Software preservation · Non-profit
Clásicos Básicos
A legally-grounded archive of historic software, served to researchers from balanced bare metal with a verified storage tier behind it.
The problem
- Hard constraints
- Rights-restricted titles must never be reachable from the public tier, by any URL.
- Every stored object needs a checksum recorded at ingest and re-verified on a schedule.
- The archive copy must be writable once and not casually deletable afterwards.
- A publication day must not be able to take the catalogue down.
- A grant-funded budget, so the bill has to be legible a year in advance.
Clásicos Básicos preserves software that is old enough to be disappearing and young enough to still be under copyright. Every title in the collection carries a different legal position: some are freely redistributable, some are held under an explicit rights agreement, and some may only be consulted on site by a named researcher.
That makes the archive an access-control problem as much as a storage problem. It also makes integrity non-negotiable: a preservation copy that has silently rotted is worse than no copy, because the catalogue still claims it exists.
The public catalogue is quiet most of the year and then very much not quiet: publishing a newly cleared collection puts more traffic on the site in an afternoon than the previous six months combined. A single machine had already fallen over once.
Architecture
Edge
Balancer pair
2 ×Super VPS Boost
Application
Catalogue nodes
2 ×Forge 4
Storage
Object store
1 ×Apex 3
Archive
Write-once copy
1 ×Apex 1
What it runs on
Edge
Balancer pair
4 vCore · 8 GB · 75 GB NVME
Two small virtual servers running the load balancer with a floating address between them. Balancing is not a product we sell — it is software the team runs, on machines sized for connection handling rather than compute.
From $18.99 per month
Application
Catalogue nodes
2 ×Forge 4
Model varies — clock and cores guaranteed · 32 GB DDR4 · 2× 2 TB NVMe
Two identical application nodes. The 4.5 GHz clock suits the mostly single-threaded catalogue search, and four terabytes of mirrored NVMe keeps thumbnail and metadata reads off the archive tier entirely.
From $93.00 per month
Storage
Object store
1 ×Apex 3
AMD EPYC 4245P · 128 GB · 2× 960 GB NVMe + 2× 15.36 TB NVMe
Thirty-two terabytes of NVMe holding the working set, with no spinning disk anywhere in the path — integrity sweeps run across the whole collection without slowing reader traffic to a crawl.
From $250.00 per month, $235.00 one-time setup
Archive
Write-once copy
1 ×Apex 1
Intel Xeon D-1521 · 16 GB · 6× 8 TB SATA
A separate machine, deliberately modest in everything but capacity: four cores and forty-eight terabytes, holding the append-only copy. Its job is to be boring and to still be there in five years.
From $130.00 per month, $126.00 one-time setup
Techniques worth naming
Content-addressed storage
Objects are stored under the SHA-256 of their own bytes, so the address is the checksum. A corrupted file cannot masquerade as a good one, and the same disk image ingested twice from two donors occupies one copy and records two provenances.
Scheduled integrity sweeps
A rolling job re-reads and re-hashes the entire collection on a fixed cycle, comparing against the ingest manifest. Bit rot is found by the archive, not by a researcher who downloaded a broken file.
Append-only retention
The cold copy accepts writes and refuses deletes. Removing anything is a deliberate, logged, two-person operation rather than something a mistyped command can do at three in the morning.
Rights-aware routing
The public tier holds no restricted bytes at all. Restricted titles live behind an authenticated path on a separate origin, so a misconfigured cache rule on the public side cannot leak one — the object is not there to leak.
Health-gated balancing
The balancer checks a real endpoint that touches the database, not a static file. A node that is up but cannot answer a query is removed from rotation instead of politely serving errors.
Dual-stack from the start
Every machine carries IPv6 alongside its IPv4 allocation, and internal traffic between tiers runs over it. University networks reaching the archive are frequently v6-first.
What changed
- A publication day is now absorbed by two catalogue nodes instead of taking one machine offline, and a node can be rebooted mid-afternoon without a maintenance window.
- Every object in the collection has a recorded checksum and a last-verified date, which is the evidence a preservation programme is expected to be able to produce.
- Restricted material is separated by architecture rather than by configuration, which is the difference between a policy and a guarantee.
- The monthly cost is four fixed line items, which is what a grant application needs it to be.
No performance percentages appear on this page. We publish figures we have measured under a named method, and these deployments are described by their architecture instead.
Ready when you are
Bring us a workload, not a shopping list
Tell us what the thing actually does and we will tell you which machines fit — including when the answer is a cheaper one than you asked for.
Virtual from $14.99/moBare metal from $24.00/mo