Archived Storage

1. The Data Classifier scans primary datasets and tags objects by access frequency, age, and retention obligation, producing a policy-driven inventory of archive candidates.
2. The Lifecycle Engine evaluates each object against transition rules (e.g., no reads in 90 days, older than the fiscal close) and marks eligible data for migration.
3. The Tiering Service applies the transition. In cloud object storage nothing is copied or moved: the object stays in the same bucket under the same key and only its storage-class attribute changes, so existing references remain valid — reads simply fail until the object is restored. (On-premises HSM and file-tiering products differ here: they do relocate the data and leave a stub behind.)
4. The Metadata Catalog retains searchable metadata—keys, timestamps, checksums, retention locks—so archived objects remain discoverable without being rehydrated.
5. The Retrieval Broker handles on-demand restore requests, choosing bulk, standard, or expedited retrieval based on the caller's urgency and the cost budget.
6. The Consumer is notified when the restore job completes, then reads the temporary copy — which lives at the same key and expires after the requested number of days, leaving the archived object untouched throughout. Automatic tiering classes behave differently: there a restore genuinely promotes the object back to a frequent-access tier, and it is re-demoted after a further period without access.

• Ultra-low storage cost vs retrieval latency of minutes to hours on the deepest tiers — or, on millisecond-access archive classes, vs higher per-GB access fees and a longer minimum-duration commitment instead.
• Automated, hands-off tiering vs the complexity of tuning lifecycle rules and avoiding premature transitions.
• Durable long-term retention vs per-request retrieval fees and early-deletion penalties.
• Reduced primary footprint vs the need to keep a rich metadata catalog for discoverability.

A national hospital network retains 18 years of MRI and CT imaging to satisfy medical-record regulations. Active scans stay on hot object storage for 90 days for clinician access; after that, an automated lifecycle policy transitions each study into Glacier Deep Archive while its DICOM metadata remains in a searchable catalog. When a physician or auditor requests an old study, a retrieval broker issues a standard restore, delivers the images within hours, and re-archives them afterward—cutting long-term imaging storage cost by roughly 80% without breaching retention mandates.

• Minimum-duration commitments: Archive tiers impose a minimum storage duration (commonly 90–365 days depending on provider and class); objects deleted or transitioned earlier still incur the full committed charge, so avoid re-tiering data that may be recalled soon after archival.
• Retrieval cost is the hidden bill: Per-GB retrieval fees plus per-request charges dominate cost when access spikes. Batch restores into bulk tiers, coalesce many small objects, and cap concurrent expedited requests against a budget to prevent runaway spend.
• Small-object overhead: Per-object metadata, minimum billable object sizes, and per-request restore fees make millions of tiny files expensive. Pack cold data into larger aggregated archives before transition and keep the manifest in the catalog.
• Catalog is the single point of discoverability: If the metadata catalog is lost, archived objects become effectively unfindable. Back it up independently, version its schema, and reconcile checksums against the archive tier periodically.
• Restore failure and expiry semantics: Rehydration is asynchronous and can fail or exceed its SLA; make restore requests idempotent, subscribe to completion events rather than polling, and handle the temporary copy expiring mid-read so consumers re-request rather than 404. Note you pay for the archived object and the restored copy concurrently, so long restore windows carry a real cost.
• Observability: Instrument transition and restore latency, retrieval bytes and request counts per policy, early-deletion penalties incurred, and catalog-to-object reconciliation drift to catch misconfigured lifecycle rules early.