Release Notes#
Release notes and known issues for Nsight Operator.
Current Release (26.3.1)#
What’s New#
Nsight Operator 26.3.1 gives you more control over when and how you profile, a clearer configuration model, easier-to-read results in the web UI and CLI, and more flexible ways to deploy in restricted environments. See the NGC collection announcement for the full release.
Clearer Configuration#
We renamed and reorganized a number of Helm values and
NsightOperatorProfileConfig fields so your configuration is
clearer and easier to work with.
Upgrading takes no extra work: your existing values files and
ProfileConfig resources keep working as-is, because the old names are
still accepted as deprecated aliases. Move to the new names whenever it
suits you.
Profiling setup is now split into two clear parts.
nsightToolConfigsdescribes what to run – the profiling tool and its settings – whileinjectionRulesdescribes which workloads to profile and how.The settings that belong to a particular workload now live with its injection rule, including which processes to include or exclude (
processIncludePatternsandprocessExcludePatterns), any extra volumes to mount, and where results and logs are written.OpenTelemetry options now use OpenTelemetry names instead of the older “OTLP mirroring” wording:
otlpMirroringEnabledbecomesopentelemetryTrace,otlpMirroringGracePeriodSecbecomesopentelemetryTraceGracePeriodSec, andotlpProxyPortRangebecomesopentelemetryProxyPortRange.
Schedule and Automate Profiling Runs#
Start profiling on a delay or at a specific time, and stop it automatically after a set duration or at a chosen time, so you can capture the right window without watching the job. Use
--delay,--start-at,--duration, and--stop-atwithprofiler-start.The
statuscommand shows when a scheduled run will start and stop.
See Nsight Cloud CLI.
Launch the Nsight Streamer from the CLI#
Open the interactive Nsight Streamer for a session directly from the CLI with
nsight_operator.py analysis streamer; it opens in your browser. It uses your current session by default, or pass--session <id>to open a specific one. Previously the Streamer could only be launched from the Cloud UI.
Session-Scoped Analyses and Reports#
The web UI and CLI now show by default only the analyses and reports for the session you are working in, so you are not sifting through results from other runs.
Filter reports by session from the CLI
list reportcommand.
Custom Labels and Annotations#
Add your own labels and annotations – globally or per component – to everything the operator deploys, so the resources fit your cluster’s cost-tracking, policy, and monitoring conventions. Updates are applied automatically when you change your configuration.
More Flexible Deployment#
IPv6 networking: deploy in IPv6-only and dual-stack clusters.
Air-gapped installs: deploy in disconnected clusters by pointing every image at your own private registry.
Host networking: run the injector webhook on the host network for clusters that require it.
OpenTelemetry Improvements#
Use your existing OpenTelemetry pipeline by opting out of the operator-created OTLP resources.
Traces now record the node each span ran on, making it easier to tie performance back to a specific host or GPU.
The OpenTelemetry-to-Nsight converter now supports generic traces, so more workloads can be viewed alongside Nsight Systems profiles.
Other Improvements#
Recipe timers in the web UI reflect actual server-side run time.
The CLI returns proper exit codes, making it safer to use in scripts and CI pipelines.
For current operational constraints and platform support details, see Known Limitations and Notes.