Flat lines are the story. Over 3.5h at 32 threads: RSS flat, fd bounded
(max 48), threads bounded (max 29), egress depth ~0, admission well under the 64 limit, 0 OOMKills,
0 restarts on all 3 flight pods (4h45m uptime). 42 samples @ 5-min, host_name-labeled (the metrics
fix held the entire soak).
Memory — no leak
Flight pod RSS — db0 ~1.1GB, db1 ~470MB, db2 ~210MB, dead flat across 3.5h (slight step-down at soak end). No growth.
File descriptors & threads — bounded
Open FDs per pod — bounded, flat (max 48 across the soak).Threads per pod — bounded, flat (max 29).
Backpressure & admission
Merge egress channel depth — ~0 throughout (no backpressure buildup under steady load).Admission in-use per pod — well under the 64 limit during steady 32-thread load.
New 0.16 gauges
cqlite.flight.warm_tables — rises on first serve, stable under load.cqlite.sstable.index_parses_total — flat on warm (no re-parse storm; the new tables_discovered gauge does not parse).
Per-pod work distribution (fan-out skew — #2680, by design)
do_get req/s per pod — db0/db2 ~160/s, db1 ~95/s. Busy pods are FIXED (not rotating).do_get bytes/s per pod — same shape. db0 4.34GB / db2 3.48GB / db1 3.06GB total.
Totals over the soak: do_get db0=2,042,139 · db2=2,053,218 · db1=1,239,138. The 2–4× skew and
the resulting ~39 qps ceiling are the known, reverted-#2680 behavior (fourth round) —
reported on the goals ladder, not as a regression.