Design an Object Storage Service With Buckets, Keys and Durable Blob Storage

Design an object storage service. Clients create named buckets and, inside a bucket, store, read, list and delete objects: opaque blobs of bytes addressed by a key, with a small amount of metadata such as size, content type and a checksum. The round was given only as "design object storage", so establishing the scope and the numbers is part of the exercise. ```hint Two kinds of data An object is a large blob plus a small record describing it. Ask whether both belong in the same storage system. ``` ```hint Durability has a price Compare what full copies and erasure coding cost in disk space, in repair traffic and in read latency before you pick one. ``` ### Clarifying Questions - What is the object size distribution: mostly small files, very large files, or both? Is there a maximum object size? - What total capacity and what read and write request rates should the design target, and in how many regions? - What durability and availability are expected? - What consistency is required: must a read immediately after a successful write return the new object, and must a new object appear in listings immediately? - Which features are in scope: overwrite, versioning, range reads, multipart upload of large objects, access control? ### What a Strong Answer Covers - A minimal API (create bucket, put, get with ranges, delete, list by prefix) and the semantics of each call - Separation of a metadata service (bucket and key to object location) from the data storage nodes - Data placement and durability: replication or erasure coding, failure domains, checksums - The write path, including when a put becomes visible, and the read path - Handling of large uploads (chunking, multipart) and of many tiny objects - Failure detection, background repair, and garbage collection of deleted or orphaned data - Capacity estimates driven by the numbers agreed in clarification ### Follow-up Questions - Most objects turn out to be a few kilobytes. What changes in the data layout? - A disk holding part of a popular object fails during a spike in reads. Walk through what happens. - How do you list a bucket with billions of keys by prefix, page by page? - How would you add cross-region replication, and what consistency would readers in the second region see?