Support Sidekiq required feature set

[casperisfine]

As discussed in redis/redis-rb#1070 (comment), I'm trying to convert Sidekiq to use redis-client, so far the missing features I found:

• url: connection param.
• id connection param (CLIENT SETNAME)
• helper for SCAN family of commands
• Redis::Namespace equivalent (this one is debatable, could potentially be implemented as some kind of middleware that has COMMAND knowledge)
• reconnect_attempts, Sidekiq sets it to 1. However if we are to implement this, we should have COMMAND knowledge to only retry readonly tagged commands.
• Empty pipelines. Sidekiq test suite allowed to noticed empty pipeline were crashing.
• The sidekiq test suite has some Redis sentinel configuration, we should support that as well (Add sentinel support #8).

[mperham]

I'm targeting Sidekiq 7.0 sometime mid-year, probably in August. Do you see this driver as stable and feature complete by then?

[mperham]

I'm happy to adjust or remove some feature usage. I tried to remove reconnect_attempts once but it turns out to be hugely useful on Redis SaaS platforms where Redis migrates or moves under the covers and Sidekiq is just supposed to recover gracefully.

[casperisfine]

Do you see this driver as stable and feature complete by then?

I believe so yes, at least for Sidekiq use case (i.e. no redis-cluster).

Also not too sure about the hiredis driver, native extensions can be tricky to test properly, but I don't think it makes that big a difference for Sidekiq, so it might be marked as "experimental" in the couple first releases.

I'm happy to adjust or remove some feature usage.

It's fine, reconnect_attempts is indeed very useful. I just want to implement it better (only retry idempotent commands). That's what I was working on this afternoon: #9, it will probably be ready tomorrow-ish.

As for Sentinel, I need to have a look, but it shouldn't be too hard.

[casperisfine]

reconnect_attempts, Sidekiq sets it to 1. However if we are to implement this, we should have COMMAND knowledge to only retry readonly tagged commands.

So I looked into having a list of commands safe to retry, but it's actually not that simple, because in many cases it also depends on the arguments and wether you care about the return value.

For example ZREM ... can be retried, but if the caller cares about the return value, then it might be incorrect to retry it.

So instead I added a #call_once method that prevent retries. So it basically allow to choose between AtMostOnce and AtLeastOnce consistency. I'll try to go over sidekiq/sidekiq#5253 to see if there's places that shouldn't be retried.

[casperisfine]

Done.