Skip to content

benchmark-k9s.md

Latest commit

 

History

History
161 lines (130 loc) · 8.42 KB

benchmark-k9s.md

File metadata and controls

161 lines (130 loc) · 8.42 KB
Raw

Performance benchmark: sofka and k9s

This test compared sofka 0.24.9 and k9s 0.51.0 on 8 September 2026. Both programs used read-only mode. The results apply to this computer, cluster, configuration, and test method. They do not establish a general performance advantage.

The test script and raw results are part of this repository.

Results

Each TUI result uses ten fresh processes per program. The memory result uses the median of five RSS samples from each process, then the median across the ten processes.

Measurement sofka median k9s median sofka range k9s range
Large pod view 1514.08 ms 3715.73 ms 1257.58-1823.58 ms 3697.79-3955.64 ms
Apply pod-name filter 66.56 ms 627.76 ms 58.17-71.34 ms 608.23-649.02 ms
Clear filter 16.84 ms 73.81 ms 16.09-20.91 ms 70.52-74.61 ms
First StatefulSets view 109.98 ms 397.34 ms 103.98-155.42 ms 386.92-472.07 ms
Process RSS 317.02 MiB 734.38 MiB 316.73-318.16 MiB 675.52-747.91 MiB
Version command 5.29 ms 45.01 ms 4.01-7.27 ms 42.15-58.93 ms
Help command 5.45 ms 45.45 ms 4.39-8.73 ms 42.49-112.21 ms

sofka had lower medians for these operations in this test. The programs use different columns, filter rules, refresh schedules, and API requests. This test does not isolate the cost of any one implementation choice.

Environment and build

Item Value
Computer Apple M3 Max
Operating system macOS, Darwin 25.6.0, arm64
Rust rustc 1.97.0 (2d8144b78 2026-07-07)
tmux tmux 3.7c
Terminal xterm-256color, 180 columns by 50 rows
Pods before / after 1446 / 1446
Pod count threshold 1373
StatefulSets before 4
sofka source commit ef10167896dc334d3e58dd9f3be3e1479605c7d4

Build command: just build-release. This runs cargo build --release. Cargo.lock did not change. sofka uses thin LTO and removes symbols. k9s came from the installed Nix package. The raw results include executable hashes and version output. This compares these builds, not matched compiler or package settings.

File Bytes MiB
sofka executable 19,077,072 18.19
k9s executable 142,893,680 136.27

The k9s launcher is a 416-byte shell script. Its size is excluded from the executable size above. All k9s time results include this launcher. sofka starts directly. Binary size reflects packaging and included functions as well as implementation.

Commands and controls

Replace YOUR_CONTEXT with the context to test.

target/release/sofka --context YOUR_CONTEXT --readonly -A pods
k9s --context YOUR_CONTEXT --readonly -A -c pods --splashless --logoless
  • Both programs used the same kubeconfig and an explicit context. The test did not change the current kubeconfig context.
  • Both used --readonly. The script sent only filter, resource navigation, and exit keys. It sent no create, edit, delete, scale, shell, or port-forward command.
  • Each TUI process had empty, separate XDG config, data, state, and cache directories. The script removed SOFKA_* and K9S_* environment overrides.
  • An isolated tmux server used /dev/null as its configuration.
  • Ten pairs ran in sequence. sofka ran first in odd pairs; k9s ran first in even pairs.
  • The benchmark build and repository checks ran outside the measurement period.
  • The operating-system file cache was not cleared. These are new process starts, not cold computer starts. The live cluster and other computer activity were not controlled.

Measurement method

Python time.perf_counter() supplied the times. tmux capture-pane supplied visible screen text. The script parsed each program's resource title, including comma-separated k9s counts. It polled with a 5 ms delay between failed checks. A screen capture had a median cost of 5.71 ms in 50 samples. The reported times include input commands, process creation where applicable, and screen observation delay. Small differences near these costs are not useful evidence of a product difference.

Large pod view: the timer started before tmux replaced the pane process. It stopped at a visible count of at least 1373 and at least 20 pod status rows. The threshold was fixed at 95% of the pod count before the test. It does not mean that all pods or metrics were loaded. The raw results include the count observed at the end of each start.

Memory: the script waited five seconds after the pod threshold, then sampled ps -o rss= -p PID five times at target intervals of 0.25 seconds. macOS reports RSS in KiB; the script divided by 1,024 for MiB. Samples within a process are not independent trials. RSS includes shared resident pages and is not private heap size. Five seconds does not establish a steady state or long-session use.

Filter: the script selected one running pod with a unique name before the test. It sent /, the full name, and Enter. The timer started before / and stopped when the table count was one and the running pod row was visible. This includes input before Enter, since filtering can start while typing. The programs use their default filter rules; this is a task comparison, not a comparison of the same matching algorithm. The clear timer ran from Escape to restoration of the pod threshold.

StatefulSets: each process opened this resource for the first time after the filter test. The timer ran from the initial : through statefulsets and Enter, until the resource title showed at least 4 objects. It includes API response time. It does not require complete metrics.

Version and help: each command had one warm-up and 100 measured runs. The script alternated which program ran first. Output went to the null device. These commands do not connect to Kubernetes. They produce different output and do not measure TUI response. Commands: sofka --version, k9s version --short, sofka --help, and k9s --help.

The raw data includes p95 values, calculated by linear interpolation at rank (n - 1) * 0.95. Ten process trials give only a limited estimate of variation; do not treat their p95 values as a reliable tail-latency estimate.

Scope and excluded attempts

The final run completed 20 process trials, with 0 recorded errors. Setup runs were excluded. An incomplete attempt was also excluded because its count parser did not accept commas. Both programs were tested again after the parser was corrected.

After the final twenty TUI trials, cleanup of an already closed tmux server failed. All TUI samples were already saved. The cleanup step was corrected, and only the command measurements and final pod count were repeated with --finish-commands. No completed TUI sample was removed. Use the same context when this option is needed; the result file does not store context identifiers.

This report replaces the earlier sofka 0.16.3 comparison. The cluster population and input timing method changed. Do not use the old and new results to claim a performance improvement or regression.

This test did not measure CPU under a fixed event rate, API request counts, network bytes, garbage collection pauses, long sessions, or other operating systems. It did not replay a fixed object set. Row movement was omitted because background screen changes can trigger a false result. Raw status-text matching was omitted because it does not establish that a row is visible.

Run the test again

just build-release
python3 scripts/benchmark-k9s.py --context YOUR_CONTEXT --pairs 10 \
  --output docs/benchmarks/NEW-RUN.json

Use --sofka and --k9s to select other builds. The script requires an existing cluster with at least 22 pods and one StatefulSet. It stores timings, counts, and build metadata. It does not store pod names or screen captures. Each new run uses the selected cluster's current objects; it is not an exact replay.