`req` is a namespace and a name, not a diff and not a copy of the object [8]. Every invocation re-reads the desired spec from scratch, recomputes what all three children should look like, and hands that to `controllerutil.CreateOrUpdate` to diff against whatever exists right now [9]. The work is identical whether the trigger was a kubectl command a millisecond ago or a process that just booted after eight hours down [10].
That is what makes a filter survivable. The architecture principles in the Kubernetes design proposals archive require level-based behaviour, correct given desired and observed state regardless of how many intermediate updates were missed, with edge-triggered behaviour "must be just an optimization" [11]. So a dropped event is a delayed repair rather than a lost one, and there is no catch-up path to forget to write because there is only the one path [12]. A predicate can throw events away precisely because nothing downstream depends on having seen them.
Which is also why the 48.5% is a number about that harness's event mix [6]. The controller under test reconciles one CRD into three children and writes observed state back to `Echo.status` [1][2][3]. The share of steady-state updates the predicate discards depends on how much non-spec churn your own objects generate. An operator whose status block is rewritten by a hot metrics loop, or whose children are annotated by three other controllers, has a different ratio, and the predicate's saving moves with the ratio, not with the predicate.
It reports that a short resync period costs zero additional API requests, because a resync is not a re-check against the cluster [5][13]. Grant that, and the no-op reconciles the predicate removed were already free at the API server. What is left is controller-side: CPU, workqueue depth, and whatever writes a reconcile issues when it does find drift. On the evidence in this material, the 48.5% is a CPU and queue lever [21]. The API-server case is the part the text does not reach: it asserts that a controller does not poll and that what it does instead is where API load comes from, then breaks off before saying so [20].
Sizing the CPU side from the reported mean: 2.71ms per reconcile is about 369 back-to-back invocations a second on one worker [4][18]. That ceiling only holds with no queue wait, and it is measured on macOS on Apple Silicon under kind v0.32.0 with controller-runtime v0.24.1 [7]. 77 timed invocations bound a maximum at 25ms [4]; they cannot resolve a tail you would care about at a thousand times the volume.
Ten injections per failure mode across the five failure modes in the summary is 50 drift events, and the material does not say how those map onto 77 timings [15][16][17]; the raw data is in the repo [19]. A second gap sits next to it: the author calls a predicate a filter with a silent cost whose cost is not the one the documentation warns about first, without printing it here [14]. Until that cost is measured, the 48.5% carries an unknown offsetting cost of its own.
