Cloud Infrastructure · Cluster
Cloud Storage
Object semantics, tiers, and consistency.
- 01 The Flat Keyspace: What an Object Store Does and Does Not Provide Object storage maps opaque keys to immutable byte strings. The operations it omits, including rename, append, and multi-key atomicity, shape every design built on top of it.
- 02 What a Successful Write Guarantees: Durability and Read-After-Write Consistency A 200 response means the bytes survive disk failure, not that every reader sees them yet. How replication, quorums, and listing indexes set the visibility rules.
- 03 Versioning, Delete Markers, and What a Delete Actually Removes Enabling versioning turns overwrite into append and delete into a placeholder. The state machine has three modes, and the suspended mode is where most surprises live.
- 04 Multipart Upload: Choosing a Part Size That Survives a Retry Parts are immutable commit units. The 5 MiB floor, 10,000-part ceiling, and per-part checksums turn part size into a decision about retry cost and memory.
- 05 Storage Tiers Are a Bet on Access Frequency: Minimum Durations and Retrieval Cost Cheaper storage classes trade per-byte price for retrieval fees, minimum storage durations, and latency. The break-even arithmetic decides which tier a dataset belongs in.
- 06 Conditional Writes: Building a Lock Out of Single-Key Compare-and-Set A precondition on an object write gives one writer a win and the rest a 412. How to use that primitive for leases, exactly-once publication, and safe fan-out.
- 07 Byte-Range Reads: Partial Responses, Alignment, and Parallel Fetch A Range header turns one object into many concurrent reads. What the 206 response guarantees, why alignment matters for throughput, and where range requests break.
- 08 Request Rate Limits and Why Key Prefixes Are Partitioned A bucket scales by splitting its keyspace across index partitions. Sequential keys hammer one partition, so the naming scheme decides the achievable request rate.
- 09 Committing a Dataset to an Object Store Without a Rename Table formats publish by writing a pointer, not by moving a directory. How snapshots, manifest lists, and garbage collection work when the commit is one small object.
- 10 Cross-Region Replication and the Window Between Two Copies Replication is asynchronous by default, so the second region lags the first. What the lag does to failover, to deletes, and to any read that crosses the boundary.