Storage or Server: Where Should Capacity Live?

Every infrastructure eventually asks where its terabytes should physically live, and the three honest answers — inside the compute, attached beside it, or on the network apart from it — produce three different systems wearing the same capacity number. The right choice follows from failure domains and growth shape more than from price, though refurbished pricing softens every option.

Inside: the storage-heavy server

An R740xd or DL380 12LFF puts the pool in the machine that uses it: shortest path, no fabric, one box to buy and power. This is the converged answer — hypervisor plus ZFS, backup software plus its repository, database plus its volumes — and for single-node estates it is usually correct: fewest moving parts, best latency, simplest mental model. Its limits are its virtues inverted: the pool shares the server's fate (maintenance, failure, upgrade cycles all touch the data), capacity caps at the chassis's bays, and a second compute node cannot natively share what lives inside the first.

Beside: shelves extend the same answer

JBOD enclosures push the internal model past its bay limit — twelve more LFF drives per shelf, same server, same software, marginal cost per bay the lowest available. The failure domain stays unified (server dies, pool pauses), which remains fine wherever the internal answer was fine. Shelves are how a storage server grows without becoming a different architecture.

Apart: NAS and SAN as their own citizens

Dedicated storage — a TrueNAS box serving NFS/SMB, an MSA serving LUNs — decouples data from any one computer's fate. Multiple hosts share it (the requirement that makes clusters, live migration and HA real), compute nodes become stateless and disposable, and storage maintenance stops touching application uptime (doubly so with dual-controller arrays). The costs are the network in the data path (10GbE minimum for serious use), one more system to administer, and the mild latency tax distance always charges. The moment two or more compute nodes need the same data, this stops being an option and becomes the design.

The framework, compressed

One node, data used where it lives: storage-heavy server; add shelves when bays run out. Multiple nodes, shared data, VM estates: separate storage — NAS for files and general virtualization, SAN/block where multipath availability is the point. Mixed reality (most shops): hot tier inside the compute (NVMe mirrors where latency pays), bulk tier apart (shared, protected, grown independently) — each byte living where its access pattern votes. Refurbished economics make the hybrid genuinely cheap: the R740xd, the shelf and the MSA are all lunch-money line items now, so the architecture can follow the workload instead of the budget.