[Feature-Request]: Cross-Slot Command Execution in ValKey Cluster

[bsmanit]

Problem/Use-case:

As we know that single key command << pipelining <<< multi key command to write/read n key from Redis, and to leverage this multi key command on Redis to get the higher ops we can enable cross slot command execution only if all multi key are owned by executing shard (not involving multiple shards for single multi key command).

Description:
Introduce a new configuration option in redis.conf, enable-cross-slot, which defaults set to no but can be set to yes when running Redis in cluster mode.
This feature enables cross-slot command execution if keys are owned by single shards.

Design:
Client: Already aware about slot range owned by each shard of Redis cluster and update with multiple topology refresh strategy.

To execute n key multi-key command, client groups this into K commands with hash ranges owned by each shards. (K is number of shards in cluster)
Then client async executes these multi key commands and join and return.

Since In above multiple nodes are involved to execute multi-key command, so their execution is not guaranteed to be atomic (so, they can actually break the atomic design of many Redis commands).

For that we can add a new config in redis.conf enable-cross-slot yes (by default no) and only considered when running Redis in cluster mode. {similar to Redis Proxy: https://github.com/RedisLabs/redis-cluster-proxy}

Alternatives Considered:
Leveraging Redis Proxy: Utilizing existing Redis Proxy solutions like https://github.com/RedisLabs/redis-cluster-proxy can mitigate the issue but adds another layer of complexity to the architecture.

Manual grouping of multi-key commands by the client based on hash slots: Given Redis's extensive 16K+ hash slots, users typically issue multi-key commands involving 10 or 100 keys. However, grouping these keys by hash slots often results in nearly single or double key multi-get commands. Unfortunately, this approach doesn't significantly enhance performance.

Additional Information:
With enable-cross-slot set to yes, Redis will ensure atomic execution of multi-key commands in individual shards, enhancing the reliability and consistency of operations.

Clients must be aware of the slot ranges owned by each shard in the Redis cluster and update their topology accordingly to leverage this feature effectively.

[madolson]

I kind of like this idea, but my thought has been it should be scoped down to "very" specific commands. Specifically, I think it should be the commands with logical fan-in and fan out behavior, like MGET, MSET, DEL, etc. Esoteric commands like like SUNIONSTORE probably don't need to support multi-slot.

With enable-cross-slot set to yes, Redis will ensure atomic execution of multi-key commands in individual shards, enhancing the reliability and consistency of operations.

I've mentioned this on other threads, but I don't think this should be a server feature, but instead a client capability they should be able to opt-in to.

[bsmanit]

@madolson Yes, I agree this should be very command-specific. The client should handle fan-in/fan-out behavior. Some clients, like Lettuce, already do this.
For example, Lettuce groups keys by hash slot for MGET and then sends them to the Redis server. With the new feature, it can group keys by the hash slot range owned by each shard instead. This will improve read throughput significantly.

[zuiderkwast]

Generic clients need to support both values this config. Do you think they should call CONFIG GET to find out if they can use crosslot commands or not?

I think it's better that it's opt-in for the client. Then only the clients that ask for it will get it, without affecting other clients and users. Idea:

CLIENT CROSSSLOT ON

It can be sent in pipeline with HELLO or AUTH just after connecting.

[zuiderkwast]

Btw, are you sure that multiple GET in a pipeline is much slower than MGET? I doubt it.

[zuiderkwast]

Third comment: I don't think it should be limited to just a few commands. That would just be another confusion. If it's enabled, it can apply to sunionstore and even to transactions affecting multiple slots.

[madolson]

Third comment: I don't think it should be limited to just a few commands. That would just be another confusion. If it's enabled, it can apply to sunionstore and even to transactions affecting multiple slots.

The original point of the cross-slot was to detect misuse. I think SUNIONSTORE where the source and target are different slots is almost always misuse. The only material case I can think of is you have an intermediary storage on the same node where you are trying to "aggregate data", and then plan on deleting it. I'm not sure that is materially that important or useful.

[zuiderkwast]

The original point of the cross-slot was to detect misuse

Yeah, so that's why it should be forbidden by default. MSET is an abuse too because it appears to the end user to be atomic, which it isn't (if the client silently splits the command and sends to different nodes).

If you get some keys from SCAN, you know they're on the same node. You may want to compute an SUNION on these keys. The alternative is to compute the union on the client side, something that can always be done even for a cross-node union, so allowing crosslot is an optimization for clients in this case too.

If we want to prevent this (because we say it's an invalid use case) we'd need additional command flags and logic just to keep track of whether this is allowed for each command. I don't like it.

[madolson]

To your point about atomicity. SUNIONSTORE should probably always be an atomic operation. I can see a world where MSET is not expected to be atomic though, because you are overwriting different keys with some value. I think the main distinction that I am seeing is that there are:

1. Read commands
2. Write commands
3. Read + modify + write commands.

That last category is the one that I don't think should be allowed to fanout and operate across slots.

[zuiderkwast]

OK so not SUNIONSTORE, but for SUNION it would be OK since it's a read command? Makes some sense now.

[madolson]

Yeah, sunion seems like it would be okay to me.

[zuiderkwast]

Great, so can we agree on this↓?

• Client opt-in (not a config).
• Allow for read and write commands, but forbid for read-modify-write commands.
• We need to add a command flag for the read-modify-write commands, or flag them as read+write, because now they're just flagged as write. (The *STORE commands and the commands with a STORE or STOREDIST argument: SORT, GEORADIUS, GEORADIUSBYMEMBER)

[madolson]

Yeah, I'm aligned with what you said.

We do theoretically have all this information already in the .json files, as they include r/w information for which keys. It's just a matter of computing it.

[PingXie]

I see the value in cross-slot operations within a single-shard setup, and it seems to provide a seamless experience for the client application.

However, I'm unclear on how cross-slot operations would work intuitively in a multi-shard setup, assuming we are not talking about fanning out to other shards on the caller's behalf, i.e., the in-proc proxy? While the client application is aware of slot mapping, it typically doesn't have control over it. The admin and user roles are usually separate, even in self-managed environments. This separation is important because the primary reason an application desires cross-slot migration is atomicity. Simply supporting cross-slot operations isn't enough; the slots relevant to the application also need to be guaranteed on the same shard.

Additionally, I'm not convinced by the performance argument. Wouldn't pipelining already achieve the desired performance improvements?

[madolson]

Additionally, I'm not convinced by the performance argument. Wouldn't pipelining already achieve the desired performance improvements?

Performance is something we can test for sure. There is a fair amount of overhead in handling the call. It also removes one edge case from clients, which is partial success within a shard. It also makes idempotency a bit easier, since either all the operations succeeds on a shard or none of them did.

I agree with one other point though, is that this becomes a lot more compelling as a proxy feature as compared to a server feature. Maybe we should create an issue for maintaining a first party Valkey proxy?

[zuiderkwast]

this becomes a lot more compelling as a proxy feature as compared to a server feature

I'm skeptical to a proxy that tries to hide the fact that there's a cluster behind it. A proxy can't do anything more than what a client can do.

What a client or proxy would need for that is cross-node transactions. (If we ever want that, we could add 2PC with key locking, like MULTI-PREPARE-EXEC.)

[madolson]

I'm skeptical to a proxy that tries to hide the fact that there's a cluster behind it. A proxy can't do anything more than what a client can do.

One point that has come up fairly frequently is that not everyone has the ability to upgrade the client, but if you put a proxy you can independently run and update the proxy.