The cluster is set up on zk{1,2,3}.eqiad.wmflabs
Description
Description
| Status | Subtype | Assigned | Task | ||
|---|---|---|---|---|---|
| Resolved | LSobanski | T95662 Implement a configuration discovery system | |||
| Resolved | Joe | T97029 integrate (pybal|varnish)->varnish backend config/state with etcd or similar | |||
| Resolved | Joe | T95656 Choose a consistent, distributed k/v storage for configuration management/discovery | |||
| Resolved | Joe | T96839 zookeeper evaluation |
Event Timeline
Comment Actions
As dynamic reconfiguration is only supported in ZK 3.5, we need to use that, and jessie only has 3.4.5 at the moment.
Comment Actions
Zookeeper is a solid, proven (https://aphyr.com/posts/291-call-me-maybe-zookeeper) distributed k-v with solid performance and reliability.
Pros:
- It's backed by a large community and different companies contribute to its development
- It's a de-facto off-the-shelf standard choice for this task
- It has quality libraries in most languages
Cons:
- it's Java
- Imagine that the preceding point has been repeated 5 times
- It uses a purpose-built protocol, so interacting with it via telnet/curl is not possible
- It shows its age, a few things are clearly better organized in more modern systems
- If libraries aren't high-level enough (like curator or kazoo) the developer has to deal with all the little nuances and the disconnection/error/retry logic, or get in serious trouble
- No multi-Dc support built in, and even using the observer nodes switching over from one datacenter is going to be painful (and ZK lacks a backup/recovery tool)
So if we want a *really* boring technology, we should pick ZK.