Moving the freeze watchdog into a Web Worker gets the report out of a dead tab

Session replay has two dependable capture paths ^[2]. One is a sampled slice of traffic. The other is any session where a JavaScript error fires ^[2]. The failures in the dev.to writeup fire nothing: a tab that degrades over hours, a scroll that stutters, a drag that lags, a main thread that stops answering ^[3].

So the sample is all that is left. The author uses 10 percent and says outright that the figure is illustrative, not any vendor's exact default ^[4]. Take it at face value and roughly nine complaints in ten arrive with nothing to watch back ^[5]. For that estimate to transfer, your plan's real sampling rate has to sit near 10 percent and your tool's replay cap has to be long enough to cover the whole session; the writeup gives about an hour as its cap, illustrative again ^[6]. An hour is plenty for a checkout flow and useless for a tab that has been open since Tuesday, where the part you want is the slow run-up before the freeze ^[6].

The second gap is the unit of measurement. Error monitors and APM tools are built around transactions: a page load, a route change, an API call, each bounded, each rolled into percentiles on a dashboard ^[7]. That shape came from servers, where a request arrives, does its work and leaves ^[7]. Percentiles cannot be traced back to one user ^[8].

The freeze detector is the part I would copy. A monitoring script inside the page runs on the same main thread that froze ^[9], so once the thread stops it cannot send anything. The watchdog instead lives in a Web Worker, survives the freeze, and posts the report itself from a snapshot the page handed over while the tab was still healthy ^[10]. The report is accurate as of that handoff, not as of the freeze ^[11]. The writeup also says to send the instant you know instead of batching ^[12], and batching is what would leave the payload queued on a thread that will not run again ^[19].

Frustration signals are the third gap. Rage-click and dead-click detection ship in current tools ^[18], and they detect frustration through clicks ^[13]. Scroll jank and sustained drag stutter produce no click, so each needs its own small detector that samples the gesture with timestamps ^[13]. That is two detectors you write and own.

Keeping every session means paying to store every session, which the author says is affordable only on infrastructure you already own ^[15], with the recorder assembled from tooling most teams have ^[16]. On sampling, the post is blunt: the fix "is not a cleverer sampling rule. It is full fidelity: keep every session" ^[17]. INP is not a substitute either. The writeup lists dead and rage clicks, scroll jank, drags and the merely-bad interaction as outside what INP measures, and argues that per-session logs match how people actually complain ^[14].