Pipeline-based script loading and retries

[marcbachmann]

This PR builds on #1497 (first three comments)
The last two commits are relevant for the new script load behavior.

Changelog

• 🚀 Scripts are loaded into the pipeline instead in advance
  • It improves performance during first load
  • Improves cluster support with scripts
• 🐛 Automatically retry failed evalsha commands if the script is missing in redis, this improves the resiliency
• 🐛 Use the same script caching logic for standalone execution of scripts
• 🐛 Only send the scripts to a socket exactly once

[marcbachmann]

Instead of doing those decisions in here, we could move it into the command.toWritable function.
But we don't have any stream context there. https://github.com/luin/ioredis/blob/b8177479c348aa4bbd467fa944d61fe9b35aec19/lib/redis/index.ts#L790

If we would pass in the stream to the toWritable function, we could maintain the script cache per cluster node or socket instead on the redis instance.

[marcbachmann]

@luin what do you think about those changes to make the scripts more reliable?
The script implementation is now much more straight forward.
Just how a socket-like object like the pipeline gets passed to the redis instance could be better at the moment. Because of that the check is a bit more complicated during write.

But in general it completely eliminates all caching issues as we just load the script once with every new connection. Previously it resulted in quite a lot of NOSCRIPT errors. I think this drawback is good enough for most cases.
We could also only use eval when multi was used. To guarantee that the script is always there in that case. For the other cases it will do the fallback.

[marcbachmann]

We could remove this fallback if order of delivery needs to be guaranteed.
It will still clear the cache for the next command that gets sent to redis.
https://github.com/luin/ioredis/pull/1499/files#diff-3618aaced6e025d978d2cdde354e756687916385e9024868d0ad3fff887c55c6R24

[luin]

I'm thinking to include this breaking change in the next major version #1504 so we can just retry custom commands in pipeline as well.

[marcbachmann]

For the pipeline we could move the buffer logic into the Redis class itself.
Maybe even persist the pipeline itself in the queue instead of calling sendCommand that often.

[luin]

Hey @marcbachmann 👋,

Sorry for the delay. I'm planning v5 in #1504. So do you mind change the target branch of this PR to v5 and resolve the conflicts?

[marcbachmann]

@luin I've rebased this.
At least one cluster.disconnect test is missing as I've removed the script based ones:
https://github.com/luin/ioredis/pull/1499/files#diff-25757dbbc9670af28c244c85ae7b36c1224b1963223299f4df48963785141a6fL415

[luin]

Awesome! Nice and clean improvements!