Set min-slices and fix worker dashboard.

[rjpower]

No description provided.

[claude]

Claude finished @rjpower's task in 1m 32s —— View job

---

Review

This PR does three things: (1) bumps min_slices from 0 to 1 for two TPU scale groups in marin.yaml, (2) replaces the worker detail's inline Gauge-based utilization section with a grid-based panel using dedicated CSS classes and larger sparklines, and (3) fixes the Sparkline component to render with a single data point or all-zero data instead of returning null. Tests are updated accordingly and a new history-accumulation test is added. Changes look correct and well-structured.

Findings:

• worker-detail.js:14 — ResourceSection and Gauge are imported but no longer used after this refactor. Dead imports.
• test_dashboard.py — The import time inside test_worker_detail_sparklines_with_history (not inside a serialized worker function) is a local import that could be at the top of the file per AGENTS.md guidelines. The existing _hold_for_heartbeats local import is fine since it's serialized to a worker.

No correctness issues or behavioral regressions found.

[rjpower]

@claude address PR feedback:

worker-detail.js:14 — ResourceSection and Gauge are imported but no longer used after this refactor. Dead imports.
test_dashboard.py — The import time inside test_worker_detail_sparklines_with_history (not inside a serialized worker function) is a local import that could be at the top of the file per AGENTS.md guidelines. The existing _hold_for_heartbeats local import is fine since it's serialized to a worker.

[claude]

Claude finished @rjpower's task in 1m 39s —— View job

---

Addressing PR feedback

• Remove dead imports (ResourceSection, Gauge) from worker-detail.js
• Move import time to top of test_dashboard.py
• Run pre-commit and tests
• Commit and push

Both issues fixed in 92b43b8:

1. worker-detail.js:14 — Removed unused ResourceSection and Gauge from the import.
2. test_dashboard.py — Moved import time to the top of the file (lines 179 and 249 had local imports in test functions that aren't serialized to workers). The import time inside _hold_for_heartbeats and _allocate_memory_and_wait was left as-is since those are serialized to workers.

Testing: Pre-commit passed. Test collection for test_dashboard.py verified (e2e tests require Playwright + a running cluster so were not executed end-to-end in CI).

[yonromai]

Approved, with one high-priority config regression to address before merge: the min_slices change in marin.yaml keeps five zoned TPU slices warm at idle. The dashboard and sparkline changes themselves look good, and the targeted worker-detail E2E tests passed locally.

Generated with Codex.

[yonromai]

Priority: High

Setting min_slices: 1 here, and again for tpu_v6e_4, turns these zoned groups into five permanently warm 4-chip TPU slices, not one global warm pool. load_config() preserves explicit min_slices on every zone-expanded group, and the autoscaler enforces that floor even with zero demand, so this change stops these pools from scaling back to zero. If the goal is to reduce cold-start latency, this needs a single dedicated warm pool or some other global mechanism rather than min_slices on multi-zone groups.

Recommended fix: revert these min_slices changes, or replace them with a design that keeps at most one intentional warm slice instead of one per zone-expanded group.

Generated with Codex.

[rjpower]

Yes, this is okay and intentional, we want to keep slices warm.