What can go wrong with the distributed system - the Hazelcast exp...
summary
Hazelcast is an in‑memory Java platform with RPC, replication and a Raft-based CP layer, and a demo runs face‑detection streaming that stores counters in a distributed map. A concrete bug: using IMap eventual consistency for job snapshots lets a 3‑node cluster lose state when node‑2 drops a replication message and node‑1 fails. Fixes discussed: read/write majority, Paxos or true linearizability via Raft; other issues include outbound queue 2 GB message bloat and micro-raft CP primitives for consistent locks.
AI-written summary · about
chapters (16)
- 0:00Introduction and speaker
- 1:13What is Hazelcast
- 2:40Real-time stream processing demo
- 4:50Replication problems in distributed systems
- 8:09Eventual consistency and CAP theorem
- 10:41Strong consistency approaches
- 13:15AP and CP subsystems in Hazelcast
- 14:21Network issues in distributed systems
- 16:11Message ordering and duplicates
- 17:58Cluster join problems
- 20:54Timestamp and logical clocks
- 23:49Outbound message queue memory issues
- 25:38Client issues with distributed locks
- 28:54Pessimistic locking in Python
- 33:10Key takeaways and resources
- 34:36Q&A session
related talks
46mWhat can go wrong in a distributed system - experience from...
30mDistributed Computing with Hazelcast
59mDistributed applications using Hazelcast
45mjava.util.concurrent for Distributed Coordination
55mHazelcast - Hitchhiker’s Guide. Grzegorz Piwowarek. Java Fest 2020 Online
47mjava.util.concurrent for Distributed Coordination
44mJava: Introduction to the Hazelcast Ecosystem
48mHow to build distributed concurrency primitives