CrossSlotPipeliningError validation seems incorrect?

[ejinotti-mdsol]

From my reading of https://redis.io/topics/cluster-spec#keys-hash-tags it seems that the magic squiggly brackets to force a specific value for hashing should be acceptable anywhere in the key. However the raise of CrossSlotPipeliningError here:

redis-rb/lib/redis/cluster.rb

Line 82 in 6542934

raise CrossSlotPipeliningError, command_keys if node_keys.size > 1

Seems to only accept the magic squigglies as a prefix:

irb(main):001:0> REDIS.pipelined { ['{zed}one', '{zed}two'].each { |k| REDIS.exists?(k) } }
=> [false, false]
irb(main):002:0> REDIS.pipelined { ['one{zed}', 'two{zed}'].each { |k| REDIS.exists?(k) } }
Traceback (most recent call last):
        1: from (irb):2
Redis::Cluster::CrossSlotPipeliningError (Cluster client couldn't send pipelining to single node. The commands include cross slot keys. ["zed", "zed"])
irb(main):003:0> REDIS.pipelined { ['one{zed}one', 'two{zed}one'].each { |k| REDIS.exists?(k) } }
Traceback (most recent call last):
        2: from (irb):2
        1: from (irb):3:in `rescue in irb_binding'
Redis::Cluster::CrossSlotPipeliningError (Cluster client couldn't send pipelining to single node. The commands include cross slot keys. ["zed", "zed"])

This is v4.3.1

[ejinotti-mdsol]

One additional piece of info: this was using a cluster with one shard and 5 replicas.

[johnduhart]

I had a brief look at this for the team, and I'm a bit suspicious of this code path:

redis-rb/lib/redis/cluster.rb

Line 290 in baf5706

node_keys = pipeline.commands.map { |cmd| find_node_key(cmd) }.compact.uniq

This is going to call find_node_key for each command individually, which results in looking up a slot (okay), and then calling find_node_key_of_slave. And then, here:

redis-rb/lib/redis/cluster/slot.rb

Lines 25 to 30 in baf5706

	def find_node_key_of_slave(slot)
	return nil unless exists?(slot)
	return find_node_key_of_master(slot) if replica_disabled?
	@map[slot][:slaves].sample
	end

A random replica will be chosen, which may result in different node_keys even though all keys are in the same slot.

[supercaracal]

It is a bug. It seems that the cross-slot validation should be skipped when commands in pipeline are sent to replicas.

[supercaracal]

Actually, forget about that. The cross-slot validation in pipelining should be geared toward primary nodes.