gw-04 — References

The primary source (read first — it's named in the JD)

• Curbing Connection Churn in Zuul — Netflix TechBlog, 2023. The per-event-loop pool + subsetting work; the ≈8× / thousands→~60 result; the low-discrepancy (Van der Corput) distribution ring. https://netflixtechblog.com/curbing-connection-churn-in-zuul-2feb273a3598
• Arthur Gonigberg, Curbing Connection Churn (companion write-up). https://arthur.gonigberg.com/2023/10/03/curbing-connection-churn/

Subsetting & balanced distribution

• Google SRE Book / Site Reliability Engineering, "Load Balancing in the Datacenter" — deterministic subsetting: why N×M doesn't scale and how subset selection keeps balance. https://sre.google/sre-book/load-balancing-datacenter/
• Van der Corput / low-discrepancy sequences — the math behind the balanced ring. (Any quasi-Monte-Carlo reference; the binary Van der Corput sequence is "bit-reverse i".)
• Twitter/Finagle Aperture load balancer — another production take on subsetting with a ring (deterministic aperture). https://twitter.github.io/finagle/guide/Clients.html#aperture-load-balancers

Pooling & keep-alive

• Go net/http.Transport — MaxIdleConnsPerHost, IdleConnTimeout, MaxConnsPerHost, ForceAttemptHTTP2; read the pooling logic.
• Envoy connection pools (per-worker-thread pools — the same per-event-loop idea) and subset load balancer docs. https://www.envoyproxy.io/docs/envoy/latest/intro/arch_overview/upstream/load_balancing/subsets
• gRPC MAX_CONNECTION_AGE / keepalive — the rebalancing-vs-churn tension from gw-02.

TLS handshake cost (why churn is a CPU problem)

• RFC 8446 (TLS 1.3) handshake; session resumption / 0-RTT.
• Cloudflare blog — TLS handshake cost, session resumption, and why connection reuse is a CPU optimization.

Tooling

• ss -s / ss -tan state established | wc -l — live connection counts.
• nstat -az | grep -i active — TcpActiveOpens (outbound connections opened) — the churn counter on the gateway side.
• wrk/wrk2 with and without --latency; toggle keep-alive to see churn (a non-keepalive client maximizes it).

Cross-lab dependencies

• Upstream: gw-01 (handshake/TIME_WAIT), gw-02 (h2 multiplexing), gw-03 (the endpoint Transport).
• Downstream: gw-06 (LB over the subset), gw-08 (EDS membership), gw-09 (EndpointSlices, pod churn).