Don't use master/slave terminology (non-breaking changes only)

[zuiderkwast]

Primary and replica are the preferred terms.

The current situation is master/replica, with slave kept as an alias of replica. Some commands like ROLE still returns "master" and "slave".

So far "master" was accepted, so there aren't yet any aliases for "master". Let's introduce aliases and deprecate master.

1. Add aliases for master (not a breaking change)
2. Delete master and slave terminology (breaking change) EDIT: After feedback, we've decided to skip the breaking changes. We can only that as a client opt-in. See [NEW] Opt-in for inclusive language (primary/replia in ROLE reply, CLUSTER SHARDS, etc.) #751.

Occurrences of "master" where we can add "primary" and make "master" an alias:

• CLIENT KILL TYPE MASTER
• CLIENT LIST TYPE MASTER
• SENTINEL GET-MASTER-ADDR-BY-NAME

Note: This issue was edited to not include any breaking changes. The initial suggestion included breaking changes, but this was not appreciated by the users. Thus, the downvotes.

[GaryPWhite]

Hi! Would this be a decent "first issue" for a techie who's interested in getting more involved with the alt/redis ecosystem? I'd be happy to use this as a way to get more familiar with the repo :)

[zuiderkwast]

@GaryPWhite Yes, that's great, but first we need to analyze and decide exactly what we need to do and in which steps. There are implications like Redis compatibility to consider.

[zuiderkwast]

@GaryPWhite Actually I think you can start with the non-breaking changes of simply adding PRIMARY aliases (or rather make MASTER an alias of PRIMARY) for these:

• CLIENT KILL TYPE MASTER
• CLIENT LIST TYPE MASTER
• SENTINEL GET-MASTER-ADDR-BY-NAME

Look into how SLAVE is already an alias of REPLICA in these commands.

[hwware]

Agree with master as an alias of primary. But suggest to hold the master and slave name, maybe they will remove in future version

[zuiderkwast]

Yes, the breaking change is a "major decision". We should analyze the implications a bit more too. There may be reasons to be compatible with other forks and implementations so we should also keep on eye on them.

We may even consider some possibility for conditional compilation or similar.

[PingXie]

I would vote for "primary" and "replica".

As for changes, I now think more and more that we need a "compat" knob. It can be either compile time or runtime (via config).

We could also consider a staged change:

1. Anything non-user facing, i.e., comments, function/variable names, etc

2. Aliasing input to support both terminologies (but also under the "compat" switch)

3. Changing output. Aliasing would not work in this case. We will default to the new terminology but it can be overridden by the user, again either compile time or runtime (pending decision)

[zuiderkwast]

Yes, 1 and 2 can be done immediately IMO, but in two separate PRs.

[stockholmux]

If there is a compat knob, it would be good to have a pre-determined sequence of deprecation.

E.g. version n introduce knob and optional new terminology, version n+1 new terminology becomes default, old terminology is deprecated, version n+2 remove old terminology.

Otherwise these things just linger. Pull off the bandaid, as they say.

[PingXie]

Pull off the bandaid, as they say

Yes we should do it at some point but I think the timing of it falls in another major decision and we need to give our users ample time to adjust. As much as I am personally in strong favor of ripping this bandaid off asap I also understand that the success of this project is very much dependent on "continuity". It takes time to turn this huge ship around.

[stockholmux]

3 major version would be enough transition time, I'd assume.

But mainly I'm just advocating have a set plan before changes happen so it doesn't languish for an undetermined amount of time.

[PingXie]

I'm just advocating have a set plan before changes happen so it doesn't languish for an undetermined amount of time.

Timing aside, I think one option, as you suggested too, is to just remove this compat knob and we are done. I have yet to think of other options because as far as output is concerned, it is a binary decision. So the plan in my opinion is more about deliberating a timeline with the help of the community, than an engineering plan to help with a smooth transition. In other words, I think #36 (comment) is the best we can do. Curious to hear other folks' thoughts.

[hpatro]

de-crapify 😂

[DarrenJiang13]

Maybe we need to consider some compatibility to SDKs like jedis.

Like "cluster nodes", it may return with "master" field and might be used by some SDKs.