S3 Glacier Flexible Retrieval

TL;DR: The classic compliance-archive tier, around since 2012: ~$0.0036/GB-month (roughly $4/TB), 82% below Standard, with a choice of retrieval speed per request — expedited (1–5 min), standard (3–5 h), or bulk (5–12 h, nearly free). For backups, closed projects, and audit archives where you can plan a few hours ahead, the savings are enormous.

The numbers

• Storage: ~$4/TB-month vs ~$23/TB on Standard
• Retrieval options: Expedited 1–5 min ($0.03/GB + $10/1,000 requests) · Standard 3–5 h ($0.01/GB) · Bulk 5–12 h (~$0.0025/GB)
• Minimums: 90-day storage duration; 40 KB per object
• Durability: 11 nines, 3+ AZs — same as Standard

Worked examples: 1 PB of genomics archives = $4,000/month vs $23,000 on Standard ($220k+/year saved). 10 TB of logs = $41/month vs $235; even the year you restore a full TB for an audit ($100 retrieval + egress), you're still far ahead.

Do this

1. Transition by age with a lifecycle rule — nobody uploads to Glacier manually:

   aws s3api put-bucket-lifecycle-configuration --bucket YOUR-BUCKET \
     --lifecycle-configuration '{"Rules":[{"ID":"archive-old","Status":"Enabled",
     "Filter":{"Prefix":"backups/"},
     "Transitions":[{"Days":90,"StorageClass":"GLACIER"}]}]}'
   

2. Restores are jobs, not downloads. restore-object stages a temporary copy (default ~1 day) once the job finishes; wire up SNS/EventBridge notifications so nobody polls, and download within the window.

3. Use Bulk for anything big and plannable — at ~$0.0025/GB it's nearly free; audit pulls submitted in the morning are ready by end of day.

4. Mass restores → S3 Batch Operations. Thousands of objects for a migration or discovery request should be one batch job, not a script of loops.

5. Test the restore workflow before you need it — the worst time to learn the two-step dance is during an incident.

Gotchas

• Expedited isn't guaranteed without provisioned capacity (~$100/month per unit); under load, expedited requests can be throttled. Don't build a DR plan on it casually.
• 90-day minimum + 40 KB minimum: churny or tiny-object data gets billed more than it saves. Aggregate small files before archiving.
• The restored copy expires (you set the window; default is short). Need it twice? Pay twice — or copy it back to Standard while staged.
• "Semi-frequent" access kills the math. Monthly retrievals mean retrieval fees plus friction; that data belongs in Glacier Instant Retrieval or Intelligent-Tiering.
• Transitions run in a daily batch — don't expect objects to move minutes after adding a rule.

Skip this if

• You need millisecond access — Glacier Instant Retrieval.
• You'll realistically never read it — Deep Archive is ~4× cheaper still.
• The data churns (rewritten or deleted inside 90 days) — Standard is cheaper despite the price tag.
• Sub-hour recovery objectives: only expedited gets close, and it's the priciest, least-guaranteed option.