Episode 2. S3. Day 2. Storage looks simple until it is not.

Key facts. S3 is object storage with eleven nines of durability, replicated across at least three Availability Zones. Storage classes: Standard, Intelligent Tiering, Standard IA, One Zone IA, Glacier Instant, Glacier Flexible, Glacier Deep Archive. Versioning, lifecycle rules, and replication move data between them.

Consistency is where Principal level answers live. S3 is strongly consistent for reads after writes and list operations. Say: S3 gives strong read after write consistency, so a missing object is never eventual consistency. Look for permissions, replication lag, or lifecycle expiry instead.

Scenario 1. Our objects are disappearing from the bucket. Nobody deleted them. Help.

Understand. Ask: versioned bucket or not? Lifecycle rules enabled? Replication or cross account writes? Exact object keys and timestamps. One prefix or the whole bucket?

Prepare. Say the data in order. Versioning status and delete markers. Lifecycle rule configuration. CloudTrail S3 data events for DeleteObject calls. Bucket policy and IAM changes. Replication metrics if cross region.

Say your translation line: Same instinct as RMAN retention and archivelog expiry: somebody or something aged the data out.

Identify. Rank the hypotheses. One: a lifecycle rule expiring current or noncurrent versions. Two: an app or script deleting with the wrong prefix. Three: a bucket policy or ownership change hiding objects. Four: replication failing on the destination side. Fix: suspend the bad rule, restore from versions, add MFA Delete on critical prefixes.

Support. Success: object count stable for seven days, a CloudWatch alarm on DeleteObject volume, and versioning plus Object Lock on the protected prefixes.

Scenario 2: 403 Access Denied on GetObject. Check in order: IAM policy, bucket policy, SCPs, KMS key grants, then Block Public Access and Object Ownership. Say: I'd check IAM, then bucket policy, then KMS grants, before touching the network. Most 403s are identity or key policy, not connectivity.

Slow uploads: check multipart thresholds, Transfer Acceleration, VPC endpoint versus NAT path, and client side parallelism.

Your Day 2 lab: create a bucket with versioning, add a lifecycle rule, upload and delete an object, then restore it from versions. Enable CloudTrail data events and find your own DeleteObject call.

Tomorrow, Episode 3: RDS and Aurora. Slow queries, failovers, and replica lag. Pause and say the four steps. Understand. Prepare. Identify. Support.