I ended up with quite a collection. Small 2-drive NAS-ish DLink DNS-320's were great. IoMega IConnect had a solid (and user-replaceable) wifi chip & numberous USB2 ports, & was very afforadble on ebay. These kind of came latter-ish, but there were great other botique SBC or simple-device options pre-ceding GuruPlug: SheevaPlug, GuruPlug, OpenRD.
Many years have elapsed. But the upstreaming work on all these machines was very great. They required very little customization. Really hope I haven't messed up the attribution here, but I believe Marvell paid Free Electrons (named Bootlin at the time) to do a ton of the driver work. It's in kernel. Each device took tended to require only a little DeviceTree configuration to work great. UBoot got upstream support, & I'd replace the onboard uboot with a modern far more featureful uboot, greatly opening boot capabilities versus the minimal/older uboot devices shipped with (usb boot? yes. ethernet boot? yes.). Devices would slowly gain a couple modest not-super-important-to-me capabilities a year, or better tuning.
I've given away, sold, & (sadly) thrown out a bunch of these devices over the years. I still have two NAS running in friends' houses as off-site connected backup devices, humming along, but mostly ignored. Still, I fully expect, if I wanted to, I could completely freshen nearly any of these systems up. Brand new kernel, totally new Debian (actually many do run modern Debians, just on old kernels). My vibe has been very good, that old hardware doesn't die, doesn't get dropped, that the graveyard is merely of my interest to give a shit. Not all SBC are such: the Marvell Kirkwood was an early 2010's vibe of hope & promise, Linux uber-alles, where some good hackers got paid to Do The Right Thing, and I think this post is a testament to how obvious, how clear, how simple that path was. These devices are still fully supported. They still work great, kernels still support them, because Marvell worked with people who knew what the heck they were doing to get support upstreamed. These devices may well never ever die. They may be able to last forever. Because someone didn't just dial it in, because it wasn't just an internal team with some heavily forked kernel shipping shit: Marvell did the right thing, they upstreamed their work. The value of that, right now, seems like it may be ever-lasting.
The graveyard isn't necessary. Good vendors upstream their stuff & it will keep working.
It's also notable & exciting to see, as I've mentioned else-where on HN recently, Qualcomm chips have gotten much more upstreaming recently. The Chromebook folk really lead the charge with do-the-right-thing-duh are-you-stupid-of-course-we-will path. But after getting a bunch of slightly upscale tablet-grade Qualcomm chips supported upstream, we're starting to see much broader classes of Qualcomm chips getting supported, & seeing growing device-trees for a variety of phones. Turns out it kind of mostly is the same hardware, no surprise, with different cores, and the decades of shitty unsupportable heavily-forked-and-useless Qualcomm-SDK horse-crap they've shovelled at people is & always has been a farce & once the ball started rolling towards upstreaming it's just kept snowballing, mostly with little thanks to Qualcomm themselves, but just just just recently, the company seems to be a adopting a not so universally useless sad pathetic view & actually stepping up every now & then to actually do a little bit of the upstreaming themselves. For a company famously reputed to require lawyers to sign off on literally every line of code, it feels like the bleak nightmare shitworld psychopathy anti-participation might finally be lifting a little bit, that good might finally be allowed to escape out into the world.
----
Regarding ESP32's & microcontroller kit- I agree, there's tons of sufficiency in microcontrollers. But that shit is just a brutal crazy world. Having a Linux capable system that can run a perfectly normal systemd is worth it's weight in gold. If you can ssh into a regular box, scp some of you python scripts in, setup a systemd.service file, and it just works, it makes the whole ESP32 dance look absurd & stupid. And it is easily within reach. I hope some-day the embedded OSes like Zephyr and NuttX have compelling counter-offers, but so far they still don't feel like general purpose OSes, where you can just install an OS and futz around in shell & set stuff up: they feel like a world where one needs extreme expertise in fantastically narrow small systems where none of the regular computing lessons & experiences port in. Is it sufficient resources? Yes. Is the experience worth your time? Absolutely the fuck not, in no way, unless you have incredibly vast time to burn. Just use Linux. I hope the embedded OS situation improves, and I actively follow & hack around in Zephyr & love it, but I cannot recommend hard enough not wasting your god damned time.