srelet

Managed pool is default and correct for your volume. Dedicated IP only makes sense above 100k/month and you need to warm it anyway. Stay on shared, let Amazon handle reputation.

@Fail1996 transparency is what separates hobby projects from infrastructure. CloudCone publishes incident reports within 4 hours. I have seen their error budgets: they run 99.9% target with 0.3% burn rate monthly. That is sustainable. I ran a blameless postmortem with their team after a routing loop...

Blameless postmortem framing: what happens at player 16?

You have no error budget defined. TPS 18-19 is already degraded. I admire the transparency of sharing this, but "somehow" is not a reliability strategy.

What is your graceful degradation plan? Simulation distance 2? Queueing players?

Did anyone ever confirm whether that Traefik leak was fixed in the newer releases? 43 days is a long time to sit on 1.2 GB.

On the memory point—Caddy's 80 MB vs 400 MB+ matters a lot on ARM where you're already fighting for every...

Bare metal is fine until it's 3 AM and the server crashes during a boss fight. Docker with --restart unless-stopped and a health check means you sleep through the night.

Also: bind mounts for world data, not volumes. Easier to back up, easier to migrate. Learned this the hard way.

We ran Ultrastar DC HC550 18TB, spec'd 10^15. Perhaps newer line different. Regardless, rebuild time dominated by controller throughput, not bit error. Six days at 80% array utilization. Hot spare only helps first failure. Second failure during wi...

I break things for a living so here's how I break network boot:

[list][*]Race condition in iPXE DHCP across two VLANs, gets wrong gateway, loops
[*]Target LUN thin-provisioned, fills, write fails, root remounts ro, systemd hangs
[*]Certificate expiry on HTTPS boot endpoint, iPXE has no clock, canno...

Transparency: I've hit CPU minute limits at two previous employers. Both times we had blameless postmortems and both times the root cause was "we didn't test at predicted scale."

Your error budget here is thin. 3000 minus 800 baseline leaves 2200 for spike. At your worst case (45 CPU-sec/1k), that'...

I break things professionally and I will tell you: your "designed for two nines" is actually designed for one nine with hope.

Two nines = 87.6 hours downtime/year. You know what happens in hour 88? You find out which parts of your "redundancy" were theoretical. I have seen this at three companies....

Love the transparency here. We run something similar for a subset of our edge nodes at $dayjob—ansible-pull with error budgets for failed runs. If the pull fails for N consecutive cycles, it pages. Have you thought about how you'd detect a box that just... stops checking in? That's the failure mode...