feat(openfeature): parse observeFullEvaluationData and hash targeting_key in flagevaluations events

[vjfridge]

Motivation

A server SDK upgrade must not silently ship subject PII (emails, account IDs, evaluation context) to the flagevaluations track. The product metric — unique-subject counts per (flag, allocation) — needs a stable per-subject identifier, not the raw subject. This PR sends a one-way SHA-256 fingerprint by default and makes full-fidelity collection an explicit per-environment opt-in.

Tracked by FFL-2962, under the server-SDK fan-out FFL-2784. The Java pilot is dd-trace-java#12042. The base track for Go was #4886. Cross-SDK contract tests were merged in system-tests#7316.

What does this PR do?

This PR makes the no-PII path the default for server-side flagevaluation events. A new UFC boolean observeFullEvaluationData selects the path.

observeFullEvaluationData	targeting_key on the wire	context.evaluation on the wire
true	raw, verbatim	included, raw
false or absent (default)	sha256_ + 64-char lowercase hex (71 chars total)	omitted

User impact:

• Upgraded SDKs stop shipping raw subject PII to the flagevaluations track unless the environment opts in.
• Unique-subject counts per (flag, allocation) stay accurate. Every SDK produces a byte-identical digest for the same subject, so hashed values join across languages.
• Any dashboard or query that reads targeting_key as the raw subject must accept the sha256_ prefix on the default path.
• The DD_FLAGGING_EVALUATION_COUNTS_ENABLED kill switch is unchanged. It still disables the track entirely.

Key design points:

• UFC read. observeFullEvaluationData sits at the root of the UFC, a sibling of environment. The custom UnmarshalJSON copies only the fields its shadow struct lists, so the field is added to both the struct and the shadow struct. Absent and explicit null both fail closed to false.
• Consent lifecycle. The evaluator snapshots consent from the exact configuration it evaluated against and stamps the value on evaluation metadata under the cross-SDK key observe_full_evaluation_data. Nothing downstream of the evaluator reads live configuration, so a Remote Config update between evaluation and flush cannot retroactively apply another environment's policy. The Java pilot hit this bug. A regression test guards both directions.
• Hashing. Unsalted SHA-256 over the raw UTF-8 bytes as received. No trimming, no case folding, no Unicode normalization. The hash runs once per aggregation bucket at flush cadence, not per evaluation.
• Aggregation. Consent is part of the full-tier bucket key, so mixed-consent evaluations never merge. The bucket entry also AND-folds consent as defense in depth. When consent is off, the context is dropped from the bucket key and never captured — the key must carry only dimensions that survive serialization.

No new configuration keys, no dependency changes, no generated files affected.

QA

L1 — Unit tests: done.

L2 — Dogfooding: done. Ran this branch against ffe-dogfooding with the staging org (dd.datad0g.com). Tracer built as v2.11.0-dev.1.

observeFullEvaluationData	targeting_key (wire)	context.evaluation	mock-intake pii_class	evaluations observed
true (default staging state)	raw (user-xxx)	present, raw	full	~100+
false (flipped for ~2 min)	sha256_<64 hex> (71 chars)	absent	protected	165 protected vs 5 full

The 5 "full" entries under the consent-off flip are the last flush cycle before Remote Config delivered the new UFC. Every event after the RC refresh is protected. The staging environment was restored to true after verification.

L3 — System-tests: done locally, still owed in CI. Ran the FEATURE_FLAGGING_AND_EXPERIMENTATION scenario from DataDog/system-tests locally against this branch (golang@2.11.0-dev.1, weblog variant net-http).
Results: 26 passed, 2 skipped, 0 failed.

• tests/ffe/test_flag_eval_evp.py: 11/11 passed, including the three tests unlocked by this PR and Test_FFE_EVP_Flagevaluation_Degradation (previously flaky, now clean).
• tests/ffe/test_flag_eval_metrics.py: 15 passed, 2 skipped (pre-existing irrelevant markers, unrelated to this PR).

A companion system-tests PR must remove the four manifest lines above to enable these tests in CI for Go.

How to reproduce the L2 verification yourself

Run the tracer against ffe-dogfooding with the staging org:

DD_TRACE_GO_PATH=/path/to/dd-trace-go dd-auth --domain dd.datad0g.com -- env DD_SITE=datad0g.com \
  docker compose -f docker-compose.yml -f local/docker-compose.go.yml up -d --build
# wait ~60s for first flush
curl -s "http://localhost:8090/logs?service=ffe-dogfooding-go&limit=1" | jq .
# open http://localhost:8080 for the dashboard

Flip consent (requires app and API keys with feature_flag_environment_config_write):

# fetch current, then PUT with observe_full_evaluation_data flipped
curl -s -H "DD-API-KEY: $DD_API_KEY" -H "DD-APPLICATION-KEY: $DD_APP_KEY" \
  "https://api.datad0g.com/api/v2/feature-flags/environments/<env-id>"

The flip touches the shared Staging environment on dd.datad0g.com. Anything else running under DD_ENV=staging picks up the change too. Flip back promptly.

Reviewer's Checklist

• Changed code has unit tests for its functionality at or near 100% coverage.
• System-Tests covering this feature have been added and enabled with the va.b.c-dev version tag. Tests already exist from system-tests#7316. The manifest activation is a companion PR. See "QA" above.
• There is a benchmark for any new code, or changes to existing code. Hashing runs once per bucket at flush cadence, off the evaluation hot path. The existing BenchmarkFlagEvaluationOTelPlusEVPParallel still covers the aggregation path.
• If this interacts with the agent in a new way, a system test has been added. No new agent interaction. Same EVP endpoint and payload shape.
• New code is free of linting errors. Run make lint locally.
• New code does not break existing tests. Run make test locally.
• Add an appropriate team label so this PR goes to the right release notes.
• All generated files are up to date. Run make generate locally. No generated files or configuration keys touched.
• Non-trivial go.mod changes: none. crypto/sha256 and encoding/hex are stdlib.

🤖 Generated with Claude Code

[datadog-prod-us1-6]

🎉 All green!

🧪 All tests passed
❄️ No new flaky tests detected

🎯 Code Coverage (details)
• Patch Coverage: 66.57%
• Overall Coverage: 64.52% (-4.04%)

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: fe6ad42 | Docs | Datadog PR Page | Give us feedback!}

[vjfridge]

System-tests results against this PR

Ran the FEATURE_FLAGGING_AND_EXPERIMENTATION scenario from DataDog/system-tests locally against this branch (golang@2.11.0-dev.1, weblog variant net-http).

To exercise the new behavior end-to-end, I locally removed the following markers from manifests/golang.yml:

• tests/ffe/test_flag_eval_evp.py::Test_FFE_EVP_Flagevaluation_ObserveFullData_Absent_Hashed: missing_feature (FFL-2784)
• tests/ffe/test_flag_eval_evp.py::Test_FFE_EVP_Flagevaluation_ObserveFullData_False_Hashed: missing_feature (FFL-2784)
• tests/ffe/test_flag_eval_evp.py::Test_FFE_EVP_Flagevaluation_ObserveFullData_True_Unhashed: missing_feature (FFL-2784)
• tests/ffe/test_flag_eval_evp.py::Test_FFE_EVP_Flagevaluation_Degradation::test_ffe_evp_flagevaluation_degradation: flaky (FFL-2676)

Results: 26 passed, 2 skipped, 0 failed

tests/ffe/test_flag_eval_evp.py: 11/11 passed, including the three tests unlocked by this PR:

• ✅ Test_FFE_EVP_Flagevaluation_ObserveFullData_Absent_Hashed
• ✅ Test_FFE_EVP_Flagevaluation_ObserveFullData_False_Hashed
• ✅ Test_FFE_EVP_Flagevaluation_ObserveFullData_True_Unhashed
• ✅ Test_FFE_EVP_Flagevaluation_Degradation (previously marked flaky — passed cleanly)
• Plus: Basic, Count, Context_Bounds, Runtime_Default, Load_Aggregation, Burst_Aggregation, High_Cardinality_Aggregation

tests/ffe/test_flag_eval_metrics.py: 15 passed, 2 skipped

• The 2 skips are pre-existing irrelevant markers unrelated to this PR (Invalid_Regex — Go validates at config load time; Nested_Attributes_Ignored).

Follow-up

A companion system-tests PR should remove the four manifest lines above to enable these tests in CI for Go.

[pr-commenter]

Benchmarks

Benchmark execution time: 2026-08-18 00:24:06

Comparing candidate commit fe6ad42 in PR branch vickie/ffl-2962-protecting-pii-in-flagevaluations-track with baseline commit 85a3c43 in branch main.

Found 0 performance improvements and 0 performance regressions! Performance is the same for 335 metrics, 0 unstable metrics, 1 flaky benchmarks without significant changes.

Explanation

This is an A/B test comparing a candidate commit's performance against that of a baseline commit. Performance changes are noted in the tables below as:

• 🟩 = significantly better candidate vs. baseline
• 🟥 = significantly worse candidate vs. baseline

We compute a confidence interval (CI) over the relative difference of means between metrics from the candidate and baseline commits, considering the baseline as the reference.

If the CI is entirely outside the configured SIGNIFICANT_IMPACT_THRESHOLD (or the deprecated UNCONFIDENCE_THRESHOLD), the change is considered significant.

Feel free to reach out to #apm-benchmarking-platform on Slack if you have any questions.

More details about the CI and significant changes

You can imagine this CI as a range of values that is likely to contain the true difference of means between the candidate and baseline commits.

CIs of the difference of means are often centered around 0%, because often changes are not that big:

---------------------------------(------|---^--------)-------------------------------->
                              -0.6%    0%  0.3%     +1.2%
                                 |          |        |
         lower bound of the CI --'          |        |
sample mean (center of the CI) -------------'        |
         upper bound of the CI ----------------------'

As described above, a change is considered significant if the CI is entirely outside the configured SIGNIFICANT_IMPACT_THRESHOLD (or the deprecated UNCONFIDENCE_THRESHOLD).

For instance, for an execution time metric, this confidence interval indicates a significantly worse performance:

----------------------------------------|---------|---(---------^---------)---------->
                                       0%        1%  1.3%      2.2%      3.1%
                                                  |   |         |         |
       significant impact threshold --------------'   |         |         |
                      lower bound of CI --------------'         |         |
       sample mean (center of the CI) --------------------------'         |
                      upper bound of CI ----------------------------------'

Known flaky benchmarks

These benchmarks are marked as flaky and will not trigger a failure. Modify FLAKY_BENCHMARKS_REGEX to control which benchmarks are marked as flaky.

Known flaky benchmarks without significant changes:

• scenario:BenchmarkOTLPTraceWriterFlush

[vjfridge]

L2 dogfooding — both consent paths verified end-to-end ✅

Ran this branch against ffe-dogfooding with the staging org (dd.datad0g.com). Tracer built as v2.11.0-dev.1.

observeFullEvaluationData	targeting_key (wire)	context.evaluation	mock-intake pii_class	evals observed
true (default staging state)	raw (user-xxx)	present, raw	full	~100+
false (flipped for ~2 min)	sha256_<64 hex> (71 chars)	absent	protected	165 protected vs 5 full

The 5 "full" entries under the consent-off flip are the last flush cycle that occurred before RC delivered the new UFC — everything after the RC refresh is protected. Restored the environment to true after verification.

Behaviors observed exactly matching the PR's contract table:

• Root-level observeFullEvaluationData is picked up by the shadow-struct unmarshaller
• Consent snapshot from the config-under-evaluation is preserved through flush (no config-update race)
• Bucket keys carry consent — mixed-consent evaluations don't merge
• Unsalted SHA-256 over raw UTF-8 targeting_key bytes; no trim/case-fold/normalization
• context.evaluation is entirely omitted from the wire payload when consent is off
• The 5→165 transition happened cleanly at the RC refresh boundary; no partial or corrupted payloads observed in between

Reproduce

DD_TRACE_GO_PATH=/path/to/dd-trace-go dd-auth --domain dd.datad0g.com -- env DD_SITE=datad0g.com \
  docker compose -f docker-compose.yml -f local/docker-compose.go.yml up -d --build
# wait ~60s for first flush
curl -s "http://localhost:8090/logs?service=ffe-dogfooding-go&limit=1" | jq .
# open http://localhost:8080 for the dashboard

To flip consent (requires app+api keys with feature_flag_environment_config_write):

# fetch current, then PUT with observe_full_evaluation_data flipped
curl -s -H "DD-API-KEY: $DD_API_KEY" -H "DD-APPLICATION-KEY: $DD_APP_KEY" \
  "https://api.datad0g.com/api/v2/feature-flags/environments/<env-id>"

Note: this touches the shared "Staging" environment on dd.datad0g.com — anything else running under DD_ENV=staging will pick up the change too. Flip back promptly.

L3 system-tests still owed per the PR's Validation status section. L1 already covers the hashed vector exhaustively (jane.doe@datadoghq.com → sha256_b4698f9b…c99e51b).

[vjfridge]

FFE Dogfooding verification — PII protection for flagevaluations events

I ran the FFE dogfooding stack against staging with the local dd-trace-go from this PR branch to verify the observeFullEvaluationData / targeting_key hashing behavior end-to-end. Everything works as designed — no failures for the Go SDK.

How I ran it

DD_TRACE_GO_PATH=~/dd/dd-trace-go \
DD_TRACE_JAVA_PATH=~/dd/dd-trace-java \
dd-auth --domain dd.datad0g.com -- \
  docker compose -f docker-compose.yml \
    -f local/docker-compose.go.yml \
    -f local/docker-compose.java.yml up --build

This builds the Go SDK against the local dd-trace-go (v2.11.0-dev.1) via go mod edit -replace, points the Datadog Agent at staging via dd-auth, and runs the evaluator dashboard at http://localhost:8080/dashboard. The mock-intake captures the EVP flagevaluation events the agent forwards so the exact wire payload can be inspected.

What I verified

1. EVP flagevaluation events carry the expected tags. Queried GET /logs/api/v2/flagevaluation?service=ffe-dogfooding-go on the mock-intake — every event has timestamp, flag.key, first_evaluation, last_evaluation, evaluation_count, targeting_key, variant.key, allocation.key, and (when enabled) context.evaluation. The staging timeseries query (context.dd.service:ffe* grouped by service) returns data for ffe-dogfooding-go, confirming delivery to the staging backend.

2. observeFullEvaluationData toggles PII protection correctly. I flipped the environment-level setting on staging (PUT /api/v2/feature-flags/environments/<env-id>) and observed the Go SDK's flagevaluation events after RC propagation (~60s; agent polls every 5s):

observeFullEvaluationData	targeting_key	context.evaluation
true (original)	raw, e.g. user-28zvjdbsi7u1	present ({country, plan, version})
false (flipped)	hashed → sha256_415a…bf73 (64 hex chars)	omitted (0/60 events)
true (restored)	raw	present (50/50)

All 60 post-propagation events with the false setting had hashed targeting_key (prefixed sha256_ + unsalted SHA-256) and zero carried context.evaluation. Restoring to true brought back raw keys + eval context, with the expected flush-straddling mix during the transition.

3. The hashing matches the spec. openfeature/flageval_pii.go does unsalted SHA-256 over raw UTF-8 bytes (no trimming/folding/normalization → byte-identical across SDKs), uses the sha256_ prefix, and returns "" for an empty targeting key so unique-subject counts aren't corrupted by an invented pseudo-subject. The privacy-preserving default (false) wins on any parse ambiguity.

Result

• ✅ Go SDK emits well-formed flagevaluation EVP events with the complete expected tag set
• ✅ End-to-end delivery to staging confirmed (timeseries query returns ffe-dogfooding-go data)
• ✅ observeFullEvaluationData=false → targeting_key hashed (sha256_ + 64 hex), context.evaluation omitted
• ✅ observeFullEvaluationData=true → raw targeting_key, context.evaluation present
• ✅ Bidirectional RC toggle works; flush-straddling transition behaves as expected
• ❌ Nothing failing for the Go SDK. All 6 test flags evaluate correctly across both PII modes.

Note: the staging environment's observe_full_evaluation_data was temporarily set to false for this test and restored to its original value (true) afterward.

[vjfridge]

System-tests validation results

Validated this branch (vickie/ffl-2962-protecting-pii-in-flagevaluations-track, 28c99401e, v2.11.0-dev.1) against the EVP flagevaluation system-tests, with the four previously-disabled EVP test classes enabled for Go on system-tests branch vickie/enable-ffe-evp-go-validation (commit 07b173dba).

End-to-end — FEATURE_FLAGGING_AND_EXPERIMENTATION scenario, tests/ffe/test_flag_eval_evp.py

11 passed in 81.07s

All 11 EVP tests pass, including the 4 newly-enabled ones:

• Test_FFE_EVP_Flagevaluation_Basic
• Test_FFE_EVP_Flagevaluation_Count
• Test_FFE_EVP_Flagevaluation_Context_Bounds
• Test_FFE_EVP_Flagevaluation_Runtime_Default
• Test_FFE_EVP_Flagevaluation_Load_Aggregation
• Test_FFE_EVP_Flagevaluation_Burst_Aggregation
• Test_FFE_EVP_Flagevaluation_High_Cardinality_Aggregation
• Test_FFE_EVP_Flagevaluation_Degradation (was flaky (FFL-2676))
• Test_FFE_EVP_Flagevaluation_ObserveFullData_Absent_Hashed (was missing_feature (FFL-2784))
• Test_FFE_EVP_Flagevaluation_ObserveFullData_False_Hashed (was missing_feature (FFL-2784))
• Test_FFE_EVP_Flagevaluation_ObserveFullData_True_Unhashed (was missing_feature (FFL-2784))

Parametric — PARAMETRIC scenario, tests/parametric/test_ffe/

24 passed, 34 xfailed, 13 xpassed, 0 failed (19.40s)

No failures. The 13 XPASS are missing_feature-marked tests that now pass (easy wins in test_configuration_sources.py and test_span_enrichment.py); those manifest declarations are outside the EVP scope and were left untouched.

SDK changes required

None. The branch satisfies the full EVP flagevaluation contract:

• targeting_key is SHA-256 hashed (with the sha256_ prefix) when observeFullEvaluationData is absent/false, and emitted raw when true.
• context.evaluation is omitted unless observeFullEvaluationData=true, and no raw PII leaks anywhere in the event.
• Aggregation, count, context bounding, runtime-default, and degradation paths all behave correctly.

system-tests change

manifests/golang.yml — removed the 4 per-class disabled declarations so the EVP PII-protection classes inherit the file-level v2.10.0-dev enablement:

   tests/ffe/test_flag_eval_evp.py: v2.10.0-dev
-  tests/ffe/test_flag_eval_evp.py::Test_FFE_EVP_Flagevaluation_Degradation::test_ffe_evp_flagevaluation_degradation: flaky (FFL-2676)
-  tests/ffe/test_flag_eval_evp.py::Test_FFE_EVP_Flagevaluation_ObserveFullData_Absent_Hashed: missing_feature (FFL-2784)
-  tests/ffe/test_flag_eval_evp.py::Test_FFE_EVP_Flagevaluation_ObserveFullData_False_Hashed: missing_feature (FFL-2784)
-  tests/ffe/test_flag_eval_evp.py::Test_FFE_EVP_Flagevaluation_ObserveFullData_True_Unhashed: missing_feature (FFL-2784)

Branch: vickie/enable-ffe-evp-go-validation on DataDog/system-tests.

⚠️ Note for re-runners: the Go build uses a local binaries/dd-trace-go folder (gitignored) which takes precedence over golang-load-from-go-get. I replaced the stale copy that was there with a fresh checkout of this branch before building.

[leoromanovsky]

Thanks. This fixes the unbounded recursive value traversal. The new depth, list, field, and cycle limits are useful. The benchmarks also cover the correct adversarial inputs.

Keeping the flag evaluation "hot path" fast even when customers mis-use is it still an important goal.

One part of my feedback remains. Top-level and nested maps still collect and sort every key before applying the 256-property limit. Evaluation-thread work therefore remains proportional to the supplied map width.

The current benchmarks show this clearly:

• 256 flat properties: 40 µs and 27 KB allocated
• 5,000 flat properties: 500 µs and 105 KB allocated
• 10,000 map properties: 837 µs and 205 KB allocated

Therefore, the claim that work is "proportional to what is kept" is not correct for maps.

Please bound map processing on the evaluation thread. For example, stop or omit the map when its property count exceeds the limit. If this remains follow-up work, please remove the bounded-cost claims and track the remaining issue explicitly.

[vjfridge]

@leoromanovsky Thanks — your remaining feedback is addressed in fe6ad42 (just pushed). Top-level and nested maps no longer collect and sort every key before applying the 256-property limit.

The change

A map (top-level or nested) whose key count exceeds the corresponding cap is now omitted entirely via an O(1) len() check, rather than collected, sorted, and trimmed:

// flattenAndPruneContext (top-level)
if len(attrs) > maxContextFields {
    return nil, []string{truncReasonMaxContextFields}
}

// copyPrunedMap (nested) + reflect-map case
if len(m) > maxStructureProperties {
    mask.add(bitMaxStructureProperties)
    return
}

The collect → sort → walk now only runs on maps already within the cap (≤256 keys), so the evaluation-thread cost is bounded by the kept set, not the supplied width. The max_structure_properties / max_context_fields reasons still fire for telemetry, and determinism is preserved (deterministically empty → same aggregation bucket).

Trade-off: there's a cliff at the cap — a 257-property structure is dropped wholesale where a 256-property structure is kept in full. Real contexts stay well under the caps, and the doc comments now state this explicitly instead of claiming "proportional to what is kept."

I also caught a related O(input) cost in the []byte branch of copyPrunedValue: the length is now checked before the string(v) conversion, so an oversized []byte is dropped without copying it.

Benchmark results

BenchmarkFlattenAndPruneContext re-run with -count=3 (Apple M4 Max). Your before numbers are from your run, so absolute values aren't directly comparable across machines, but the over-cap cases collapse by ~4,000–8,000×:

sub-benchmark	before (your run)	after (this branch)
flat_at_cap_256	40 µs / 27 KB	~21.6 µs / 27.5 KB
flat_over_cap_5000	500 µs / 105 KB	~12 ns / 16 B
wide_structure_10000	837 µs / 205 KB	~101 ns / 64 B
wide_list_10000	—	~20.1 µs / 40.7 KB
deep_chain_100	—	~636 ns / 520 B
cyclic	—	~326 ns / 456 B
oversized_byteslice_1MB (new)	—	~175 ns / 368 B (no 1 MB copy)

In-bounds cases are unchanged; the sort now only ever runs over ≤256 keys. The over-cap map path is now O(1) — just the len() check.

[vjfridge]

Re-tested with ffe-dogfooding and system-tests again locally and it's all looking gtg!