Building a Performant Configuration-Store Layer
Talkat FoundationDB Summit 2019 · on Monday, 2019-11-18· at 14:30· Grand Ballroom 8, Marriott Marquis San Diego Marina·
summary
Uses a deterministic prefix derived from the transaction read-version masked+bit-reverse buckets to spread appends across storage servers and shows production graphs where hotspots shift every few minutes instead of saturating two servers. Preserves ordering by adding a conflict-range previous-bucket dirtying invariant that forces out-of-order commits to retry. Cuts median transaction latency ~25% by caching and reusing read/commit versions to skip GRV when proxy latency is low. Handles multi‑MB to multi‑GB values with a chunked-ownership protocol using a GC row and atomic handoff.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 1:33Change-Logs - Hot Shards Use-cases: Indexing, cache_sync etc.
- 3:10Change-Logs: Unbalanced Mutations
- 4:17Change-Logs: Row Structure Approach • Switch SS faster, small bursts are okay • Unique, non-contiguous prefix (buckets)
- 6:13Change-Logs: Ordering Issue
- 9:22Minimizing Latency Latency sources
- 13:54Handling Large Values: Add()
- 16:36Recap
related talks
35mFast on paper, slow in reality: what we got wrong about performance
35mSystems @Scale Summer Q&A Week 3
41mTransactions: myths, surprises and opportunities
24mStronger Semantics for Low-Latency Geo-Replicated Storage
31mReimagining the Future: Where the Market is Evolving
18mAvoiding Conflicts in Highly-Concurrent Applications
59mDistributed Abstractions
01h 12mScaling distributed ledgers: Achieving arbitrary throughput and sub-second latency