CAP Theorem - two decades and few clouds later
summary
CAP is not a triangle label but a set of per-operation trade-offs: S3 strong read-after-write (announced 2020) gives read-after-write semantics while bucket metadata remains eventual; Spanner TrueTime combines two‑phase commit, Paxos and bounded clock uncertainty to achieve linearizability and reports ~99.999% SLA; Cassandra tunable consistency and lightweight Paxos let users pick consistency vs latency; Kafka uses producer acks and unclean leader election so availability choices can lose messages. Configuration and SLAs, not static CP/CA/AP tags, decide behavior.
AI-written summary · about
chapters (42)
- 0:00Intro
- 0:57Triangle makes CAP easy to remember
- 2:42CAP is proven
- 3:26What about the tools?
- 4:09Quick reCAP
- 6:07CAP is about the decision making
- 7:54Node failure is not a network partition!
- 9:31Linearizability (strict consistency)
- 9:48on a single copy of data (single node)
- 10:41Consistency is a spectrum of models and guarantees
- 10:51Performance & fault tolerance trade-offs
- 11:47What about ACID?
- 12:17ACID: Transactions CAP: Operations
- 12:40ACID Consistency is about the invariants
- 13:45Cloud hash-table for blobs (key-value)
- 13:57Buckets & objects
- 17:17Google Spanner
- 18:30If you're rich, the rules apply a little less to you...
- 19:052PC + 2PL + Paxos
- 23:00Differential availability
- 25:42Availability issues easy to recover from
- 26:11Unpredictable consistency is the worst nightmare
- 27:24Defer conflict resolution to the reads (last write wins)
- 28:02Tunable consistency
- 30:11Lightweight transactions (Paxos)
- 31:37Distributed append-only data log (and beyond!)
- 32:08Kafka was build for scalability
- 33:07Consistency recipe: leader & follower(s)
- 34:29Consumers polling for data
- 37:49Consistency does not guarantee durability!
- 38:04Improving availability
- 38:31Unclean leader election
- 39:14KIP-392: Follower fetching
- 40:30Config & params define CAP trade-offs
- 40:57Summary
- 41:17Consistency and availability are spectrums
- 41:54High availability is not enough for A
- 42:07Eventual consistency is a spectrum too
- 42:23Partition handling strategy matters
- 42:42SaaS: SLA beats CAP
- 43:05Availability trade-offs are the easiest
- 43:39CAP missing parts: durability, latency, ...
related talks
01h 15mCAP Theorem — two decades and few clouds later
20mCAP Theorem Reloaded
41mUnderstanding the CAP Theorem with Apache Kafka
57mScaling the CAP Theorem for Modern Distributed Systems
47mBeing transactional in 2019 - what does this actually mean?
54mDay 1 Session 1
30mApache Cassandra: What’s New and What’s Next
13mThe ACID BASE Conundrum