$ whoami
jamesarch — Staff SRE / Platform Engineer
10 yrs infra · 8 yrs full-remote · China → anywhere async
$ uname -a
production 80B+ daily requests 95k+ enterprise clients 10k+ live agents 200+ global nodes
×ばつ2R + internal_replication → 36-59% silent read loss per query · fixed via hot-reload · 19 TB reclaimed [memory] glibc per-thread arena fragmentation under ~2k msg/s json churn → mimalloc 2000 MiB/pod → 250 MiB/pod (-87%) no leak no restart [ha-platform] 10k-agent websocket mgmt: single instance → redis-backed multi-replica (presence · pub-sub routing · sigterm drain) → rolling restart agent downtime <1s vs full-restart = 2200 agents drop [incident-492] 7600 overseas agents offline · machines alive · no app error root: oversea dns line → broken cloudflare cname → 403 on /ws/agent fix: dig +subnet confirmed · direct A record · 4 min full recovery [aiops] hermes-agent (NousResearch) deployed in production ops extended with custom MCP server: 154 infra skills (vsphere · ansible · metrics) 3-tier model fallback · full sqlite audit log per agent action · zero inbound ports">
$ cat /proc/impact
[observability] zabbix 3934 nvps · profiled rcache utilization (0.4% actual vs. overconfig)
→ cpu 36→12 cores (-67%) load 175→32 (-82%) nvps flat zero restart
[data-integrity] clickhouse 4-shard cluster misconfigured as ×ばつ2R + internal_replication
→ 36-59% silent read loss per query · fixed via hot-reload · 19 TB reclaimed
[memory] glibc per-thread arena fragmentation under ~2k msg/s json churn
→ mimalloc 2000 MiB/pod → 250 MiB/pod (-87%) no leak no restart
[ha-platform] 10k-agent websocket mgmt: single instance → redis-backed multi-replica
(presence · pub-sub routing · sigterm drain)
→ rolling restart agent downtime <1s vs full-restart = 2200 agents drop
[incident-492] 7600 overseas agents offline · machines alive · no app error
root: oversea dns line → broken cloudflare cname → 403 on /ws/agent
fix: dig +subnet confirmed · direct A record · 4 min full recovery
[aiops] hermes-agent (NousResearch) deployed in production ops
extended with custom MCP server: 154 infra skills (vsphere · ansible · metrics)
3-tier model fallback · full sqlite audit log per agent action · zero inbound ports
$ ls -la ~/oss/
createrepo_rs — Rust RPM repo metadata generator
thread::scope + crossbeam parallel scan · in-memory SQLite · single VACUUM INTO flush
254 pkg: 10.5s → 0.078s (×ばつ) · musl static · .deb .rpm debian-src · 4 MB scratch docker
packaged in COPR / AUR / Debian
browser-autos — cloud-native headless browser automation API
TypeScript · multi-arch Docker (browserautos/browser-autos) · production-grade
ghostty-patched — personal ghostty build
new_tab_with_command · link-file · concurrency deadlock fix on link-click
VOUCHED contributor → ghostty-org/ghostty
$ gh pr list --author jamesarch --state merged --repo upstream
kubernetes/kubernetes ✓ merged
nushell/nushell ✓ merged
aya-rs/aya ✓ merged
actix/actix-web ✓ merged
rust-course ✓ merged
$ cat ~/.config/runs-in-prod
# things I build
lang rust · go · python · shell
own createrepo_rs browser-autos ghostty-patched ansible-over-wss vsphere-iac
# things I operate at scale (80B req/day · 10k agents · 200+ global nodes)
container kubernetes (k3s + multi-cluster) docker containerd
platform hermes-agent mcp-server openai-compat-gateway
observe zabbix 7.0 victoriametrics clickhouse prometheus grafana vector alertmanager
infra ansible opentofu vsphere+govmomi hyper-v kvm routeros
network wireguard quic/quinn mtls iptables/conntrack/netfilter
$ cat ~/.config/approach
ai-native workflow:
heavy use of ai agents for coding · root-cause analysis · operational automation
my half: problem discovery · architecture decisions · production validation · red lines
prefer boring reliability over clever fragility
prefer static binaries over dependency piles
prefer observable failures over silent magic
prefer terminal workflows over dashboards
$ echo $LOCATION
China · full-remote · open to overseas / async-first