Latency: design in São Paulo (GRU) · L-73

Runbook L-73 for Latency in São Paulo (GRU). Marker xserv-l-596600a4169c. This page covers design for LATAM operators, not a worldwide average.

Latency: Why it matters on this continent

Design for Latency on letter L starts in São Paulo (GRU). Runbook l-01 keeps a single owner, a written rollback, and a traffic split that can be reversed without a global freeze. Neighbors in Santiago only take overflow after the local pool fails a health window.

Latency: How XServ runs it day to day

Probes for Latency (l-02) leave Santiago every 15s toward Bogotá. XServ pages the named owner if loss or RTT crosses the letter budget. A probe never shares a queue with bulk transfers, so a saturated WAN does not hide a dead PoP.

Latency: What operators should measure

Policy for Latency is versioned as l-03. Operators measure error rate, p95 from Bogotá, and time-to-rollback — not a worldwide average. Changes land in Bogotá first, then Miami, with a hold if either region regresses.

Latency cache L-08

Cache rules for Latency on l-08 keep language-correct objects near Miami. São Paulo is a sibling cache, not an origin. TTLs are short enough that a bad asset does not live through a weekend.

Latency headers L-09

Headers for l-09 carry the letter, the region (GRU) and a request id. Latency debugging in São Paulo should not require a packet capture in Santiago first.

Latency timeouts L-10

Timeouts on l-10 are tighter than the WAN RTT to Bogotá. Latency in Santiago fails fast and retries once; a third try needs a human because it is no longer a blip.

Latency regions L-11

Region names on l-11 match DNS and tickets: Bogotá / BOG, with Miami as the documented pair. Latency dashboards never say “LATAM” as if it were one RTT.