Earlier quoted context omitted.
Yes...there were a couple of generations of what we would recognize as vaguely 'modern' computers (say...roughly ENIAC to the IBM 704/709) built completely out of stuff that looked like this: http://www.righto.com/2018/01/examining-1954-ibm-mainframes-...
Yes, that was the first all-electronic generation after the very earliest relay designs. They were shockingly unreliable and incredibly expensive. Tubes have a very low mean time between failure, so any design that uses tubes exclusively can't work for more than short periods without breaking down - possibly minutes, maybe hours, probably not days, and absolutely not months or years. And each failure means a cycle of…
Tube reliability improved radically over the 15-20 years tube computers were a thing; it had too. And just like you point out about silicon, reasonable thermal management became recognized as important to tube reliability and designs changed accordingly. MTBF was lower than a modern computer, but they certainly ran for days or weeks and more. And debugging was usually fairly quick as you ran some diags that pinpointed the module (not single tube) that failed and replaced the whole thing.
I have an acquaintance with a Bendix G15 that still runs. Admittedly, the G15 is much simpler than an IBM 700, but it's a nearly 65 year old tube machine.