<p><strong>TLDR:</strong> American Express published a detailed breakdown of two zero-downtime payments network migrations, using canary routing and shadow traffic replay as the core techniques. Cilium turned 10 and now holds over 60% CNI market share, with 1.0.19 shipping ztunnel mTLS and eBPF-based kube-proxy replacement. QCon London surfaced real concerns about autonomous coding agents leaking secrets and ballooning costs. Spring Batch and a from-scratch Kafka commit log in Go round out the practical backend reading today.</p>
<hr>
<h2>American Express: two payments network migrations, zero downtime</h2>
<p>Amex published a detailed post-mortem on migrating live card authorization traffic twice without taking downtime. The first migration moved from a legacy platform to microservices using a Global Transaction Router for connection migration, shadow traffic replay for pre-production validation, and canary routing starting at 1%. The second migration reused the same playbook to move between Kubernetes environments, adding Envoy Proxy and a custom control plane for multi-region canary drain-and-refill. The key takeaway is that rollback was treated as a first-class requirement from day one, not an afterthought.</p>
<h2>Cilium turns 10, now runs AI training clusters at Microsoft and Google</h2>
<p>Cilium hit its 10th anniversary with 60%+ CNI market share, 1,010 contributors, and 24,000 GitHub stars. The 1.19 release adds ztunnel mTLS integration, stable Multi-Pool IPAM, and Hubble flow log aggregation. Zynga’s case study is worth reading: they eliminated conntrack scaling issues entirely by replacing kube-proxy with eBPF-based load balancing. Microsoft and Google are now using Cilium as the networking data plane for large-scale AI training clusters.</p>
<h2>QCon London: coding agents are getting more capable and more dangerous</h2>
<p>Birgitta Böckeler’s QCon London keynote covered the shift from vibe coding to autonomous agent swarms, and it wasn’t all positive. She cited a real incident where an unsupervised agent leaked secrets to an NPM registry via a prompt injection attack. She also flagged rising costs as a practical constraint on agent-based development. Her risk framework — mistake probability, impact, detectability — is a useful mental model. The closing point landed well: AI coding amplifies whatever engineering practices you already have, good or bad.</p>
<h2>Martin Kleppmann at QCon London: local-first as a cloud sovereignty strategy</h2>
<p>Kleppmann is speaking at QCon London 2026 on local-first software as a concrete response to Europe’s dependency on US cloud providers. The framing is geopolitical as much as technical, positioning local-first architectures as infrastructure sovereignty rather than just an offline-first UX pattern. Worth watching if you’re thinking about data residency or multi-cloud resilience.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Spring Batch at scale:</strong> Practical walkthrough of chunk-oriented processing, JobRepository-backed crash recovery, and scaling options including Remote Chunking with Kafka and Remote Partitioning.</li>
<li><strong>Kafka commit log from scratch in Go:</strong> Covers store files with length-prefixed binary records, memory-mapped index files via gommap, and file segmentation — full source included.</li>
<li><strong>DCB-compliant event store in SQLite:</strong> Four-column table plus a conditional INSERT handles optimistic locking; the post explains how to adapt the same pattern for PostgreSQL using SERIALIZABLE isolation.</li>
<li><strong>Feature flags and CI gaps:</strong> If CI only tests the default code path, enabling a flag in production hits untested logic. Fix: force both flag-on and flag-off paths in every CI run.</li>
<li><strong>Rollback that didn’t actually roll back:</strong> Schema had already migrated forward, so old app code couldn’t read it. The fix is bidirectional migrations with explicit backward paths.</li>
<li><strong>AI-generated code in production:</strong> CodeRabbit data puts AI code at 1.7x more bugs than human-written; Veracode found 45% contains OWASP Top 10 vulnerabilities. Practical mitigations: test-first prompting, architecture context files, N+1 query audits, and targeted observability.</li>
<li><strong>JavaScript dependency bloat:</strong> Three root causes — legacy engine targets, atomic one-liner packages, and stale ponyfills. Tools like knip and the e18e CLI can help identify what to cut.</li>
<li><strong>BFF pattern decision guide:</strong> Four criteria worth checking before committing to a Backend For Frontend: device-specific UX needs, number of third-party providers, shared business logic across clients, and auth/session security requirements.</li>
<li><strong>Starlette 1.0 released:</strong> Minor breaking changes from 0.x; Claude Skills was used to bridge the gap since the model isn’t yet trained on the 1.0 changes.</li>
<li><strong>Five years of systems reading groups at Microsoft:</strong> Practical lessons — start small, let scope grow organically, use multi-session series for depth, and make it easy to show up unprepared.</li>
</ul>


Server-side development, system design, API patterns, and distributed architecture.

Backend Digest

A curated backend engineering digest covering: American Express's detailed account of two zero-downtime payments network migrations using canary routing, shadow traffic replay, and rollback-first design; Cilium's 10th anniversary with 60%+ CNI market share and new eBPF/mTLS features; QCon London talks on autonomous coding agent risks (secret leaks, cost overruns) and local-first software as cloud sovereignty; plus practical notes on Spring Batch at scale, building a Kafka commit log in Go, SQLite event stores, feature flag CI gaps, AI-generated code quality data, and the BFF pattern.

American Express migrates payments twice with zero downtime, Cilium hits 10 years and 60% CNI market share