OK Log: Distributed and Coördination-Free Logging
summary
A distributed logging system that forgoes global coordination and instead writes ULID with millisecond prefix‑prefixed records into segment‑per‑connection files on ingestors, then has store nodes poll and pull oldest segments, replicate to a configurable replication factor 2–3, dedupe by ULID, and compact overlapping files into time‑nonoverlapping segments. It implements three durability modes (fast/fsync/bulk), defaults segments to ~16MB or 1s, keeps replication transactions short (seconds), and tolerates node loss by re‑polling and timeout.
AI-written summary · about
chapters (34)
- 0:00Intro
- 1:54My incredible journey
- 2:27Why is it hard
- 2:50You know this
- 3:37Fundamentally hard
- 6:13It's phenomenological
- 10:38Observability
- 13:27Well-defined scope
- 14:41Actual requirements
- 15:42Products
- 17:14Why is Prometheus good?
- 18:24High-level goals
- 20:16Constraining the problem
- 21:03Query interface
- 21:29A note on scale
- 22:16Easy to operate
- 24:09Coördination-free
- 24:43In practice
- 25:20The importance of writes
- 25:54Ingestion
- 26:57Durability modes
- 27:44Component model
- 31:12Priorities
- 33:04Replication
- 35:12Transactional coördination
- 35:44Failure during gather
- 36:28Failure during replicate
- 37:12Dupes don't matter
- 37:34Dupes get compacted
- 37:50Failure during acks
- 38:12Segment files
- 39:12Goal of compaction
- 39:52Adding ingest capacity
- 41:07Adding query capacity
related talks
21mLogDevice At Scale
46mLogDevice: durable and highly available sequential distributed storage system
15mXLL: Cross-Layer Logging for Data Deduplication in Consensus-Based Storage
14mLoLKV: The Logless, Linearizable, RDMA-based Key-Value Storage System
34mDistributed systems 102
19mScalog: Seamless Reconfiguration and Total Order in a Scalable Shared Log
52mThe Computer Science behind a modern distributed data store
17mDiscard-Based Garbage Collection for Distributed Log-Structured Storage Systems in...