tinaworm
Member
OP
- Joined:
- Jun 2024
- Posts:
- 84
- From:
- Riga, Latvia
You have provisioned a server. You have installed services. You have not tested failure modes. This is a mistake.
- Test disk I/O with fio, not dd. Sequential writes mislead. Random 4k matters.
- Test memory with stress-ng --vm 2 --vm-bytes 80%. Observe OOM behavior.
- Test network with iperf3 to multiple regions. Single-point tests deceive.
- Test CPU stability with mprime for 24 hours. Thermal throttling hides in short runs.
- Test your provider's reaction to abuse reports. Send one yourself. Measure response time.
Pro tip: schedule tests before production data exists. Recovery from corrupted databases is not a learning experience.
- Test your backup restoration. A backup you cannot restore is not a backup.
Document everything. Compare against specifications. Accept nothing without verification.
-t
worst bandwidth, best stories
olya
Member
Always Free
- Joined:
- Jun 2024
- Posts:
- 159
- From:
- Warsaw, PL
Solid foundation. Here is my expansion for Oracle free tier specifically, though most applies elsewhere.
1. Check CPU steal with `top` then `1`. Ampere A1: expect 0.0% steal. X86 "always free": expect 10-30% during US daytime. Heads up: this is not throttling, this is oversubscription.
2. Disk test: `fio --name=randread --ioengine=libaio --iodepth=32 --rw=randread --bs=4k --direct=1 --size=4G --numjobs=4 --runtime=300 --group_reporting`
Pro tip: Oracle's block volumes have burst credits. First 10 minutes look amazing. Run 30 minutes for truth.
3. Network: `iperf3 -c iperf.he.net -R` from multiple instances simultaneously. I have seen 500 Mbps become 50 Mbps under cross-AZ load.
4. Hidden throttling check: run `stress-ng --cpu 4 --timeout 3600` and watch `cpufreq-info`. If frequencies drop after 20 minutes, you have thermal or power limits not in the spec sheet.
My A1 instance held 3.0 GHz for 45 minutes, then dropped to 2.6. Not advertised. Not illegal. Just hidden.
licensing is a suggestion
