Implementation for transaction

[supercaracal]

This issue is just opened to purpose to put together related pull requests. It will be closed when they are merged or closed.

• feat: support transaction #277
• Fix MULTI transactions calling the block twice #294
• Support watch & unwatch properly #295
• Implement "pinning" cluster connection to a single node - RedisClient::Cluster#with #298
• Considerations:
  • the determination of the node that the transaction should be sent to
    • The client can't determine the node until the first command included keys is passed.
  • the handling redirects duaring the resharding
    • The client should not redirect any errors caused by user side bugs.
  • the optimistic locking with the watch command
    • The watch command can watch multiple keys. However it should be called to a single node in a sharded cluster. So in practically, it allows only keys to a single slot.
  • the key validation
    • There is a discussion whether should do it proactive or reactive.
    • I think the client doesn't need the proactive validation if the server do the strict validation.
  • the compatibility between the redis-clustering gem
    • I think the redis-cluster-client gem should not be complex in excess due to the redis gem.
  • the design
    • We should keep the interface same as the redis-client as possible.
    • We shouldn't bring any complexity to the code base just for the feature. Once the code messed up, we'll be hard to maintain it.
      • Of cource all bugs should be fixed.
    • Our gem shouldn't affect in excess by the redis-clustering gem.

[KJTsanaktsidis]

OK - I think I'm going to close all three of those PR's (each of which try and solve the whole implementation at once, in three different ways), and begin opening smaller PR's for some of the prereq work to get cluster transactions working slightly better.

• Perform cluster bookeeping tasks _before_ determining retry #300
• Refactor key-slot conversion logic #301

[KJTsanaktsidis]

OK, next up, is a new test helper & some further refactoring of the error handling logic in the router:

• Add infrastructure for capturing Redis commands in tests #304
• Handle timeouts to blocking_v specially via a custom client class #305

(Thanks for looking at the previous two so promptly!)

[KJTsanaktsidis]

I opened #306 as an alternative to handling blocking_v specially in response to your feedback.

[supercaracal]

• Delete special handling of blocking command timeouts #306

The above pull request has been merged. Thank you.

[KJTsanaktsidis]

And for my next trick, now these two blocks of code are identical, so we can refactor this:

• Refactor redirection error handling into one block #307

[KJTsanaktsidis]

Two more PR's for your consideration @supercaracal :)

• Add middleware to ensure Router only handles errors for its clients #310
• Handle unknown subclasses of RedisClient::Error in clustering redis/redis-rb#1244

[KJTsanaktsidis]

OK, couple more @supercaracal ! This is code you've mostly looked at before I think but I think they can now be merged in because all of the prerequisites are done.

• Implement RedisClient::Cluster#with (again!) #311 - This implements #with again, without any kind of wrapping. The raw RedisClient instance is returned to the caller, which means the whole thing is now about 5 lines.
• Implement RedisClient::Cluster::Command#extract_all_keys #312 - this implements an internal extract_all_keys method for getting all the keys out of a command. I want this for performing ahead-of-time checks on keys passed in to pinned clients, but it's probably a useful thing to have in the gem anyway.

[KJTsanaktsidis]

This week's PR's @supercaracal

• Fix error identification middleware calling into the wrong super #315 - Fix a simple bug from the error identification PR previusly
• Validate key slots used with RedisCluster::Client#with #314 - More complicated: implements raising TransactionConsistencyError when mixing up slots.

As always, thank you! 🙇