Blog

Notes from the network.

Post-mortems, latency experiments and the occasional opinion about configuration management.

38points of presence
41 Tbpsglobal capacity
99.99%edge availability
22 msmedian TTFB

2026-09-18 · Network

Anycast is not magic: what actually changes when you move to the edge

Anycast shortens the first hop. Everything after it — cache state, origin pools, TLS reuse — decides whether users feel faster. A field guide to the parts that matter.

Read note →

2026-08-02 · Engineering

Cutting tail latency with connection reuse

p95 is a connection-management problem more often than a compute problem. Here is how pooling, keepalive tuning and request coalescing moved our p95 by 38%.

Read note →

2026-06-11 · Reliability

A practical guide to regional failover

Failover is easy to draw and hard to test. We share the health-check settings, hold-down windows and game-day scripts we use in production.

Read note →

Recent notes

2026-09-18 · Network

Anycast is not magic: what actually changes when you move to the edge

Anycast shortens the first hop. Everything after it — cache state, origin pools, TLS reuse — decides whether users feel faster. A field guide to the parts that matter.

We publish the measurement method with every post: the window, the sample size and the region. If a number here cannot be reproduced from the dashboard, it does not belong in the post.

Written by the reliability team · 6 min read

2026-08-02 · Engineering

Cutting tail latency with connection reuse

p95 is a connection-management problem more often than a compute problem. Here is how pooling, keepalive tuning and request coalescing moved our p95 by 38%.

We publish the measurement method with every post: the window, the sample size and the region. If a number here cannot be reproduced from the dashboard, it does not belong in the post.

Written by the reliability team · 6 min read

2026-06-11 · Reliability

A practical guide to regional failover

Failover is easy to draw and hard to test. We share the health-check settings, hold-down windows and game-day scripts we use in production.

We publish the measurement method with every post: the window, the sample size and the region. If a number here cannot be reproduced from the dashboard, it does not belong in the post.

Written by the reliability team · 6 min read

Put your next deployment closer to users.

14-day trial, no card, full network. Migration help included.