Netflix moves 30,000 jobs to Flink autoscaler

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

🎬 Netflix moves 30,000 jobs to Flink autoscaler

🛠️ Encore lets coding agents test on real infra

🐍 Python adds yield from to async generators

🔍 New tool traces CI pipelines

☸️ Karmada graduates as CNCF project

Plus: 🎁 8 other news you might like, 🧰 6 tools, and 📚 5 papers.

🎬 Netflix moves 30,000 jobs to Flink autoscaler LINK

• Netflix is migrating over 30,000 streaming jobs across multiple AWS regions to the open-source Apache Flink Autoscaler, after its older cluster-level system proved poorly suited to complex, stateful pipelines with operators that have differing processing needs.
• The new autoscaler estimates each operator's true processing rate from throughput and busy time, then walks the job graph to set parallelism per vertex, an approach from FLIP-271 that cut one team's annual Flink compute spend by 58%.
• Netflix wired the autoscaler into a Spring Boot service using Temporal workflows instead of the Flink Kubernetes Operator, tuned JobManager metrics for 3,000-subtask jobs, and set a 0.45 utilization target below Flink's 0.7 default to curb aggressive rescaling.

🛠️ Encore lets coding agents test on real infra LINK

• Combining Daytona and Encore now gives coding agents isolated, persistent sandboxes where they build and verify complete backends against real infrastructure primitives rather than mocks, with the same declarations running unchanged in production.
• Encore's Infrastructure SDK lets agents declare databases, queues, buckets, caches, secrets, and cron jobs alongside code; Daytona spins those up locally in the sandbox, avoiding separate Docker Compose setups or shared cloud credentials.
• Because the primitives are type-safe, agents catch invalid declarations and event-type errors in the same loop, and can connect to the Encore MCP server to call endpoints, inspect database state, and retrieve traces.

🐍 Python adds yield from to async generators LINK

• Python approved PEP 828, adding yield from to async generators in Python 3.16 so delegating to another async generator no longer requires a hand-written async for loop.
• The interpreter now handles delegation end to end, forwarding values, exceptions, and close requests to the inner generator through .asend(), .athrow(), and .aclose(), matching how sync generators have behaved since Python 3.3.
• The PEP also adds a .value attribute to StopAsyncIteration, so an async generator's return statement finally carries a value the delegating caller can capture, closing a long-standing gap in asynchronous code.

🔍 New tool traces CI pipelines LINK

• An OpenTelemetry Collector with the alpha `githubreceiver` component can now trace CI pipelines across a whole GitHub org, converting `workflow_run` and `workflow_job` events into OTLP spans without touching any workflow files.
• A single org-level webhook feeds the collector, which maps each workflow to an outer span, jobs and steps to nested children, with span and trace IDs hashed deterministically from run and check-run IDs so steps can attach their own telemetry.
• Export the spans to Tempo, Jaeger, or Datadog via OTLP; every repo including future ones is covered instantly, though you must expose a public endpoint, secure it with IP allowlisting, and include a dummy `scrapers` entry or config validation fails.

☸️ Karmada graduates as CNCF project LINK

• Karmada, the open source engine for running applications across multiple Kubernetes clusters, clouds, and regions without modifying them, has graduated from the CNCF, its highest maturity tier, announced today at KubeCon China 2026.
• The v1.19 release advances multi-component scheduling for distributed AI training jobs and promotes priority-based scheduling to Beta, enabled by default, so critical workloads get placed first across the fleet.
• Karmada extends the standard Kubernetes API with centralized placement, propagation, failover, and multi-cluster autoscaling, and is used in production by Bloomberg, Trip.com, Alibaba Cloud, Huawei, and Wellhub for hybrid capacity and multi-region resilience.

Other news you might like

• Linux 7.3-rc2 ships heavier than expected, Torvalds blames AI for itLINK

• Apple invites developers to submit browser interoperability ideas for Interop 2027LINK

• GPT-6 Astra beat Portal start to finish without human help in under 24 hoursLINK

• China Merchants Bank Wins CNCF End User Case Study Contest for Unifying AI Training and Inference on KubernetesLINK

• OpenAI’s new model costs 2.5x more per token — and developers are saving money anywayLINK

• Logitech launches $99 MX Keypad for coding and AI workflowsLINK

• Astonishing mod runs DLSS 5 on a second GPU to boost neural-rendered FPS up to 127% — game renders on one card, neural post-processing runs on the other, much like dedicated PhysX GPUsLINK

• Forget Wi-Fi moisture sensors—this ESP32 project works offline and costs half as muchLINK

🧰 Trending tools

Dograh: open-source voice agent platform with visual flow builder, telephony, human transfer, and self-hosted QA monitoring across 30+ model integrationsLINK

Zro: routes coding requests to open-source models like GLM-5.2, DeepSeek V4 Flash, and Kimi K3 through one endpoint without storing your dataLINK

CartAI: automates checkout across merchant sites through one API, handling payments, shipping, and order tracking without building per-site integrations.LINK

ditto.site: converts any public URL into clean Next.js or Vite code, deterministically extracting components, design tokens, fonts, and hover states via a free API.LINK

Replay QA: records browser sessions with full devtools access and converts them into shareable bug reports, adding QA coverage to AI-built apps without writing tests.LINK

Browzer: connects to your codebase to auto-generate self-healing docs, blogs, quickstarts, and videos, freeing DevRel teams to focus on community and growth.LINK

📚 Trending papers & reports

AI bug-fixing tools confabulate false reasoning in about 73% of code repairs, with only 21% to 56% of fixes passing developer tests, meaning automated patches often look plausible while being fundamentally wrong.LINK

Agent skill libraries turn an AI assistant's past task attempts into a tiered playbook of proven moves, letting it complete household simulation tasks ~87% of the time while avoiding failure-prone shortcuts learned from bad runs.LINK

Scrum Master hiring becomes necessary once Agile software teams grow past six members, and companies still lack any standard way to train or measure the soft skills these coordinators need.LINK

EmoPyLab is a free toolkit that lets teams run and compare hundreds of optimization methods across many competing goals on modern chips, then pick the best trade-off, with results locked in for reproducibility.LINK

Explaining model decisions gets more accurate by fixing an overlooked reference point in popular attribution tools, letting anyone see which inputs actually drove a prediction and how confidence builds through the system.LINK

See you tomorrow for a new dose of ☕️ Devshot!