EC2 Instance Store
EC2 Instance Store is physical NVMe storage directly attached to the host server running your instance, delivering the highest possible IOPS on AWS with zero network latency since there's no network hop involved. Unlike EBS, data is permanently lost whenever the instance is stopped or terminated, making it suitable only for temporary data the application can regenerate — never for anything that must survive a restart.
A video-transcoding pipeline uses i3 instances with Instance Store as high-speed scratch space for intermediate rendering files during a job, since those files are disposable the moment the job completes and the raw NVMe speed cuts transcode time meaningfully.
Speed vs Durability Trade-off
Instance Store on i3 instances can hit over 3 million IOPS, versus a ceiling of 256,000 IOPS on the fastest EBS type (io2 Block Express) — but a simple stop (not even a terminate) wipes it completely, since the instance may resume on entirely different physical hardware.
Common MistakeUsing Instance Store for database files because it's faster. It works fine until routine maintenance stops the instance and every byte of data disappears. Use EBS for anything that must survive a restart.
Frequently Asked Questions
How does instance store differ from EBS in terms of persistence and use case?
Instance store is physically bolted to the host server, so it delivers the lowest latency and highest IOPS AWS offers on EC2 with no network hop, but it's ephemeral: stopping, terminating, or the instance migrating to different hardware wipes the data instantly. EBS, by contrast, is network-attached block storage that persists independently of the instance lifecycle and can be detached and reattached elsewhere.
What's a safe production use for instance store given its data-loss risk?
Use it for scratch space: temporary buffers, local caches, swap, or shuffle data in a distributed compute job (like Spark or Hadoop worker nodes) where losing the data just means recomputing it. The common mistake is treating it like durable disk for databases or logs — a routine instance stop (not just a crash) is enough to erase everything on it, with no recovery path.