feat(payload): adopt AtLeastOneHundredth for Config::unique_tag_ratio

[goxberry]

What does this PR do?

Change the public dogstatsd::Config::unique_tag_ratio field from f32 to
the AtLeastOneHundredth alias of BoundedProbability<{ f32::to_bits(0.01) }>.
The try_from impl enforces the finite + [0.01, 1.0] invariant at
deserialize time, so the redundant MIN_UNIQUE_TAG_RATIO..=MAX_UNIQUE_TAG_RATIO
range check in common::tags::Generator::new is removed. The WARN-level
check for values in [MIN, WARN_UNIQUE_TAG_RATIO] is preserved.

The new type threads through MemberGenerator::new,
common::tags::Generator::new, dogstatsd::common::tags::Generator::new,
and opentelemetry::common::TagGenerator::new. The OTel UNIQUE_TAG_RATIO
constant becomes a const AtLeastOneHundredth via a match on try_new,
with unreachable!() in the Err arm (the bound is provably valid at
const-eval time). MAX_UNIQUE_TAG_RATIO, no longer referenced outside
tests, is now #[cfg(test)].

At the comparison site in <common::tags::Generator as Generator>::generate,
.get() extracts the inner f32 so RNG sequences and bit-exact output are
preserved.

Motivation

Fifth and final per-field PR in Phase 1 of the BoundedProbability rollout.
Pushes range validation to deserialize time once and removes the runtime
range check from the tag-generator constructor. AtLeastOneHundredth is
chosen over Probability to preserve the existing 0.01 floor and over
AtLeastOneTenth so that in-the-wild configs in [0.01, 0.10) keep
deserializing.

Verification

cargo test -p lading-payload: 250 passed, 0 failed.

cargo clippy -p lading-payload --all-targets -- -D warnings: clean.

Criterion comparison vs. parent (goxberry/probability-multivalue-pack-prob)
for cargo bench -p lading-payload --bench dogstatsd:

Bench	Median Δ	Criterion verdict
dogstatsd_setup	+1.59%	within noise
dogstatsd_throughput/1MiB	+2.27%	flagged regressed
dogstatsd_throughput/10MiB	+0.83%	no change
dogstatsd_throughput/100MiB	+1.17%	within noise
dogstatsd_throughput/1GiB	-0.88%	no change

.get() is const fn returning a copied f32, so the IR at the comparison
site is bit-identical. The 1 MiB regression is most plausibly machine-state
noise (the bench was run right after a CPU-heavy OTel job had warmed the
system); the larger sizes do not reproduce it. OTel benches were skipped
because the OTel hot path is the same common::tags::Generator::generate
already exercised by dogstatsd, and the OTel call sites only pass
UNIQUE_TAG_RATIO to TagGenerator::new (init time, not hot path).

Related issues

Phase 1 of .claude/probability-type-rollout-plan.md, building on:

• payload config: add new Probability<const u32: MIN_AS_BITS> type #1874 (introduces BoundedProbability)
• PR 1: TimestampConfig::probability
• PR 2: ValueConf::float_probability
• PR 3: Config::sampling_probability
• PR 4: Config::multivalue_pack_probability

Additional Notes

Wire format is unchanged: BoundedProbability continues to round-trip as a
bare f32 via serde(into = "f32", try_from = "f32"). The
MIN_UNIQUE_TAG_RATIO and WARN_UNIQUE_TAG_RATIO constants are retained
because the WARN-level check still references them.

[goxberry]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• feat(payload): adopt AtLeastOneHundredth for Config::unique_tag_ratio #1881 👈 (View in Graphite)
• feat(payload): adopt Probability for Config::multivalue_pack_probability #1880
• feat(payload): adopt Probability for Config::sampling_probability #1879
• feat(payload): adopt Probability for ValueConf::float_probability #1878
• feat(payload): adopt Probability for TimestampConfig::probability #1877
• feat(payload): improve BoundedProbability ergonomics for adoption #1876
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[datadog-prod-us1-3]

✨ Fix all issues with BitsAI

⚠️ Warnings

🚦 3 Pipeline jobs failed

Continuous integration | Rust Actions (Check/Fmt/Clippy) (macos-latest, fmt)     

🔧 Fix in code (Fix with Cursor).

Compilation error in dogstatsd.rs:278: expected one statement but found multiple.

Continuous integration | Rust Actions (Check/Fmt/Clippy) (ubuntu-latest, fmt)     

🔧 Fix in code (Fix with Cursor).

Formatting errors detected in dogstatsd.rs and common.rs. Please fix the formatting differences.

Changelog Check | changelog-check     

🛟 This job is unlikely to succeed on retry. Please review your pipeline configuration.

No changes to CHANGELOG.md detected. Add 'no-changelog' label if this is intentional.

Useful? React with 👍 / 👎

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: 4e9ba85 | Docs | Datadog PR Page | Give us feedback!}

[goxberry]

PR 5: Adopt AtLeastOneHundredth for Config::unique_tag_ratio

Context

This is the fifth and final per-field PR in Phase 1 of
.claude/probability-type-rollout-plan.md. PRs 1–4 already wired
Probability into TimestampConfig::probability, ValueConf::float_probability,
Config::sampling_probability, and Config::multivalue_pack_probability.

The remaining f32 probability/ratio in lading_payload is
dogstatsd::Config::unique_tag_ratio. This PR changes its type from f32 to
AtLeastOneHundredth (= BoundedProbability<{ f32::to_bits(0.01) }>) and
threads the typed value down through the tag generator constructors so that
range validation happens once at deserialize time rather than at
Generator::new time.

AtLeastOneHundredth is used instead of Probability because the existing
runtime check enforces a floor of 0.01 (MIN_UNIQUE_TAG_RATIO), and using
AtLeastOneTenth would tighten the bound and reject in-the-wild configs in
[0.01, 0.10). The decision is locked in .claude/probability-type-rollout-plan.md
("Decisions locked in").

Branch: goxberry/probability-unique-tag-ratio, stacked on top of
goxberry/probability-multivalue-pack-prob via gt.

Assumptions

• AtLeastOneHundredth is already defined in lading_payload/src/common/config.rs:184.
• BoundedProbability::try_new is const fn, so the OTel
  UNIQUE_TAG_RATIO: f32 = 0.75 const can become const AtLeastOneHundredth
  via match AtLeastOneHundredth::try_new(0.75).
• The wire format is unchanged: BoundedProbability round-trips as a bare
  f32 via serde(into = "f32", try_from = "f32").
• The MIN_UNIQUE_TAG_RATIO, WARN_UNIQUE_TAG_RATIO, and
  MAX_UNIQUE_TAG_RATIO constants stay (used by tests and the warn-check;
  the rollout plan's "Out of scope" section pins this).
• Hot-path bit-exactness is preserved by calling .get() at the comparison
  site, exactly as PRs 1–4 did.
• Test code currently passes raw f32 values to Generator::new; the
  minimal change is to wrap each call with AtLeastOneHundredth::try_new(x).expect(...).

Files to modify

1. lading_payload/src/dogstatsd.rs

   • Import: add AtLeastOneHundredth next to Probability in the
     crate::common::config use list.
   • Config::unique_tag_ratio field: f32 → AtLeastOneHundredth. Update
     the rustdoc to drop the "between 0.10 and 1.0" claim (the type enforces
     [0.01, 1.0]).
   • Default for Config: replace unique_tag_ratio: 0.11 with
     unique_tag_ratio: AtLeastOneHundredth::try_new(0.11).expect("0.11 is in [0.01, 1.0]").
   • MemberGenerator::new: change the unique_tag_ratio: f32 parameter to
     AtLeastOneHundredth. The value is forwarded unchanged to
     tags::Generator::new.

2. lading_payload/src/common/tags.rs

   • Import: use crate::common::config::{AtLeastOneHundredth, ConfRange};.
   • Generator::unique_tag_probability field: f32 → AtLeastOneHundredth.
   • Generator::new signature: parameter unique_tag_probability: f32 →
     AtLeastOneHundredth.
   • Remove the now-redundant MIN_UNIQUE_TAG_RATIO..=MAX_UNIQUE_TAG_RATIO
     range check (lines 191–195). The type enforces [0.01, 1.0] at
     deserialize time.
   • The WARN check (lines 197–201) keeps its current behavior. Call
     .get() on the parameter for the range membership test:
     (MIN_UNIQUE_TAG_RATIO..=WARN_UNIQUE_TAG_RATIO).contains(&unique_tag_probability.get()).
   • Hot path at line 277:
     let should_reuse = choose_existing_prob > self.unique_tag_probability.get();
   • Update rustdoc on Generator and Generator::new to drop stale
     "between 0.10 and 1.0" prose; the type encodes the lower bound.
   • Tests in this file pass raw f32 to Generator::new. Wrap the constant
     literal 1.0 calls (lines 334, 373) and the proptest-generated
     unique_tag_ratio (line 423) with AtLeastOneHundredth::try_new(x).expect(...).

3. lading_payload/src/dogstatsd/common/tags.rs

   • Import: add AtLeastOneHundredth next to PoolKind / tags.
   • Generator::new signature: unique_tag_probability: f32 →
     AtLeastOneHundredth. Forward to inner tags::Generator::new unchanged.
   • Tests pass raw f32 to inner tags::Generator::new. Wrap the literal
     1.0 calls (lines 181, 227, 266, 351) and the proptest-generated
     unique_tag_ratio (lines 306, 400) with
     AtLeastOneHundredth::try_new(x).expect(...).

4. lading_payload/src/opentelemetry/common.rs

   • Import: use crate::common::config::{AtLeastOneHundredth, ConfRange};.
   • UNIQUE_TAG_RATIO constant: f32 = 0.75 → AtLeastOneHundredth = match AtLeastOneHundredth::try_new(0.75) { Ok(p) => p, Err(_) => panic!(...) }.
   • TagGenerator::new signature: unique_tag_probability: f32 →
     AtLeastOneHundredth. Forward to inner tags::Generator::new unchanged.

5. lading_payload/src/opentelemetry/log/templates.rs and
   lading_payload/src/opentelemetry/metric/templates.rs

   • No source changes — call sites already pass UNIQUE_TAG_RATIO, whose
     type is updated centrally.

Hot-path call site (one site for this PR)

Per the rollout plan's table, the only hot-path comparison for this field is:

• lading_payload/src/common/tags.rs:277 — change
  choose_existing_prob > self.unique_tag_probability to
  choose_existing_prob > self.unique_tag_probability.get().

.get() is const fn returning a copied f32, so the generated code is
identical and RNG sequences remain bit-exact.

Out of scope (explicitly)

• No YAML/JSON wire-format change.
• No substitution of OpenClosed01 < p with sample_bernoulli.
• No removal of MIN_UNIQUE_TAG_RATIO, WARN_UNIQUE_TAG_RATIO, or
  MAX_UNIQUE_TAG_RATIO.
• No reordering or "tidy-up" of adjacent code.

Verification

1. Type check + lints: ci/scripts/check-features.sh or
   cargo check -p lading_payload --all-features and
   cargo clippy -p lading_payload --all-features -- -D warnings.
2. Unit + proptest suite for the affected crate:
   cargo test -p lading_payload. This re-runs the proptests in
   common::tags and dogstatsd::common::tags that vary
   unique_tag_ratio over [WARN, MAX), which exercises the
   wrap-into-AtLeastOneHundredth call paths in tests.
3. Bit-exact byte output for default config: confirm by running
   cargo bench -p lading_payload --bench dogstatsd once before and once
   after the change. Expected delta is noise — no allocations added, no
   new branches, .get() inlines away. Capture the criterion summary in
   the PR description.
4. OTel benches per the rollout plan's "Efficiency verification" for PR 5:
   cargo bench -p lading_payload --bench opentelemetry_log,
   --bench opentelemetry_metric, --bench opentelemetry_traces. Same
   expectation: noise-level delta.
5. Deserialize-time rejection: a runtime sanity check via
   serde_yaml::from_str::<Config>("...unique_tag_ratio: 0.005\n...")
   should now fail with a "below lower bound 0.01" error rather than
   succeeding and then failing later in Generator::new. (Optionally
   captured as a new unit test; matches the pattern already added in PR 1
   for timestamp.probability at dogstatsd.rs:1003.)

Commit / submit

• Commit message follows the established pattern of PRs 1–4:
  feat(payload): adopt AtLeastOneHundredth for Config::unique_tag_ratio
  with a short body explaining the type change, the threaded constructors,
  the removed redundant range check, and the bit-exact .get() hot-path
  preservation. Include the Co-Authored-By trailer.
• gt create --all to branch from goxberry/probability-multivalue-pack-prob
  then gt submit to push and open the stacked PR (matches the existing
  Graphite stack already initialized with main as trunk).