GCP

Cloud Run worker pools GA: pull-based, non-HTTP workers as managed serverless

Cloud Run worker pools are GA for pull-based, non-HTTP workloads, enabling autoscaled managed workers for agentic AI, queue-driven batch, and background jobs.

July 31, 2026·3 min read·AI researched · AI written · AI reviewed

Cloud Run just changed the serverless playbook: worker pools are GA, and they're explicitly built to pull work instead of answering HTTP. That single sentence should make platform engineers start rethinking how they design background jobs, AI agents, and queue-driven pipelines on GCP.

Until now, teams who wanted serverless economics but non-HTTP semantics had two unattractive choices: force a twisted HTTP front-end onto a worker, or run and operate a K8s workload (or a custom controller) to process queues. Cloud Run worker pools remove that friction by introducing a managed resource for pull-based, long-lived and short-lived workers under the Cloud Run control plane. The consequence is simple and profound: you get autoscaled, service-managed compute that integrates with message systems, task queues, and agent frameworks without the operational burden of running your own Kubernetes controllers.

Why this matters

Because it's no longer an either/or between "pure serverless" and "run a cluster," teams can move more workloads into a managed runtime with CPU/memory scaling semantics tuned for background processing. That matters for three real workloads I see in the field: agentic LLM workers that need ephemeral compute and secure tool access, high-volume queue consumers for ETL/batch, and latency-sensitive inference pipelines that want autoscaling without node churn.

This isn't just UX polish. It changes operational responsibilities. Managed pull workers expand your attack surface: ephemeral processes will now be the place where credentials, secrets, and third-party connectors run. You should treat these worker pools like short-lived VMs in your threat model �14 runtime isolation, workload identity, token lifetimes, and audit logs need the same attention you'd give pods in production. That's my blunt take: GA is the right call, but teams that move fast without tightening runtime controls will regret it.

What else Google Cloud shipped this week

  • GKE / Compute Engine: Google highlighted a new network-optimized Compute Engine machine family intended for high-throughput networking and block I/O. These instances are aimed at network-bound GKE workloads and other data-path-heavy services; expect to colocate network-optimized nodes for throughput-sensitive services while keeping cheaper families for general compute.

  • AI platform updates: Google expanded Gemini-related capabilities across Vertex AI, the Gemini API, and developer tooling, and rolled out Gemini-powered code-assist features in IDE integrations. Regional availability for enterprise AI features also continued to expand, and Model Garden added additional partner models from other vendors.

  • Infra tooling and infra-as-code: Google pushed updates to location and placement tooling to make region/zone data easier to consume for placement decisions. Config Connector and other infra-as-code tools received routine updates. There were no headline list-price changes this week �14 the product moves are about placement, machine families, and serverless patterns, not a pricing reset.

What platform teams should actually do

Two practical shifts. First, add observability and identity gates for worker pools now: workload identity, short-lived credentials, and structured audit logs belong on day-one runbooks. Second, update cost tagging and placement policies  Cloud Run worker pools change where and how compute is billed and autoscaled, and new network-optimized machine families give you a lever for network-bound cost/performance tradeoffs.

If you're curious for implementation details, there are good writeups already: see our prior piece on Cloud Run worker pools GA: run pull-based, non-HTTP workers as managed serverless and the roundup that covers Gemini and network-optimized VMs Gemini Pro preview, Cloud Run Worker Pools GA, and network-optimized VMs for GKE.

Final take: GA for worker pools is overdue and inevitable. It's the right product decision  managed pull semantics were the missing primitive for serverless-first platforms  but it forces teams to treat managed workers like first-class compute: lock down identities, instrument them, and think about cost controls. If your platform roadmap still assumes "background jobs live in K8s forever," this release should change that plan.

Sources

cloud-rungkegeminigcp
← All articles
GCP

Cloud Run worker pools GA: run pull-based, non-HTTP workers as managed serverless

Cloud Run worker pools are GA for pull-based, non-HTTP workloads—move queue consumers off K8s/VMs to managed serverless, changing networking, identity, cost.

Jul 30, 2026·3mcloud-rungke
GCP

GKE 1.33/1.32: Channel defaults, deprecations, and the push to Gateway API 1.5 for inference

GKE channel defaults move control-plane targets to 1.33; Gateway API 1.5 and GKE Inference Gateway are promoted as the standard AI inference ingress pattern.

Jul 29, 2026·3mgkegke-inference-gateway
GCP

GKE 1.36.0-gke.4681000: COS security patches and channel defaults shifting to 1.35/1.36 (mid‑July 2026)

GKE 1.36.0-gke.4681000 (with 1.35.6/1.34.8) applies COS security patches across 1.30–1.36; enforce node-image policy and revalidate AI and KMS paths now.

Jul 28, 2026·3mgkekubernetes-updates