<p><strong>TLDR:</strong> Iranian drone strikes in early March 2026 damaged AWS data centers across UAE and Bahrain, exposing the limits of multi-AZ resilience assumptions. Spotify’s internal agent Honk is now merging 1,000 PRs every 10 days across their entire codebase, with PR review — not code generation — as the new bottleneck. Google Cloud’s Memorystore for Valkey 9.0 hit GA with some genuinely useful performance gains. Microsoft expanded Fabric IQ to expose its semantic layer via MCP, targeting the multi-agent data consistency problem.</p>
<hr>
<h2>AWS Middle East outage from kinetic attacks</h2>
<p>Drone strikes in early March 2026 damaged three AWS data centers across ME-CENTRAL-1 (UAE) and ME-SOUTH-1 (Bahrain), taking out two of three AZs in UAE and one facility in Bahrain. The incident makes a point that’s easy to dismiss until it isn’t: multi-AZ protects against hardware failures and localized disasters, not coordinated physical attacks that can hit multiple zones in the same geographic region simultaneously. AWS told affected customers to migrate workloads to US, Europe, or Asia Pacific — which immediately created data residency compliance headaches for regulated industries. Multi-cloud isn’t the answer here either, since the risk is regional rather than provider-specific. If you’re running anything in Middle East regions, multi-region architecture isn’t optional anymore.</p>
<h2>Spotify’s Honk agent: 1,000 merged PRs per 10 days</h2>
<p>Spotify presented Honk at QCon London 2026 — an internal AI agent that runs continuous large-scale code migrations across their entire codebase. Before Honk, deterministic scripts handled 70% of migrations and cut timelines from nearly a year to under a week, but the remaining 30% needed LLM reasoning for edge cases. Honk wraps the full dev lifecycle — requirements, generation, build, test, iteration — and integrates with Slack for triggering migrations from anywhere. The interesting architectural call: separating agent runtime from verification runtime. The new constraint is PR review throughput, not code generation, which is pushing Spotify toward codebase standardization and tooling changes to make reviews faster and agent output more predictable.</p>
<h2>Google Cloud Memorystore for Valkey 9.0 goes GA</h2>
<p>Valkey 9.0 is now generally available on Google Cloud with some concrete performance numbers: up to 40% higher throughput from pipeline memory prefetching, 20% from zero-copy responses, and 200% throughput improvement for BITCOUNT and HyperLogLog operations via SIMD. The new commands are worth noting — HEXPIRE adds granular hash field expiration, GEOSEARCH BYPOLYGON enables polygon-based geospatial queries, and DELIFEQ provides atomic conditional delete for distributed locking without reaching for Lua scripts. Existing customers can upgrade in-place with no downtime.</p>
<h2>Microsoft Fabric IQ exposes semantic layer via MCP</h2>
<p>Microsoft expanded Fabric IQ to publish its business ontology through MCP, making it accessible to AI agents from any vendor. The problem being solved is real: in multi-agent deployments, different agents built on different platforms end up with conflicting definitions of the same business entities — what counts as a “customer” or an “order” varies by system, and that causes decision failures at runtime. The expansion also consolidates Azure SQL, Cosmos DB, PostgreSQL, MySQL, and SQL Server under a single Database Hub management plane. Analysts flagged integration complexity and organizational readiness as open questions. The broader signal for data engineering teams: the semantic layer is becoming production infrastructure, not a BI concern.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>API reliability report (215+ services):</strong> OpenAI averaged one incident every 2.5 days in January 2026; Stripe ran near 99.99% uptime; AWS DynamoDB cascaded failures into 141 downstream services; median resolution time was ~90 minutes but tail events cause disproportionate damage.</li>
<li><strong>Grab’s multi-agent support system:</strong> Built on LangGraph and FastAPI, the system routes Slack requests through an Enhancement Agent (GitLab MRs, test runs) and four specialized investigation agents, cutting issue resolution from hours to minutes and reclaiming several FTE-equivalents of engineering bandwidth.</li>
<li><strong>Kubernetes custom metrics scaling:</strong> 64% of organizations use HPA but only 20% scale on custom metrics; CPU and memory are lagging indicators that miss real demand for queue-backed workers and latency-sensitive services — queue depth and tail latency are better signals.</li>
<li><strong>OpenAI Codex internals:</strong> Codex is built on codex-1, a fine-tuned o3 variant; when MCP proved insufficient for streaming progress and mid-task approval pauses, OpenAI built a custom JSON-RPC App Server protocol powering CLI, VS Code, web, desktop, and third-party IDE integrations from one codebase.</li>
<li><strong>Musinsa AI-native hiring:</strong> The Korean fashion platform redesigned junior engineering interviews around a deliberately vague concurrency problem (100 simultaneous requests, only 1 succeeds), evaluated via a 3-tier automated pipeline that checks if the service runs, tests correctness under load, and analyzes prompt history and git history to distinguish engineers who understand what they built.</li>
<li><strong>Observability cost growth:</strong> Production telemetry analysis found 82% of data points missing <a href="http://service.name" target="_blank" rel="noopener nofollow">service.name</a> attributes, 75% log duplication, and PII leaking via auto-instrumentation — the root cause of growing observability bills is instrumentation quality, not vendor pricing.</li>
<li><strong>LLM observability on one database:</strong> A unified approach using OpenTelemetry GenAI Semantic Conventions and GreptimeDB stores traces, metrics, and conversation logs in a single backend, enabling SQL joins across traces and conversation logs via shared trace_id/span_id.</li>
<li><strong>Client ID Metadata Documents (CIMD):</strong> A new OAuth mechanism where clients host metadata at a stable HTTPS URL used as their client_id, eliminating pre-registration with every authorization server; MCP is adopting CIMD as the preferred default for client registration via SEP-991.</li>
<li><strong>Temporal sensitive data patterns:</strong> Two patterns for keeping PII out of Temporal workflow history — payload encryption (data encrypted before reaching Temporal, decrypted only by workers) and the claim check pattern (large payloads stored in S3, only a reference key passed through the workflow).</li>
<li><strong>Databricks acquires Neon for ~$1B:</strong> Confirms that serverless Postgres and database infrastructure are the binding constraint for agentic AI systems, not the models themselves.</li>
<li><strong>Mastodon OTel setup:</strong> A single engineer runs OpenTelemetry Collectors across two Kubernetes deployments handling up to 10 million requests per minute using one Collector per namespace, tail-based sampling at 0.1% for successful traces and 100% for errors, managed via OTel Operator and Argo CD.</li>
</ul>


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

Backend Digest

A tech news roundup covering several major stories: Iranian drone strikes in early 2026 damaged AWS data centers in UAE and Bahrain, exposing the limits of multi-AZ resilience for regional physical threats. Spotify's internal AI agent Honk is now merging 1,000 PRs every 10 days across their codebase, with PR review throughput — not code generation — becoming the new bottleneck. Google Cloud's Memorystore for Valkey 9.0 reached GA with up to 40% throughput gains from pipeline memory prefetching, new commands like HEXPIRE and DELIFEQ, and SIMD-powered improvements for BITCOUNT and HyperLogLog. Microsoft expanded Fabric IQ to expose its semantic layer via MCP, targeting data consistency problems in multi-agent deployments. Additional items cover API reliability benchmarks, Kubernetes custom metrics scaling, OpenAI Codex internals, observability cost drivers, and Databricks acquiring Neon for ~$1B.

AWS Middle East regions hit by drone strikes, Spotify's Honk merges 1,000 PRs every 10 days