Okay, I know this is going to get some strong reactions, but after running workloads on all three majors for the last two years, I’ve had to switch my default.
For critical stateful applications, I’m now consistently choosing Google Cloud’s persistent disks over AWS EBS. The reliability difference, especially around multi-attach and consistent performance, is just too stark.
My turning point was a series of automated failover tests for a high-availability Postgres setup. On EBS gp3 volumes, we'd occasionally see latency spikes during the attachment process after a simulated zone failure. Not catastrophic, but enough to cause a few seconds of application timeouts. The same test pattern on GCP's SSD persistent disks? The transition was consistently smooth, with no observable performance hit.
I’ve been digging into why, and a few things stand out:
* **Multi-attach behavior:** GCP’s multi-writer option for regional persistent disks feels more mature. The read/write coordination seems to handle contention better in my stress tests.
* **Performance consistency:** The baseline performance for their SSD tiers appears more predictable under variable I/O patterns. I see fewer "noisy neighbor" type dips.
* **Snapshot/Clone speed:** Creating a new disk from a snapshot is noticeably faster in my experience, which speeds up recovery and staging environment spins.
I’m curious if others have done similar head-to-head comparisons. Are my observations an outlier? Has anyone run long-term metrics on failure rates or performance variance between the two?
For reference, my primary use cases are databases (Postgres, MySQL) and write-heavy analytics pipelines. The cost is roughly comparable, so for me, it’s coming down to operational stability.
Your point about latency spikes during EBS attachment is interesting. I've observed similar behavior, but I'd trace it more to the specific volume type and instance configuration than the platform's inherent reliability.
In my benchmarks, EBS io2 Block Express volumes with multi-attach enabled show attachment times and failover latency comparable to GCP's regional disks, but at a significantly higher cost. The gp3 performance variability you mention is documented; it's a trade-off for the lower price point. GCP's consistency likely comes from their different provisioning architecture, where performance is more directly tied to allocated capacity rather than being burst-based.
The real question is whether that consistency holds under sustained, maxed-out IOPS for stateful services like Kafka or large OLTP databases. My own tests showed GCP disks could hit a performance ceiling more abruptly during sustained write-heavy loads, whereas io2 volumes offered more headroom. So it's less about universal reliability and more about which failure mode your application tolerates worse: variable latency during attachment or a stricter throughput limit under peak load.
Test it yourself.