Storage service, access-pattern, and size economics

Storage price comparisons are valid only after the access model is fixed. Object storage exposes keys and object APIs; shared file storage exposes a hierarchical namespace and filesystem semantics; block storage presents a device to a host. A workload that needs POSIX file operations cannot select S3 solely for a lower capacity rate, and a database volume cannot replace its block device with EFS merely because several clients could mount it. Record operating systems, protocols, locking, attachment scope, persistence, and failure domain. Then quantify logical data, provisioned capacity, copies, snapshots, growth, request count, IOPS, throughput, retrieval, and transfer. EBS is AZ-scoped block storage; EFS supplies shared Linux NFS semantics; FSx families provide purpose-built file systems such as Windows SMB; S3 supplies elastic object capacity. Performance is another contract. General-purpose gp3 can decouple provisioned performance from volume size within its current envelope, whereas a demanding transactional profile may justify io2. Keep exact limits and rates freshness-sensitive and use measured latency, queue depth, throughput, and IOPS to choose.

A useful storage total-cost equation has more than GiB-month. Add requests, retrieval, minimum billable object or duration effects, provisioned IOPS and throughput, snapshots, replicas, cross-boundary transfer, data processing, backup growth, and operational labor. For provisioned storage, distinguish business bytes from allocated bytes and filesystem free space. For consumed-capacity services, include access and throughput modes rather than assuming that elasticity removes performance planning. Model the steady month and a recovery month because restores or replica rebuilds can change request, transfer, and performance demand. A low unit rate can lose when it requires excess provisioned capacity to obtain throughput, or when every small object creates disproportionate request work. Conversely, premium performance is waste when measurements remain inside a cheaper envelope. Security and resilience constrain cost optimization: encryption key access, retention locks, and required multi-AZ or cross-Region copies are not optional line items. The optimization target is the smallest valid architecture over the declared horizon, not the fewest storage products.

Instrument capacity and access before changing tiers. For block volumes, collect used filesystem space, provisioned size, read/write IOPS, throughput, queue depth, burst behavior, and the instance-side EBS ceiling. For EFS, inspect stored bytes by class, throughput demand, metadata-heavy operations, and client behavior. For FSx, add protocol-specific features, deduplication where appropriate, free capacity, and maintenance constraints. For S3, count objects, requests by operation, object-size distribution, lifecycle transitions, retrievals, and version growth. Snapshot creation does not by itself prove restore time, so test a representative restore and filesystem expansion. Delete an unattached volume only after ownership and retention are verified. A database that needs 6 TiB today might still require additional free-space headroom and snapshot growth; that headroom is not equivalent to waste. Reversal conditions should be explicit: EFS becomes preferable when many Linux clients truly require one namespace, FSx for Windows when SMB and Active Directory behavior are mandatory, and S3 when all consumers can use object semantics. These boundaries keep rightsizing from changing the application contract.