allow read from replicas in pipeline

[duke-cliff]

Can you merge this and make a 2.1.1 version first?
Using mget is not acceptable as it's adding too much cost to us(3x as the expected cluster size).

[Grokzen]

@duke-cliff Yes i can make a 2.1.1 releaes after i merge your two PR:s

[Grokzen]

Also even if all the current tests actually works, do you know know that the current pipeline algorithm actually works out in basically all pipeline cases even with read-only mode? I have no tests for this and this PR do not include any extension to the tests or at least any manual steps or a validation script to validate that it actually works out as we expect and that it handles a lot of the error scenarios and failover scenarios that actually can happen that the algorithm have to take care of.

[duke-cliff]

@Grokzen Yup, it works. The redirect logic is working the same as before, the only difference is when it could get a connection from the slaves for a slot. The rest logic is the same to how you deal with exceptions/redirects(moved/ask) on masters:

You can see our CPU was dropped from 16% to 5-6% because of this change.

[duke-cliff]

I will have more optimization for this PR coming. testing locally first.

[jerroydmoore]

When I merged master into this branch, the diff went away. I think this PR might be safe to close.