[Work in Progress] Cluster-wide Multi-key commands

[nilanshu-sharma]

No description provided.

[github-actions]

✅ Pull request no significant performance differences ✅

Summary

New baseline 'pull_request' is WITHIN the 'main' baseline thresholds.

Full Benchmark Comparison

Comparing results between 'main' and 'pull_request'

Host '7e3d460d803a' with 4 'x86_64' processors with 15 GB memory, running:
#17~24.04.1-Ubuntu SMP Mon Dec  1 20:10:50 UTC 2025

ValkeyBenchmarks

Client: GET benchmark metrics

Malloc (total): results within specified thresholds, fold down for details.

Malloc (total) *	p0	p25	p50	p75	p90	p99	p100	Samples
main	52	55	55	57	58	58	58	5
pull_request	53	54	55	57	57	57	57	5
Δ	1	-1	0	0	-1	-1	-1	0
Improvement %	-2	2	0	0	2	2	2	0

Client: GET benchmark | parallel 20 | 20 concurrent connections metrics

Malloc (total): results within specified thresholds, fold down for details.

Malloc (total) *	p0	p25	p50	p75	p90	p99	p100	Samples
main	52	56	57	59	61	63	63	24
pull_request	51	54	58	59	60	64	64	22
Δ	-1	-2	1	0	-1	1	1	-2
Improvement %	2	4	-2	0	2	-2	-2	-2

Connection: GET benchmark metrics

Malloc (total): results within specified thresholds, fold down for details.

Malloc (total) *	p0	p25	p50	p75	p90	p99	p100	Samples
main	4	4	4	4	4	4	4	8
pull_request	4	4	4	4	4	4	4	7
Δ	0	0	0	0	0	0	0	-1
Improvement %	0	0	0	0	0	0	0	-1

Connection: GET benchmark – NoOpTracer metrics

Malloc (total): results within specified thresholds, fold down for details.

Malloc (total) *	p0	p25	p50	p75	p90	p99	p100	Samples
main	4	7	8	9	9	9	9	7
pull_request	4	7	8	9	9	9	9	7
Δ	0	0	0	0	0	0	0	0
Improvement %	0	0	0	0	0	0	0	0

Connection: Pipeline array benchmark metrics

Malloc (total): results within specified thresholds, fold down for details.

Malloc (total) *	p0	p25	p50	p75	p90	p99	p100	Samples
main	14	18	25	29	29	29	29	7
pull_request	14	18	25	29	29	29	29	7
Δ	0	0	0	0	0	0	0	0
Improvement %	0	0	0	0	0	0	0	0

Connection: Pipeline benchmark metrics

Malloc (total): results within specified thresholds, fold down for details.

Malloc (total) *	p0	p25	p50	p75	p90	p99	p100	Samples
main	7	24	24	25	25	25	25	7
pull_request	7	24	24	25	25	25	25	7
Δ	0	0	0	0	0	0	0	0
Improvement %	0	0	0	0	0	0	0	0

HashSlot – {user}.whatever metrics

Malloc (total): results within specified thresholds, fold down for details.

Malloc (total) *	p0	p25	p50	p75	p90	p99	p100	Samples
main	0	0	0	0	0	0	0	17
pull_request	0	0	0	0	0	0	0	18
Δ	0	0	0	0	0	0	0	1
Improvement %	0	0	0	0	0	0	0	1

ValkeyCommandEncoder – Command with 7 words metrics

Malloc (total): results within specified thresholds, fold down for details.

Malloc (total) *	p0	p25	p50	p75	p90	p99	p100	Samples
main	0	0	0	0	0	0	0	826
pull_request	0	0	0	0	0	0	0	828
Δ	0	0	0	0	0	0	0	2
Improvement %	0	0	0	0	0	0	0	2

ValkeyCommandEncoder – Simple GET metrics

Malloc (total): results within specified thresholds, fold down for details.

Malloc (total) *	p0	p25	p50	p75	p90	p99	p100	Samples
main	0	0	0	0	0	0	0	1952
pull_request	0	0	0	0	0	0	0	1950
Δ	0	0	0	0	0	0	0	-2
Improvement %	0	0	0	0	0	0	0	-2

ValkeyCommandEncoder – Simple MGET 15 keys metrics

Malloc (total): results within specified thresholds, fold down for details.

Malloc (total) *	p0	p25	p50	p75	p90	p99	p100	Samples
main	0	0	0	0	0	0	0	374
pull_request	0	0	0	0	0	0	0	377
Δ	0	0	0	0	0	0	0	3
Improvement %	0	0	0	0	0	0	0	3

[adam-fowler]

I've had an initial look at this.

• There is a lot of memory allocation.
• The public protocol API needs some work. I'm not certain how it can be improved but if feels a bit hacky
• Have you considered how this is going to work for pipelined commands?

[adam-fowler]

Suggested change

	return slotOrder.map { (indices: slotIndices[$0]!, slot: $0) }
	return slotIndices.map { (indices: $0.value, slot: $0.key) }

[adam-fowler]

Do the slots indices need to be ordered? If not you can replace this code with

for (i, key) in keys.enumerated() {
    let slot = HashSlot(key: key)
    slotIndices[slot, default: []].append(i)
}

[adam-fowler]

Is it worth not converting to the response type and parsing the RESP in the assemble function. This could avoid multiple allocations, if a response is converted to an Array. In the case of MGET you end up rebuilding a RESPToken anyway.

[adam-fowler]

Suggested change

	func subCommand(for indices: [Int]) -> Self
	func createSubCommand(for indices: [Int]) -> Self

[adam-fowler]

Suggested change

	static func assemble(
	static func assembleResults(

[adam-fowler]

Is it worth having an early out when there is only one partition which calls the standard code path?