Question about Sentinels with network partitioned

[HeartSaVioR]

Hello!

I have been collaborated with Jedis, and I'm interested in Sentinel.

While commenting to issue, I'm suddenly wondered about how we retrieve master with Sentinels with network partitioned.

(https://github.com/xetorthio/jedis/pull/557#issuecomment-35495815)

My scenario starts at 3 Sentinels (A, B, C) and Redis Server (M, S).

A. Sentinels A, B, C monitors Redis Server with M/S replication, and also Client monitors 3 Sentinels by subscribing "+switch-master".
B. Suddenly network partition occurred -> Sentinels B, C + Redis M/S, Client / Sentinel A.
Sentinel A suddenly can't connect.
C. Redis Master Server suddenly shut down. B, C Sentinels agree to failover and changes their master. 
(Sentinel A cannot know that)
D. Client changes master based on "+switch-master" message.
E. Network partition is resolved, so Sentinel A is now reachable.

F. Client talks to Sentinel A who is master.

I have two questions about this scenario.

1. When F occurred, is there no probability to respond M is master? Is there no timing issue?

2. Can Client always get current master to any Sentinel correctly? (IDONTKNOW is surely OK.)

Or is there a information about current Sentinel's epoch?

Actually scenario could be applied to Sentinel A suddenly shut down and revive.

Please leave an answer so that I can think deeply on Sentinel support.

Thanks!

[HeartSaVioR]

I found curious PR from github redis repo.

https://github.com/antirez/redis/pull/821

Does it make sense?

Thanks!

Sincerely.

Jungtaek Lim (HeartSaVioR)

2014년 2월 20일 목요일 오전 10시 44분 8초 UTC+9, HeartSaVioR 님의 말: