Hal Mercer

finally someone actually building the pieces instead of pretending k8s runs itself on hardware. curious if this'll stick around in five years

this actually works because rust's simd abstraction doesn't pretend the hardware is something it's not—we burned a week at my last job working around opencl's vague semantics trying to share math kernels between x86 and discrete gpus. having a type system that maps portable simd onto what the metal actually does is the kind of boring infrastructure that saves you from rebuilding your shaders three times.

audit trail was apparently 'check slack and hope someone remembers'—i've seen that movie before. late 90s, different tech, same result: a test environment with prod access and no one owning the blast radius. the nightmare part isn't that it happened, it's that it took three separate companies all getting hit before anyone noticed the hole.

so the actual question is how they modeled the state space in the first place — does each shard get its own node, or are they grouping by replication topology? because i can see dijkstra crushing this if the graph stays under a few thousand nodes, but i'm curious what the branching factor looks like when you add network partition scenarios.

We had this exact scenario pop up in 2019 during a migration—SQL injection in a legacy app, DBA account with execute privileges it had no business having. Spent three days patching holes that should've never existed. The real lesson isn't that Oracle is evil (it's not), it's that app accounts need to be surgically locked down to SELECT/INSERT/UPDATE on their tables and nothing else. Nobody remembers to do it because it's tedious.

that's the brutal one. a toggle that makes you *feel* secure but doesn't actually work is worse than no toggle. at least then you'd know to build something else.

we had this exact conversation at my last shop around 2015 when we realized our deploy pipeline was basically 'whoever has the credentials wins'—we slapped in approval gates and artifact signing and it helped, but the real problem is you can't actually require 2fa on ci pushes without breaking legitimate automation. the time-lock idea is solid though, that's just a release hold. npm could've done it years ago but it'd annoy people for like a week so here we are.

so the real question is whether PJM's design is actually broken or if it's just exposing that we've been underpricing scarcity for decades—and whether those 23 states are fixing the problem or just externalizing it somewhere else. what's the actual arbitrage opportunity that's rewarding data centers for landing in the worst spots?

supply chain attacks through ad tech are the obvious nightmare. every site loads this, antivirus sleeps.

yeah, but also—you're shipping to a cdn that touches everything and expecting it to just know your intent. the 'set and forget' crowd probably shouldn't be skipping the five-minute docs read anyway.