Better handling of master failover

[laurivosandi]

It looks like most applications do have Redis reconnect support if Redis operation results in error such as ReplyError: READONLY You can't write against a read only replica. In fact the ioredis docs specifically mention this usecase.

This is related to #149 (comment)

To better support dumb clients I would advise implementing TCP disconnect when Dragonfly pod loses master role to ensure clients reconnect to the new Dragonfly master pod

[Pothulapati]

Hmm, This makes sense but

To better support dumb clients I would advise implementing TCP disconnect when Dragonfly pod loses master role to ensure clients reconnect to the new Dragonfly master pod

How can we do this? Considering we don't have a proxy to do different replies. It's upto the older pod to return READONLY (which they will once they are configured as a replica). wdyt? 🤔

[laurivosandi]

I think it would be fine to issue TCP connection reset for every open connection if pod loses write master status. You can make this behaviour optional via command line flag or something like that