perf: optimize Ring.Get() when a large number of instances are in a state for which the extended replication set is requested

[pracucci]

What this PR does:

The Ring.Get() is very inefficient when a large number of instances are in a state for which the extended replication set is requested. This is not a new issue, and there was a previous optimization attempt in #672 by @56quarters. Unfortunately the issue is not fixed yet and we can see it when the magnitude of instances increases.

In this PR I propose an optimization based on the fact that looking up small maps is slower than just searching for corresponding items in slices (maps become fasters when the number of unique items increases, but for few items slices are just faster).

The result of this optimization is not huge, but I think it's still a step forward. It's still very slow when we have a large number of instances in a state for which the extended replication set is requested, compared to the case all instances are ACTIVE.

goos: darwin
goarch: arm64
pkg: github.com/grafana/dskit/ring
cpu: Apple M3 Pro
                                                         │ new-before.txt │            new-after.txt            │
                                                         │     sec/op     │    sec/op     vs base               │
Ring_Get/with_zone_awareness-11                             1008.1n ± 19%   553.1n ±  8%  -45.13% (p=0.002 n=6)
Ring_Get/one_excluded_zone-11                                624.6n ±  1%   393.0n ± 12%  -37.08% (p=0.002 n=6)
Ring_Get/without_zone_awareness-11                           347.1n ±  4%   334.5n ±  1%   -3.62% (p=0.002 n=6)
Ring_Get/without_zone_awareness,_not_enough_instances-11     268.5n ±  1%   256.2n ±  1%   -4.56% (p=0.002 n=6)
Ring_Get_OneZoneLeaving/one_zone_leaving_=_true-11          3739.3µ ±  5%   903.0µ ±  2%  -75.85% (p=0.002 n=6)
Ring_Get_OneZoneLeaving/one_zone_leaving_=_false-11          1.905µ ±  6%   1.169µ ±  2%  -38.62% (p=0.002 n=6)
geomean                                                      2.734µ         1.643µ        -39.91%

                                                         │ new-before.txt │             new-after.txt             │
                                                         │      B/op      │     B/op      vs base                 │
Ring_Get/with_zone_awareness-11                              0.000 ± 0%       0.000 ± 0%        ~ (p=1.000 n=6) ¹
Ring_Get/one_excluded_zone-11                                0.000 ± 0%       0.000 ± 0%        ~ (p=1.000 n=6) ¹
Ring_Get/without_zone_awareness-11                           0.000 ± 0%       0.000 ± 0%        ~ (p=1.000 n=6) ¹
Ring_Get/without_zone_awareness,_not_enough_instances-11     0.000 ± 0%       0.000 ± 0%        ~ (p=1.000 n=6) ¹
Ring_Get_OneZoneLeaving/one_zone_leaving_=_true-11         109.6Ki ± 0%     124.3Ki ± 0%  +13.45% (p=0.002 n=6)
Ring_Get_OneZoneLeaving/one_zone_leaving_=_false-11          0.000 ± 0%       0.000 ± 0%        ~ (p=1.000 n=6) ¹
geomean                                                                 ²                  +2.13%               ²
¹ all samples are equal
² summaries must be >0 to compute geomean

                                                         │ new-before.txt │            new-after.txt            │
                                                         │   allocs/op    │ allocs/op   vs base                 │
Ring_Get/with_zone_awareness-11                              0.000 ± 0%     0.000 ± 0%        ~ (p=1.000 n=6) ¹
Ring_Get/one_excluded_zone-11                                0.000 ± 0%     0.000 ± 0%        ~ (p=1.000 n=6) ¹
Ring_Get/without_zone_awareness-11                           0.000 ± 0%     0.000 ± 0%        ~ (p=1.000 n=6) ¹
Ring_Get/without_zone_awareness,_not_enough_instances-11     0.000 ± 0%     0.000 ± 0%        ~ (p=1.000 n=6) ¹
Ring_Get_OneZoneLeaving/one_zone_leaving_=_true-11           19.00 ± 0%     27.00 ± 0%  +42.11% (p=0.002 n=6)
Ring_Get_OneZoneLeaving/one_zone_leaving_=_false-11          0.000 ± 0%     0.000 ± 0%        ~ (p=1.000 n=6) ¹
geomean                                                                 ²                +6.03%               ²
¹ all samples are equal
² summaries must be >0 to compute geomean

Which issue(s) this PR fixes:

N/A

Checklist

• Tests updated

---

Note

Improves performance of Ring.Get() when many instances extend the replica set, especially under zone-awareness.

• Replaces distinctHosts slice/map usage with a hybrid stringSet (util.go) to reduce lookups/allocations; adds comprehensive tests
• Switches per-zone tracking from maps to slices indexed by zone index (totalHostsPerZone, examinedHostsPerZone, foundHostsPerZone); updates canStopLooking() signature accordingly
• Adds explicit zoneIndex resolution with validation and panics on misuse if zone-awareness disabled
• Adjusts main selection loop to use stringSet.len()/contains()/add() and new stop condition
• Expands and parameterizes BenchmarkRing_Get_OneZoneLeaving (larger instances, more zones, and both with/without one zone leaving) with deterministic assertions

Net effect: substantial reductions in Ring.Get() time in pathological scenarios; no API changes.

^{Written by Cursor Bugbot for commit 9f1c505. This will update automatically on new commits. Configure here.}

[56quarters]

LGTM, nicely done.

[56quarters]

Maybe assert that setMap is nil before any adds?

[pracucci]

Done in 2b36903

[alexweav]

LGTM