratio — peak 4, far under the 64 limit
blocking_tasks_in_use — bounded, settles to 0
blocking_tasks_in_use — bounded, settles to 0