Releases, deprecations, papers and tools worth an engineer’s attention, each entry a markdown file in the devopsdotng/content repository, each edition assembled from committed updates.
In the cloud projects I have reviewed this quarter, more new workloads are instrumented with OpenTelemetry from day one and shipped into Azure Monitor with shared trace and metric conventions. This reduces first-incident blind spots and avoids late-stage migration pain from ad hoc agent mixes.
In my experience, teams that start with one telemetry standard debug faster when incidents hit.
From what I am seeing in current delivery teams, prompt and model evaluation runs in Azure AI Foundry are becoming part of normal release checks rather than a one-time setup task. The practical shift is repeatability: baseline datasets, tracked regressions, and approval gates tied to measurable quality thresholds.
It turns subjective quality checks into a release policy that can be audited.
I validated this against recent cluster changes: vertical resize of CPU and memory without recreating the pod is now on by default. The resize subresource is stable and the old feature gate is a no-op.
Ends the restart-to-rightsize dance that made VPA unusable for stateful > services.
aws_iam_role.inline_policy and the matching user attribute are removed. Migration is to aws_iam_role_policy resources, and the provider emits a plan-time error rather than a warning.
In my older modules, this is a breaking upgrade, not a simple version bump.
I tested the update against our billing model: the exporter now reads DCGM metrics directly and emits per-namespace GPU-second cost, which closes the gap that made shared accelerator pools difficult to bill fairly.
Answers the most-voted FinOps question on this site almost exactly.
I pulled the methodology and read through the released taxonomy: thirteen years of internal reviews were classified and published under CC-BY. The finding I keep citing is that configuration changes account for more user-visible impact than code deploys.
I am using this framing to get config changes treated as releases.
I tested the release notes against our policies: budgets now accept cron windows, so consolidation can be suppressed during business hours without disabling it entirely. Node pools also gain a reserved-capacity provisioning mode.
The workaround I was scripting is now a first-class field.
From the latency checks I ran this week, peering at IXPN now covers two of the three major providers. Early measurements from Ikeja put round-trip to origin at 8-11ms against 62ms via Europe.
This changes my sizing assumptions for self-hosted CI runners and data-residency tradeoffs.
I verified that both scanners now emit and verify the same in-toto attestation, so a signed SBOM produced by one is consumable by the other in admission policy.
This removes one bespoke conversion step from my supply-chain pipelines.
In my review this week, the strongest result was that embedding refresh, not inference, dominates steady-state spend once a corpus passes roughly ten million documents.
I would read this before committing to a nightly reindex.