Raw data — Sky.Live under load on x86 GCE, 2026-08-15

Analysis and every stated condition live in ../../skylive-remote-validation.md, "The active result". These are the files it was derived from.

Targets were sky-bench-micro (e2-micro) and sky-bench-small (e2-small), us-central1-a, project settleby, both running examples/26-ui-showcase at commit ba3c3b1d. Both instances were deleted; see teardown.txt.

file	what
`idle-{micro,small}-samples.tsv`	300 s idle baseline, 5 s sampling, no Ops Agent
`{micro,small}-noagent.tsv`	100/250/500 sessions × 3 repeats
`{micro,small}-lowN.tsv`	1/25/50 sessions × 2 repeats
`micro-AGENT.tsv`	25/50/100 × 2, with the Ops Agent installed
`micro-rss-n500-r1-memexhaustion.txt`	1 Hz trace of the run where the e2-micro ran out of memory
`teardown.txt`	deletion + verification, verbatim

Column contract

*-noagent.tsv, *-lowN.tsv, micro-AGENT.tsv:

level         sessions requested
repeat        repeat index at that level
idle_rss_kb   app RSS after restart, before load (median of 5 reads)
load_rss_kb   app RSS under load (median of last 40 of ~110 1 Hz samples)
delta_kb      load_rss_kb - idle_rss_kb
established   sessions the generator actually established -- THE DIVISOR
kb_per_session delta_kb / established
conn_app_max  peak established TCP connections on :8000 (~2 per session)
throughput    interactions/sec
p50/p95/p99   interaction latency, ms, INCLUDING a ~111 ms UK->us-central1 RTT
err_rate      fraction of interactions not returning a patch set
valid         generator's own validity flag

idle-*-samples.tsv is scripts/skylive-observe-remote.sh output:

ts rss_kb vmsize_kb threads conn_app conn_pub proc_jiffies
cpu_total cpu_idle load1 req_total msg_total mem_avail_kb

micro-rss-n500-r1-memexhaustion.txt is epoch rss_kb conn_app mem_avail_kb at 1 Hz.

Two traps in this data

conn_app is not a session count. It runs at ~2× sessions: each session holds an SSE stream and a keep-alive connection for its event POSTs. Use established.

kb_per_session at level=1 is not a per-session cost. It is the app's fixed first-request growth (~11–13 MB) divided by one. The per-session figure is the slope across levels, not any single row's ratio; see the analysis doc.