TL;DR: EFS Archive is the same POSIX filesystem at ~$0.008/GB-month instead of Standard's ~$0.30 — a 97% discount for files you touch once or twice a year. Files age into it automatically via lifecycle policy; when something cold is read, EFS restores it in 3–5 hours (no expedited option) and charges ~$0.03/GB. Compliance archives and never-restored "insurance" backups are the canonical wins; anything an application reads on demand must stay out.
The numbers
| EFS tier | $/GB-month | Access |
|---|---|---|
| Standard | ~$0.30 | instant, multi-AZ |
| Standard-IA | ~$0.016 | instant + ~$0.01/GB fee |
| One Zone / One Zone-IA | ~half the multi-AZ price | instant, single-AZ |
| Archive | ~$0.008 | 3–5 hours + ~$0.03/GB fee |
Worked examples: 1 TB cold = $300/month → $8 ($3,504/year saved). Field case: 15 TB of never-restored EFS backups at $4,500/month → 13.5 TB archived + 1.5 TB on IA ≈ $132/month (~$52k/year).
Do this
-
Check what's genuinely cold: per-tier storage in
describe-file-systems/ CloudWatch, and be honest — "when did anyone last read files older than six months?" -
Add the Archive step to your lifecycle policy (files reach it automatically; nobody archives by hand):
lifecycle_policy { transition_to_ia = "AFTER_30_DAYS" transition_to_archive = "AFTER_90_DAYS" }Console: EFS → file system → Lifecycle Management. Start at 90–180 days and tighten once metrics prove the data is cold.
-
Keep "transition back to Standard on first access" on for anything that might get re-read repeatedly — otherwise each cold read pays the fee again.
-
Watch two CloudWatch signals for a month: per-tier
StorageBytes(files actually migrating) and Archive-tier IO (a climbing line means your threshold is too aggressive).
Gotchas
- 3–5 hours, no expedited option. A user clicking a file served from Archive waits hours. If the runtime can't tolerate that, the file doesn't belong there — period.
- Scanners reset the clock. Nightly antivirus, search indexers, and full-filesystem backup tools count as "access" and keep everything hot. Exclude them or lengthen thresholds.
- "A few times a month" is IA's job, not Archive's — Archive fees plus latency lose to IA's instant access at that cadence.
- Archive vs deleting: for data with no plausible future use, the comparison isn't Archive vs Standard — it's Archive vs $0.
- Filesystem requirement check: if nothing actually needs POSIX/NFS semantics, S3 Glacier Deep Archive is ~8× cheaper still — the bigger win may be leaving EFS entirely.
Skip this if
- The application reads the data on a schedule or on user demand — retrieval latency is product-breaking.
- Backups you'd need within the hour in a disaster — keep DR-critical copies in IA or Standard.
- You're unsure whether data is cold — get visibility first with a conservative 180-day policy and the per-tier metrics, then tighten. The policy mechanics live in EFS Lifecycle Management.