#zookeeper #shared-state

zkstate

Share a struct across multiple instances of an application using zookeeper

15 releases

0.1.23 Sep 23, 2021
0.1.22 Sep 23, 2021
0.1.19 Aug 21, 2021
0.1.13 Jul 5, 2021
0.1.9 Mar 28, 2021

#75 in Concurrency

Download history 3/week @ 2021-06-28 46/week @ 2021-07-05 14/week @ 2021-07-19 14/week @ 2021-07-26 26/week @ 2021-08-02 31/week @ 2021-08-09 67/week @ 2021-08-16 4/week @ 2021-08-23 2/week @ 2021-09-06 11/week @ 2021-09-13 73/week @ 2021-09-20 4/week @ 2021-09-27 37/week @ 2021-10-04 34/week @ 2021-10-11

89 downloads per month
Used in jotty

Apache-2.0

22KB
327 lines

zkstruct

Simple way to share a struct with zookeeper. You can subscribe to changes and handle data changes.

This is a prototype, and should not be considered stable or be used for production purposes. You're going to do that anyways, but FYI.

overview

ZkStruct is designed to provide a struct that can act as a shared state.

We're building a little agent that will monitor a list of processes on the host. The agent uses ZkStruct for its state. An external actor can come along and modify that state object (either using ZkStruct or something else) to update the list of monitored processes. Let's assume this is a Vec<String>, where each String is a glob of the process we want to match.

This external actor modifies our Vec from vec!["systemd*", "rsync*"] to vec!["systemd*", "rsync*", "firefox*"]. The library will register the change, either from a Watcher event or refresh event, and update the internal state struct. The library will do a diff on the local state, emit all changes to a channel you can subscribe to, and then update the local state.

reads

Reads using .read() are not guaranteed to be fresh, and might be out of date. There is an eventual guarantee that the data will be consistent, as reads will fail if the object has not been verified consistent within a certain time (default is 30 seconds).

Reads using .c_read() are guaranteed to be fresh, as the local object version is compared to the zookeeper version. This will, of course, be slower as the client needs to request data from zookeeper and compare it before it will be returned, but you can be sure that the data being returned is correct.

writes

There are two write guarantees:

  1. You will always update the latest version of the object. If the local version of the object does not match the version in zookeeper, the update request will fail with an error.
  2. Only one update operation can take place at a time, either locally or among clients. An exclusive write lock is used both internally, and in zookeeper to prevent any unexpected updates or data loss.

Update Stream

Any changes, either local or remote, can be consumed and acted on. All clients can subscribe to a stream of changes to the local object. This can be done using the .update_handler method to attach a closure that will be caused everytime there is change registered.

There are 4 change types:

  1. Added
  2. Modified
  3. Removed
  4. Unchanged

Each Enum will come with the names of keys updated, and serde_json::Value body that contains the values added or removed. Modified changes will return the old and new values. Unchanged results are not enqueued and therefore not returned.

shared_state.update_handler(|msg| {
    println!("got stage change message: {:?}", msg);
});

Changes are emitted before the local object is updated, but the local object can be updated before all messages in the channel can be consumed.

zookeeper structure

/<dir>                  Where the Data will be stored
/<dir>/payload          JSON Payload
/<dir>/listeners        Ephemeral records with the IDs of every listener active

Dependencies

~3–4.5MB
~85K SLoC