Skip to content

v2.11.0 — Pod shell session recording + playback

Choose a tag to compare

@vurulkan vurulkan released this 08 Oct 16:40
· 30 commits to main since this release

Every pod shell (exec) session is now recorded as an asciicast v2 file
and can be replayed by admins in the browser.

Recording

  • Terminal output (stdout + stderr) and resizes are streamed to
    /<session_id>.cast. Typed input is not recorded as
    input events; whatever the shell echoes back is part of the output.
  • The shell shows a red "This session is being recorded" banner that
    warns that secrets printed to the screen are stored too. The banner
    appears only when the server confirms the session is recorded.
  • Fail-open: a disabled, full or broken recorder never blocks the
    shell; the session just isn't recorded and the reason is audited.
    The 5-minute idle timeout is unchanged.

Admin → Recordings

  • Settings, editable without a restart: on/off, retention days,
    per-session cap, total quota, free-space floor, disk policy. The
    panel shows disk usage and warns when the limits are hit.
  • List: filters for user, namespace, pod and date range, server-side
    paging, in-progress / truncated badges, Play / Download / Delete.
  • Player: asciinema-player bundled from npm and lazy-loaded (no CDN).
    Play/pause, seek, 1x/2x/4x, skip idle time.

Disk protection

  • Recordings share the /data PVC by default. A total quota (2 GB) and a
    free-space floor (512 MB) are checked before every session and every
    30 s while sessions run, so the volume never fills up under app.db.
  • Disk policy evict_oldest (default) deletes the oldest recordings to
    make room, audited as recording.evicted. Policy stop keeps them and
    opens new shells unrecorded.
  • A daily retention purge also removes orphan .cast files. After a
    crash, the next start closes the recordings that were left open.

Audit

  • pod.exec.session.recorded (session_id, size, dur_ms, truncation)
    carries the same request_id as pod.exec.start / end.
  • Also: pod.exec.session.record_skipped / record_failed,
    recording.view / download, recording.delete.,
    recording.settings.update.
    , recording.evicted, recording.purge.
  • recording.delete.* records whose recording was deleted (owner, start
    time, size, cluster, container). The row and the file are gone
    afterwards, so the audit entry is the only remaining trace.
  • Every recording action (view / download / delete) also writes its own
    "recording.action" console log line with request_id. This line does
    not depend on LOG_INCLUDE_AUDIT.

API (admin only)

  • GET /api/admin/recordings, GET /api/admin/recordings/{id},
    GET /api/admin/recordings/{id}/cast (?download=1),
    DELETE /api/admin/recordings/{id},
    GET / PUT /api/admin/recordings/settings

Deploy

  • New env vars SESSION_RECORDING_ENABLED / DIR / RETENTION_DAYS /
    MAX_SIZE_MB / MAX_TOTAL_MB / MIN_FREE_MB / DISK_POLICY. All except
    DIR only seed the settings on first boot; the admin UI owns them
    after that.
  • deploy/deployment.yaml documents them (commented). deploy/pvc.yaml
    and deployment.yaml include a commented, optional dedicated
    recordings PVC.
  • CSP script-src adds 'wasm-unsafe-eval' for the player's WebAssembly
    terminal. It allows WebAssembly compilation only, not JS eval.
  • No ClusterRole change is needed.

Fixes

  • The active cluster is now reconnected after a restart. Before this
    fix, a cluster activated through the multi-cluster API was not
    restored on boot, and the dashboard showed "Service Unavailable"
    until an admin re-activated it.
  • The Admin page title is translated (it was always the English tab
    label).

Image: ghcr.io/vurulkan/kubernetes-openconsole:2.11.0 (also :latest), linux/amd64.

🤖 Generated with Claude Code