Antidote - a Scalable and Consistent Transactional Data Store
summary
Antidote is a working research database: a distributed key-value store with CRDTs, atomic multi-key writes, snapshot reads and data partitioning, implemented in Erlang implementation as part of the SyncFree project. It claims Causal consistency as the strongest available guarantee under partitioning, implements Clock-SI and a geo-replicated variant (Cure) with per-DC vector clocks, and achieves about 100k ops/s in a 3‑DC × 25‑node benchmark on Grid5000. Prototype limitations: default full replication and ongoing work on selective replication and search indexing.
AI-written summary · about
chapters (26)
- 0:00Intro
- 0:47vision
- 2:30The Role of the Datastore
- 4:08Antidote in a nutshell
- 6:12Why Erlang?
- 7:56Enabling technologies
- 9:03CRDTs in Antidote
- 10:15Consistency at different levels
- 11:39Highly Available Transactions (HATS)
- 13:31Guarantees
- 16:01Interactive Transaction API
- 17:17Static Transaction API
- 19:31Not causal
- 21:01Causal consistency
- 26:43Cure
- 27:26Causal+ Consistency Protocols
- 28:06Related Work: implementation
- 29:11Outlook: Partial replication
- 31:20Architecture
- 31:45(Physical) Nodes
- 33:28Preliminary benchmarks
- 35:08Evaluation: Scalability
- 37:51Protocol buffer interface
- 38:48Things on our agenda
- 41:33The Antidote Team
- 42:09Open positions
related talks
44mAntidoteDB: highly available, transactional database
35mData Types for Eventually Consistent Databases
57mLiving in eventually consistent reality
50mCRDTs in Practice
17mJust-Right Consistency
16mScaleDB: A Scalable, Asynchronous In-Memory Database
22mHappily Ever After: A CRDT Fairy Tale
52mTransactions: Myths, Surprises and Opportunities