Two jobs arriving at the Jetson together do not degrade gracefully. The pool is sized for one loaded model or one active render, so a second loader puts the box into thrashing and neither job finishes [15]. That is why nothing addresses the Jetson directly and everything lands in a durable queue with a gateway in front of it [17].
Where the gateway lives is what picked the implementation. The Pi 5 carries DNS and dozens of small services that cannot go down, and the author's own rule is that nothing on it may pin the CPU long enough to take DNS with it [16]. Then the memory: 5 GiB in use plus 4.4 GiB of swap is 9.4 GiB of committed pages before a scheduler exists at all [2][1]. An orchestrator that runs its own agent and control loop on that box is competing with DNS for pages the box does not have, which is the basis on which the write-up says k3s and Nomad were dropped [2].
The replacement keeps systemd for the part systemd already does well. Enforcement stays in the kernel: each job runs as a subprocess in its own systemd-run scope, each resource class maps to a slice with memory and CPU limits, and an OOM kills the one job in that slice rather than the host [7]. Only arbitration is bespoke: four priority classes from interactive to best-effort, a manifest-driven schedule, exponential backoff, a dead-letter path for anything a human has to look at [5], and parking instead of burning retries when a dependency is down [6]. The isolation is borrowed while the admission logic is written by hand, and that is the right ratio between the two.
The unfinished part is the deadline promise. The scheduler will only start a queued job ahead of a fixed window if it can promise the job finishes before that window closes [10], and the first estimator was the median of the last five runs [11]. One job is bimodal: about twenty minutes when it finds work, about five seconds when it does not [12]. Two of its three recorded runs were quiet, so the estimate landed at six seconds against a real run of twenty-one minutes, understating it by roughly 210x (1,260 seconds divided by 6) [12][2]. On that number the scheduler would have started the job thirty seconds before a workday window, about 1,230 seconds short of what it promised [13][3]. Deleting the two quiet runs held the number for twenty minutes, until a fresh quiet run pulled it back down [14]. Two quiet runs out of three are enough to make a median confident and wrong.
The RTX 4070 Ti Super box marks where the gateway's authority stops. Its render work never enters the queue, because it was never competing for the Jetson's model slot, and routing it through would only add latency [21]. The one LLM call both machines need does queue, because for that call they are contending [21]. That is also the transfer test. An admission gateway earns its code when one indivisible resource has more than one claimant and the control plane has no memory to spare. If your accelerator holds two models, or your scheduler host is not swapping, the k3s comparison comes out the other way. Three machines here cost roughly what a decent laptop does [22], and at that budget a scheduler's own resident set is a first-order cost rather than a rounding error.
