Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

Free Kubernetes Client Sextant Cut Memory per Object from 26 KB to 5.2 KB

Sextant, a free Rust-based desktop Kubernetes client and a Lens alternative, no longer stores every cluster object as a bloated JSON tree. Previously, a pod of about 2.5 KB took 26 KB in memory; now it takes 5.2 KB. For 50,000 pods, that means 127.5 MB instead of 481 MB. The trade-off is speed: loading the full list takes 10.3 s instead of 9.0 s. For tools built for large clusters, memory and speed should be measured together.

AI-processed from Habr AI; edited by Hamidun News
Free Kubernetes Client Sextant Cut Memory per Object from 26 KB to 5.2 KB
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1The idea came from an external pull request, one of the first the project has received.
Key metric #2The first table rows appear in 94 ms instead of 77 ms.
Key metric #3The measurement was made on a debug build; comparisons with Lens and Freelens are still ahead.

If you have opened a desktop Kubernetes client on a large cluster, you know the problem: a window that simply shows a list of pods gradually eats gigabytes of memory. Sextant, a free Lens alternative built on Rust and Tauri 2 with no account, subscription, or telemetry, was lagging behind competitors on this metric. The author admitted as much in the first article about the project. Now, after a week of 13 releases (from 0.20.0 to 0.25.5, October 21 to 28), he has paid off that debt. The idea for the fix came from an outsider.

What was wrong with memory

Previously, the Sextant core stored every cluster object as a JSON tree. That is convenient for development, but wasteful. Imagine writing an address on a sheet of paper, but having to store it in a cabinet where every word gets its own cell: the meaning is the same, but you need many times more space.

In version 0.25, objects are stored compactly. The measurement was made on a test with pods of about 2.5 KB each:

  • memory per object: 26.0 KB → 5.2 KB, five times less;
  • stored objects for 50,000 pods: 481 MB → 127.5 MB, about 3.8 times less;
  • full list of 50,000 pods: 9.0 s → 10.3 s;
  • first table rows appearing: 77 ms → 94 ms.

The author states the cost plainly: the full list became slightly slower, by 1.3 seconds. While the table fills, the user sees rows after 94 ms instead of 77 ms, a difference that is barely noticeable.

Where the idea came from and what else shipped

This week was the first time external pull requests went into releases. Among them were table filters, sidebar settings, and a Windows fix for Helm console windows flashing on screen. The idea for compact object storage also came from an external PR, and after it, memory per tracked object dropped about fivefold.

Part of version 0.25 grew out of reader requests. These include search across all columns, tabs for pods and node resources, and operation without permission to list namespaces. The last one will be familiar to many people at large companies: a developer has access to a couple of namespaces but not to the list of all of them. Now the "All namespaces" option means "all accessible ones," and instead of an error, a "No access" screen appears with a "Show anyway" button.

In addition, integrations shipped for OpenShift, Operator Lifecycle Manager, k3s/RKE2, and vcluster. Alerts from Prometheus and Alertmanager appeared, along with logs from remote pods via Loki. For AI agents, version 0.25.5 improved the built-in MCP server: it can now work when the user has restricted access rights (RBAC). The trigger was a discussion in which a user connected an agent to a cluster with access to only a couple of namespaces and described where MCP got in the way. Each item became a separate issue, and all five went into the release.

How the code is written and what to take from it

All Sextant code is generated by Claude Code. The human sets tasks, makes architectural decisions, and does code review, and every change goes through repeated testing. Each feature has an issue with requirements and a wiki page.

The practical takeaway for those building tools for large clusters: storing raw JSON trees is expensive. A compact representation yields a multiple-fold memory gain for a small cost in time. Before accepting a trade-off like this, measure both parameters on realistic data volumes. Look at both total memory and how quickly the user sees the first data.

There are also limitations. The author stresses that the measurement was made on a debug build and on pods of fixed size. On release builds and in live clusters, the numbers may differ. A repeat comparison with Freelens and Lens on release builds is still ahead, so declaring victory over competitors on memory is premature. Still, the direction is clear: the project is free, will stay free, and is developing noticeably faster than one might expect from a personal experiment.

Story

Key Questions Answered

What was wrong with memory?

Previously, the Sextant core stored every cluster object as a JSON tree. That is convenient for development, but wasteful. Imagine writing an address on a sheet of paper, but having to store it in a cabinet where every word gets its own cell: the meaning is the same, but you need many times more space. In version 0.25, objects are stored compactly.

Where the idea came from and what else shipped?

This week was the first time external pull requests went into releases. Among them were table filters, sidebar settings, and a Windows fix for Helm console windows flashing on screen. The idea for compact object storage also came from an external PR, and after it, memory per tracked object dropped about fivefold. Part of version 0.25 grew out of reader requests. These include search across all columns, tabs for pods and node resources, and operation without permission to list namespaces.

How the code is written and what to take from it?

All Sextant code is generated by Claude Code. The human sets tasks, makes architectural decisions, and does code review, and every change goes through repeated testing. Each feature has an issue with requirements and a wiki page. The practical takeaway for those building tools for large clusters: storing raw JSON trees is expensive. A compact representation yields a multiple-fold memory gain for a small cost in time. Before accepting a trade-off like this, measure both parameters on realistic data volumes.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…