Disk IO: the hidden spec nobody checks

My PostgreSQL instance on a "SSD" Contabo VPS started grinding last month. Queries that took 50ms jumped to 800ms. I assumed query plan regression, spent two days optimizing.

Then I ran fio.

• Sequential read: 450 MB/s — "fine"
• Sequential write: 380 MB/s — "fine"
• 4K random read QD1: 4.2 MB/s — not fine
• 4K random write QD1: 1.8 MB/s — catastrophic

The "SSD" was a Samsung PM863a SATA behind an LSI 9211-8I in IR mode. No cache. 4K random latency spiking to 120ms under concurrent load. YMMV, but IMO this is why "SSD VPS" marketing without IOPS specs is meaningless. Take it with a grain of salt, but check your actual hardware before blaming your database.

I have since moved to OVHcloud's NVMe tier. Same price bracket, 4K random QD1: 85 MB/s. Night and day.

What are others seeing? Please include fio params so we compare apples to apples.

Nossa, my "SSD" from Hostinger was worse than HDD! 4K random read 0.9 MB/s, caramba! I thought it was my application, spent one week optimizing queries that were not the problem. Double negation in their marketing: "not slow" — well, it is not not slow, it is terrible! Warm recommendation to avoid.

Did you test your restore?

Because here's the thing: fast disk with no backup is a disaster waiting for Tuesday. Before you chase NVMe, check your 3-2-1:

• 3 copies of data
• 2 different media types
• 1 offsite

I ran fio on my "SSD" KnownHost. Results were mediocre but acceptable. Then I tested restoring a 20GB dump. Took 4 hours. The restore is where SATA RAID cards hurt — sustained mixed IO, not the clean patterns fio tests. My checklist now includes restore timing, not just benchmark numbers.