A municipal open-data warehouse bursts temporary scratch to faster media during hyperparameter sweeps. Which storage practice is this?

A municipal open-data warehouse that bursts temporary scratch to faster media during hyperparameter sweeps is practicing performance tier selection: intensive, short-lived scratch rides a higher-performance tier for the sweep, then can cool back to cheaper media. That keeps sweep iteration fast without permanently paying NVMe rates for all temporary data. Tier selection is how civic AI platforms align media cost with workload intensity over time.

Pinning aged sweep archives to higher-performance tier for long-term retention runs fails by concept—it strands cold data on premium media long-term rather than bursting active scratch to faster media during the sweep. Replicating burst scratch on higher-performance tier across zones for standby recovery fails the stem—it provides cross-zone standby copies rather than selecting a higher-performance tier for burst scratch during intensive sweep runs. Enabling inline deduplication on burst scratch to extend higher-performance tier capacity fails the burst performance need the stem describes—deduplication stretches capacity but does not select a higher-performance tier for burst scratch during intensive sweep runs.

Exam caveat: Domain 2.3 storage performance evaluation includes using faster tiers when the AI phase demands it. Burst tiering is not a license to leave sensitive open-data scratch unprotected; security and cleanup still apply after the sweep.

Operational check: automate promotion of sweep scratch to the fast tier with quotas and TTLs; measure trial step time versus tier; reclaim or demote scratch after sweeps so the open-data warehouse does not permanently strand capacity on premium media.