Take the two rates together. Five hundred events a second is 30,000 a minute, and 20MB of heap growth across that minute [1] works out to roughly 0.7KB retained per event [1]. That is the right order of magnitude for a small record plus its proxy handlers plus a queued render job, so the account is at least internally consistent, though it is still arithmetic about someone else's payload.
Run the same rate out to the stated freeze at the two-hour mark and it implies about 2.4GB of net retention [2]. For that to be literal, the collector has to be reclaiming close to nothing, which means a live reference still reaches every batch ever ingested. The dev.to walkthrough attributes that to internal proxy references lingering across component re-renders and closure boundaries [13]. An unreachable proxy is collectable like any other object, so the retaining edge has to come from somewhere, and the article's own cause list supplies two candidates: index keys that confuse the patch algorithm during high-frequency updates, and child components registering global listeners they never remove in onUnmounted [7].
The proxy count reads as a real cost. The retention does not read as a leak. Five thousand records producing tens of thousands of wrappers [5] implies at least four proxies per record [5], because the multiplier is per nested object rather than per row: in the sample interface the metrics object and the tags array each get their own traps [15]. Every update then walks that structure, schedules queue work, and pushes the virtual DOM to diff the whole tree [6]. That is a tax proportional to live data, which is a different failure from unbounded retention of dead data.
The replacement ingest path is cheaper per item and dearer per call. It assigns a fresh array by spread on every call, and once length passes 2,000 it reassigns again with slice(-1500) [10]. Under a 500-per-second feed the buffer burns its 500 records of headroom in one second [3], so the trim fires roughly once a second on top of one whole-array copy per call. Call ingestBatch per event rather than per window and you are copying up to a million array slots a second on the main thread [4]. Windowing the socket is what makes that copy affordable, and nothing in the pattern does it for you.
Note also that capping the buffer at 1,500 records would cut memory under deep reactivity as well [10]. It is a capacity decision that happens to travel alongside the shallowRef refactor [8], independent of it.
For 20MB a minute to be your number, the account needs nesting comparable to the sample record [15], an append path that treats the stream as an immutable collection and never drops old rows [12], and rows that survive in the v-for as real DOM nodes, which is the retention the article pairs with virtualisation in its own title [14]. Change the third and the proxy fix stops being the dominant term.
Heap snapshots are tedious, and here they are the only witness: the detached-node count filtered by constructor is the thing that separates the two failures [11].
