#node #state #fork #github #delta #own #failure

datacake-chitchat-fork

A fork of https://github.com/quickwit-oss/chitchat used in datacake. YOU SHOULD USE THE MAIN LIBRARY NOT THIS ONE IN YOUR OWN CODE.

2 releases

0.5.1 Dec 6, 2022
0.5.0 Dec 6, 2022

#41 in #failure

Download history 21/week @ 2024-11-24 26/week @ 2024-12-01 26/week @ 2024-12-08 39/week @ 2024-12-15 1/week @ 2024-12-22 4/week @ 2025-01-05 14/week @ 2025-01-12 8/week @ 2025-01-19 2/week @ 2025-01-26 18/week @ 2025-02-02 23/week @ 2025-02-09 21/week @ 2025-02-16 22/week @ 2025-02-23 16/week @ 2025-03-02 5/week @ 2025-03-09

66 downloads per month
Used in 7 crates (3 directly)

MIT license

135KB
3K SLoC

chitchat

This crate is used at the core of Quickwit for

  • cluster membership
  • failure detection
  • sharing configuration, and extra metadata values

The idea of relying on scuttlebutt reconciliation and phi-accrual detection is borrowed from Cassandra, itself borrowing it from DynamoDB.

A anti-entropy gossip algorithm called scuttlebutt is in charge of spreading a common state to all nodes.

This state is actually divided into namespaces associated to each node. Let's call them node state.

A node can only edit its own node state.

Rather than sending the entire state, the algorithm makes it possibly to only transfer updates or deltas of the state. In addition, delta can be partial in order to fit a UDP packet.

All nodes keep updating an heartbeat key, so that any node should keep receiving updates from about any live nodes.

Not receiving any update from node for a given amount of time can therefore be regarded as a sign of failure. Rather than using a hard threshold, we use phi-accrual detection to dynamically compute a threshold.

References

Dependencies

~4–13MB
~137K SLoC