Skip to content

CloudHero Weekly / Issue 004

Rethinking isolation costs and cloud recovery

Developments from to

AWS’s new Nitro Sandboxes give teams paying for dedicated instances to isolate workloads a reason to revisit that capacity plan. OpenSearch’s managed recovery can remove replication work, while Revolut’s postmortem shows why recovery tooling needs its own resilience. For self-hosted AI agents, weigh control over session data against operating another production platform. [1], [6], [8], [9], [10]

Nitro Sandboxes put dedicated-node spending back under review

For platform teams running untrusted builds, agents or tenant workloads.

On 9 November, AWS launched Nitro Sandboxes: opted-in EKS pods can run in separate hardware-isolated VMs while sharing an EC2 instance. Teams separating workloads primarily for security now have an alternative to evaluate before buying more dedicated capacity. [1]

Availability covers supported instances across commercial AWS regions. There is no sandbox surcharge, but EC2 and applicable management fees remain. Savings depend on achievable workload density and operating overhead. [1], [2]

EKS Auto Mode includes the runtime; AWS’s launch walkthrough requires Kubernetes 1.36 or later. Outside Auto Mode, teams must build a sandbox-enabled AMI and enable Nitro Enclaves in the launch template. [3], [4]

Isolation does not reserve physical cores: CPU remains time-shared. Confidential mode restricts interactive debugging and requires policies that check attestation before releasing secrets. Customers retain responsibility for application images and security policies. [5]

Before expanding the fleet: start with workloads whose dedicated nodes exist for isolation. Compare startup latency, contention and operational access under representative load. Consolidate only if the result meets both the security requirement and the capacity budget.

OpenSearch removes replication work, but leaves the recovery runbook

For owners of business-critical search and analytics.

AWS’s 10 November managed cross-region disaster recovery launch could reduce custom replication work in an existing recovery project. It adds no feature surcharge, although the second domain and cross-region transfer remain billable. [6]

Eligibility requires same-account domains on OpenSearch 3.7 or later, with matching versions for switchover. Commercial regions are supported; warm/cold storage and Optimized-engine domains are excluded. AWS publishes a recovery point objective (RPO) below one minute and a recovery time objective (RTO) below 20 minutes. Application recovery must be tested separately. [6], [7]

The delivery estimate still needs configuration synchronization, failover initiation and application write redirection. Index templates, ingest pipelines and roles are not replicated. A promoted domain can contain the data but lack the configuration needed to process new writes. [7]

Proposed acceptance gate: in staging, reproduce the production pipeline configuration, promote the secondary and redirect application traffic. Require successful writes, expected transformations and correct permissions before accepting the recovery design. Querying copied data alone is insufficient.

Revolut’s unnoticed quorum degradation blocked failover

For SRE and database teams operating quorum-based systems.

Revolut’s 13 November postmortem adds evidence about Google Cloud’s 20 October London outage. Some five-member coordination groups already had one member offline without an alert. The zone failure removed two more, leaving two voters where three were required to promote a database primary. The main database-metrics visualization instance also depended on the failed zone. [8]

Normal application service had concealed reduced failure tolerance.

Before closing maintenance: verify restored coordination membership and zone-loss tolerance after replacing a member. In a proposed staging exercise, make one voter unavailable, then remove the zone containing the most remaining voters. The first pass condition is an alert reaching an operator before the second failure. Recovery tools must also remain accessible.

Self-hosted Claude agents shift control and operating work together

For enterprises that need agent state on their infrastructure.

Anthropic’s 12 November private beta extends self-hosting beyond tool sandboxes to the Managed Agents API, agent loop and session storage. It offers a packaged alternative to building that control plane. Model inference still runs at Anthropic or the selected provider. Hosting session history locally does not bring inference onto your infrastructure. [9], [10]

Customers run the server, database and workers, and install releases themselves. The self-hosted version lacks webhooks and MCP OAuth token-refresh flows. Teams that need those features must account for polling and credential-renewal work. [10], [11]

Before approving a pilot: require an operating plan with named owners for upgrades, database restoration and credential renewal. Approve destinations for model requests. Confirm beta access and required features before committing delivery dates.

If the requirement is only local tool execution, evaluate the narrower self-hosted sandbox option first.

Worth watching: EU cloud procurement proposals extend to utilities

For European utilities and procurement teams.

An EU Council report dated 13 November describes negotiating changes discussed on 10 November under the proposed Cloud and AI Development Act. Cloud public-procurement provisions would extend to contracting entities covered by Directive 2014/25/EU, including electricity and water suppliers. A proposed delegated power to extend risk-assessment requirements more broadly to private entities would be removed. [12]

These remain proposals. For long-lived cloud contracts, procurement and legal teams should track entity scope and evolving assurance criteria before fixing tender requirements.

What deserves attention first

  1. Quorum-based services: verify membership and zone-loss tolerance before the next maintenance change.
  2. Dedicated isolation capacity: platform owners should request a workload-specific consolidation comparison before expanding the fleet.
  3. OpenSearch recovery projects: revise replication estimates, retaining configuration parity and application-write testing as acceptance gates.
  4. Agents requiring local state: approve a pilot against an operating plan and permitted inference destinations.
  5. EU utility tenders: monitor proposed scope before setting long-term supplier requirements.

Sources

  1. AWS, 9 November 2026: Nitro Sandboxes launch
  2. AWS, 10 November 2026: Nitro Sandboxes architecture, pricing and availability
  3. AWS, 9 November 2026: Nitro Sandboxes on EKS and version prerequisite
  4. AWS documentation: EKS sandbox setup requirements
  5. AWS documentation: Sandbox modes, attestation and shared responsibility
  6. AWS, 10 November 2026: Managed OpenSearch cross-region disaster recovery
  7. AWS documentation: OpenSearch recovery requirements and replication exclusions
  8. Revolut, 13 November 2026: 20 October outage technical overview
  9. Anthropic, 12 November 2026: Self-hosted Claude Managed Agents announcement
  10. Anthropic documentation: Self-hosted Managed Agents responsibilities
  11. Anthropic documentation: Differences from the hosted service
  12. Council of the EU, 13 November 2026: Cloud and AI Development Act progress report, sections 8–9

Continue the weekly briefing

Follow CloudHero Weekly

Cloud, AI and DevOps developments assessed against reliability, security, cost and delivery decisions.

Follow the weekly briefing

Contact CloudHero

Tell us what you're working on

Our sales team replies by email within 1 business day.

Topic

We use your details only to reply. Privacy policy