Filesystems · Cluster
Distributed Storage
Erasure coding, replication, and placement.
- 01 Failure Domains and the Arithmetic Behind Replication Factor How a copy count becomes a durability estimate, why that estimate collapses when failures correlate, and what placement must do to keep the arithmetic honest.
- 02 Consistent Hashing and What Changes When a Node Joins Why modulo placement reshuffles almost every key on a membership change, how a hash ring limits the movement, and where the ring still concentrates load.
- 03 Reed-Solomon Codes: The k-of-n Arithmetic and Its Repair Cost How a Reed-Solomon code turns k data fragments into n, why any k reconstruct the whole object, and what the encoding matrix demands of the arithmetic.
- 04 Quorum Intersection and Why R + W Must Exceed N The arithmetic that guarantees a read sees the latest write, how sloppy quorums break it, and what read repair has to do to restore the guarantee.
- 05 Merkle Trees, Range Digests, and Finding Divergence Without Reading Everything How a tree of hashes narrows a comparison between two replicas to the differing leaves, what it costs to maintain under writes, and why the leaves must align.
- 06 CRUSH Placement: Deterministic Mapping Over a Weighted Hierarchy How a pseudorandom placement function turns an object identifier and a cluster map into a device list, why straw selection replaced simpler bucket types, and what changes on topology edits.
- 07 Repair Traffic in Erasure-Coded Sets and What Regenerating Codes Change Why rebuilding one fragment costs k fragments of reads, how the repair bandwidth bound is derived, and what local parity and regenerating codes buy against it.
- 08 The Degraded Window: Rebuild Rate, Exposure, and the Second Failure How the time between a device loss and the completion of its rebuild becomes a durability multiplier, and why the repair rate cannot be set independently of it.
- 09 Chain Replication: Why the Head Writes and the Tail Reads How ordering replicas into a chain turns replication into a single ordered path, why reads are served at the tail, and what the head and tail failures require of the membership protocol.
- 10 Copysets and the Tail of Random Placement Why uniformly random replica placement produces objects whose copies share a failure set, how many nodes one rack loss can take with it, and what bounded placement changes.