Memory Leak Investigation — src/live/index.html

Symptom

Chrome tab memory grows from ~150MB to ~500MB+ over 30 minutes. Observed on /live/?kiosk page.

Root cause

Globe.gl's ring digest never releases ring data objects. Every ring object passed to globe.ringsData() is retained in the digest's internal Map permanently, even after being removed from subsequent calls.

Verified with FinalizationRegistry:

Before fix: ring=52/0 (52 created, 0 ever garbage collected)
After fix: ring=291/263 (normal finalization)

Bug report filed at ../globe.gl/BUG.md.

Fix

Replaced dynamic ring creation/destruction with a fixed object pool. The pool is passed to globe.ringsData() once at init. On updates, we mutate pool objects' lat/lng properties in place. Since object identity never changes, the digest's add/remove path is never triggered, avoiding the leak entirely.

Additional fix: Tween leak

globe.pointOfView(target, duration) accumulated Tween objects in an internal group that never pruned finished tweens. Each call with duration > 0 added a tween permanently. In kiosk mode this was called on every shield event.

Replaced with smoothPointOfView() — custom requestAnimationFrame animation that calls globe.pointOfView(coords, 0) (duration=0 = no tween created).

Investigation approach

Systematic isolation via console overrides, monitoring performance.memory and renderer.info.memory every 30s, and FinalizationRegistry tracking.

Round 1: Isolate subsystem

Test	`js:total` trend	Conclusion
Baseline	+5–25 MB/min, accelerating	Leak confirmed
`eventSource.close()`	Flat	Leak tied to SSE event handling
`eventSource.close()` fresh	Flat at 132	Render loop, audio, clock don't leak
`onmessage = noop`	~0.3 MB/min	Not EventSource buffer
`onmessage = parse-only`	Flat	Not JSON.parse

Round 2: Isolate handler function

Test	`js:total` trend	Conclusion
`updateEventHistory = noop`	Growing	Not DOM updates
`smoothPointOfView = noop`	Growing	Not camera animation
`processShieldData = noop`	Flat	Leak is in processShieldData
`processEcomscanData` alone	Flat	ecomscan not leaking

Round 3: Isolate within processShieldData

Test	`js:total` trend	Conclusion
`globe.ringsData = noop`	Growing	Not just ringsData call
All globe methods + handlers nooped	Growing	Not globe.gl API calls
All nooped, fresh reload	Flat at 128	Contaminated by prior state

Round 4: Object-level GC tracking

Object type	Created/Finalized	Leaking?
Arc	52/44	No
Glow	52/44	No
Ring	52/0	Yes — never GC'd

Round 5: Confirm ring digest is the cause

Test	Ring GC	`js:total`	Conclusion
`globe.ringsData = noop`	68/51 (normal)	Flat	Ring digest retains objects
Ring pool fix (mutate in place)	291/263 (normal)	Stable ~110MB	Fixed