8 min read · 12 sources

DigitalOcean’s Inference Router now refuses to pick a cheap model if switching would blow up your prompt cache. The update, announced this week, makes the router cache-aware: when choosing between models for a request, it factors in the cost of breaking a warm context cache alongside price and quality. The router also adds an explicit model-affinity header for session pinning and a routing-budget knob that caps how much extra cost a model switch can incur.

The pitch is that agentic workloads hammer the same large prompt payloads over and over, which turns caching from a marginal optimization into a primary driver of cost and latency. DigitalOcean cites 90%+ cache-hit rates on Kimi K3 workloads, OpenAI’s claim that cached prompts can cut latency by up to 80%, and Anthropic’s use of prompt caching in Claude Code. For teams running agentic systems, the practical upshot is that “cheapest model wins” routing is wrong when the same 200k tokens of context are getting re-billed every turn.

A July 23 Azure West US outage lasted about five hours after a planned optical-device repair, a defective blast-radius analyzer, and per-device safety checks combined to withdraw all egress routes at once.

AWS Glue 6.0: 30% cheaper, Iceberg v3, and a VARIANT type that flattens your JSON problem

Source: aws.amazon.com ↗

AWS Glue 6.0 is generally available in every region where Glue operates, priced 30% below prior versions with no API changes required. The runtime is modernized end-to-end: Apache Spark 4.1, Python 3.13, Scala 2.13, and full Apache Iceberg v3 via Iceberg 1.11.0.

The Iceberg v3 piece is the part that will reshape ETL schemas. Glue 6.0 ships a VARIANT data type with shredding, plus Geometry and Geography types, nanosecond-precision timestamps, and explicit handling of unknown-typed fields. In practice, shredding means you can store and query nested JSON, application logs, and event payloads without flattening them into tables first, and Iceberg v3 will read them efficiently on the way out. Glue 6.0 also adds Spark Declarative Pipelines for declarative ETL authoring and Arrow-native Python UDFs/UDTFs to cut pandas-row serialization overhead. For anyone running Spark jobs against S3, the 30% price drop combined with the runtime upgrade is the kind of change that makes you rebuild your cost model.

Docker Verified Publisher applications are now self-serve

Source: docker.com ↗

Docker moved its Verified Publisher program to a self-serve flow inside Docker Hub, removing the previous sales-team gatekeeping while keeping manual review on every application. The badge now covers all content types on Docker Hub, including MCP servers, models, sandboxes, and agents — not just container images.

Approved publishers get a verified badge, priority search ranking, and analytics that show which companies are pulling their images, with usage broken out by domain so it can flow directly into a sales pipeline. Current verified publishers include Google, Microsoft, AWS, Datadog, and Grafana Labs. For vendors distributing images on Docker Hub, the discoverability boost and the new analytics tier are the actual product; the self-serve flow is the friction removal.

Anatomy of a five-hour Azure outage

Source: surfingcomplexity.blog ↗

Microsoft has published a post-incident review for a July 23 Azure West US networking outage that took roughly five hours to resolve. Three things had to go wrong at once, and they did.

A planned optical-device repair kicked off a blast-radius analysis that was supposed to bound the scope of the maintenance. A defect in that analyzer incorrectly expanded the scope to every optical device egressing the datacenter. Then a safety validation checked the devices individually rather than evaluating the aggregate effect of isolating them all at once. The result was simultaneous route withdrawals that disconnected West US from the internet while physical links and adjacencies still appeared healthy, which is why the failure initially looked like a monitoring problem. The rollback automation, of course, depended on the connectivity that had just been withdrawn.

The post lists five contributing factors: flawed blast-radius analysis, inadequate aggregate safety validation, misleading health signals, incomplete event logging, and rollback automation that depended on the disrupted path. For SREs, the lesson is that per-component checks pass while the system-level invariant is being violated, and that “everything looks healthy” is not a useful signal during a partial failure.

Kubernetes microservice orchestration in Octopus

Source: octopus.com ↗

Octopus’s Easy Mode series adds a Kubernetes microservice orchestration pattern for teams that have many independently deployable services and want to promote them as a unit. The pattern uses a parent orchestration project containing “Deploy a Release” steps that sequence promotion of multiple microservice projects between environments.

The walkthrough uses the Octopus AI Assistant Chrome extension on Octopus Cloud, and the example wires up a token account, a Docker Hub feed, and a “Mock K8s” target using the octopusdeploy/worker-tools:6.5.0-ubuntu.22.04 image with client-side apply. The interesting bit is the project-triggers-project model: there is no external orchestrator, and the release order is encoded directly in the parent project. For teams whose release order lives in shell scripts and a Slack thread, that is a real upgrade.

Two repos worth a look: Databasus and Aerospork

Source: github.com ↗

Databasus is a free, self-hosted PostgreSQL backup tool aimed at low RPO/RTO and Point-in-Time Recovery. It supports physical backups via PostgreSQL’s native incremental mechanism (full, incremental, and continuous WAL streaming), logical compressed dumps, and cron scheduling. It does real restore verification by spinning up a database container and checking row counts per table, and retention is configurable for fixed duration, fixed count, GFS, and size caps. Storage targets include local, S3, Cloudflare R2, Google Drive, NAS, Dropbox, SFTP, and Rclone, with notifications over email, Telegram, Slack, Discord, Teams, Mattermost, and webhooks.

Aerospork is a Swift fork of the AeroSpace i3-style tiling window manager for macOS, MIT-licensed and targeting macOS 13.0+ with Swift 6.0. It is configured via TOML and ships a SwiftUI MenuBarExtra settings GUI. The fork exists because upstream monitor matching by name, regex, or index is fundamentally broken for duplicate panels. Aerospork adds a MonitorFingerprint that matches displays on per-display UUID first, then EDID vendor/model/serial, then name and size, with debounced reconfiguration to handle staged DisplayLink dock connections. It also adds a 50ms debouncer that coalesces accessibility event bursts and a no-op skip in MacApp.setFrame to cut framebuffer writes, plus on-demand workspace creation and release to keep long sessions from drifting.

CISA's BOD 26-04 replaces CVSS with risk-based SLAs

Source: datadoghq.com ↗

CISA’s Binding Operational Directive 26-04 codifies risk-based vulnerability prioritization and patching timelines for federal agencies using four variables: public exposure, KEV catalog membership, exploit automation feasibility, and technical impact (total versus partial control). The most severe combinations require remediation within 3 calendar days plus forensic triage. Timelines stretch to 14, 21, or 30 days as risk drops, and vulnerabilities that do not qualify are deferred to the next system upgrade.

For federal suppliers, this is the first time CISA has replaced CVSS-driven defaults with exposure- and exploitability-driven SLAs at the directive level, and the timelines are calendar days, not business days. Datadog’s Runtime Prioritization Engine is positioned to map runtime context onto the same four variables, which is the kind of thing that is about to be table stakes for any vendor selling to federal agencies.

AI agents on call inherit control-system failure modes

Source: surfingcomplexity.blog ↗

A piece on the Surfing Complexity blog argues that pushing AI agents into first-responder roles inherits control-system failure modes that may exceed the incidents they are meant to mitigate. The argument runs through Ashby’s Law: handling more system states requires more complexity, which is what lets agents tackle generic on-call work in principle, but the same complexity makes it harder for humans to reason about what the agent did when something goes wrong. The post references Boris Tane’s “On-Call is Now Theatre” piece and an OpenAI BlackHat talk, and points to Air France 447 and the 737 MAX as cautionary precedent for complex-control failures.

The most dangerous scenario, the author argues, is an agent attempting remediation in an unexpected way, making an already complex outage worse before humans enter the loop, and forcing responders to diagnose the combined behavior of the software system and the agent itself. For reliability teams shipping agentic runbooks, the practical question is what your incident review looks like when half the actions on the timeline were taken by a model that you cannot fully explain.

"There continue to be reasons for software to be slow"

A counter-argument to Dan Luu’s claim that LLMs make optimization universally worthwhile. The author walks through the form of the argument — property X used to cost $A over budget $B, and post-LLMs it costs $A/N < $B — and shows how each premise can fail in practice, even assuming rational economic actors with perfect knowledge. Teams may tolerate slower software, cut engineering budgets, deprioritize performance, or face larger costs in testing, migration, correctness, and long-term maintenance. AI-generated optimizations can overfit benchmarks or introduce reliability problems, so faster code generation does not automatically translate into faster production software. The same critique is applied to blanket claims about testing and formal methods scaling with LLM cost reductions.

A2A joins MCP under the Linux Foundation, and a sandbox for local coding agents

Source: devopsdigest.com ↗

Google’s Agent2Agent protocol has joined the Agentic AI Foundation, the same home as Anthropic’s Model Context Protocol. A2A is an open agent-to-agent communication standard where agents publish structured “agent cards” for capability discovery and task delegation, donated to the Linux Foundation with AWS, Cisco, Google, Microsoft, Salesforce, SAP, and ServiceNow, and merged with IBM’s Agent Communication Protocol in August 2025. A2A v1.0 shipped in March 2026 with multi-protocol bindings, version negotiation, multi-tenancy, and signed agent cards for cryptographic identity. For multi-agent systems, MCP plus A2A now form complementary standards (tools and data versus agent-to-agent) under neutral governance.

Octopus argues that shared corporate AI agents do not necessarily need sandboxes if their tools are secure, but local agents running on developer machines with arbitrary prompts and full environment access absolutely do. Confirmation prompts are described as a poor substitute for real policy enforcement, on the order of MFA fatigue. The proposed fix is a Vagrant-built VM sandbox that confines the local agent’s blast radius while still allowing IDE workflows, with a final Vagrantfile on GitHub and an Easy Mode Claude Agent example. For engineers running coding agents locally, the post makes a concrete case for VM-level isolation over prompt-based gating and ships a runnable starting point.