[1/4] feat(power): add the DCGM power artifact data layer

[edwingao28]

What

Data layer for multinode DCGM power artifacts: the file formats, parsing, and validation that a follow-up collector runtime writes and an offline validator audits. No runtime behavior changes — nothing imports this package yet; it lands first so the collector PR stays reviewable.

Contents (src/srtctl/core/power/)

• contract.py — schema version, filenames, reason codes, sample-gap limit (MAX_SAMPLE_GAP_SECONDS), finite-number checks. Single source of truth for the on-disk contract.
• parser.py — Prometheus scrape → per-GPU power readings via prometheus_client.parser (new dependency prometheus-client>=0.20.0; parsing the exposition format by hand is strictly worse).
• samples.py — append-only CSV sample stream: writer, reader, observed-device derivation.
• topology.py — expected-device construction from the Slurm process topology; device identity keys; role/het-group resolution.
• manifest.py — manifest.json writer: producer identity, expected windows, validation rollup, atomic writes.
• windows.py — measurement-window validation: coverage against the sample stream, gap checks, running→interrupted conversion.

Evidence

The full series (this contract plus the collector and windows that follow) has been running in production CI on real multinode deployments:

• GB200 and GB300 1P1D disaggregated deployments (4 prefill + 4 decode GPUs), N=3 independent full concurrency ladders (c1–c128) per platform; 8/8 measurement windows strictly valid in every run.
• Reproducibility across the N=3 runs: J/output-token CV ≤ 1.59% (GB200), ≤ 0.67% (GB300).
• The exact tree this series is cut from also passed a GB200 canary where the manifest's stored rollup was recomputed offline from raw artifacts and agreed to float precision.

Happy to attach a sample artifact bundle (samples CSV + windows + manifest) to the PR if useful for review.

Tests

tests/test_power_artifacts.py — 45 tests over the contract: round-trips, malformed-input reasons, device-set mismatches. Window coverage/gap edge cases are exercised by the consumers' suites in [3/4] and [4/4], which drive the real validate_expected_windows (including boolean-typed concurrency/schema_version traps — bool is an int subclass and must not key as concurrency 1).

pytest tests/ passes (same 3 pre-existing environment-dependent failures as main on my machine: fingerprint CPU probe, apply_mock CPU allocation, mcp_spec explain-field).

Series

• fix(processes): guard the post-kill wait in ManagedProcess.terminate #287 fix(processes): guard the post-kill wait in ManagedProcess.terminate (independent fix, used by [3/4])
• [1/4] this PR — artifact data layer
• [2/4] [2/4] feat(telemetry): collect multinode DCGM power into strict artifacts #289 — collector runtime (writes these artifacts during sweeps)
• [3/4] [3/4] feat(power): add the measurement window contract and stage plumbing #290 — measurement window contract + stage plumbing
• [4/4] [4/4] feat(power): add the offline artifact validator and telemetry docs #291 — offline validator CLI + docs

[FrankD412]

@edwingao28 -- thanks for the PR. I'm taking a look and running initial reviews. I'm planning to have feedback for you sometime Monday or early on Tuesday; I'm feeling under the weather today. Does that timeline work for you?

This includes #288 #289 #290 #291

[edwingao28]

@edwingao28 -- thanks for the PR. I'm taking a look and running initial reviews. I'm planning to have feedback for you sometime Monday or early on Tuesday; I'm feeling under the weather today. Does that timeline work for you?

This includes #288 #289 #290 #291

Yes, that timeline works for me. Thanks for taking a look, and hope you feel better soon!

[FrankD412]

Code Review

Confirmed Bugs

windows.py (_scan) — When 3+ window files share the same (benchmark_type, concurrency) key, the first file (A) ends up in artifact_errors twice. When B arrives, both A and B are correctly flagged as MEASUREMENT_WINDOW_DUPLICATE. When C arrives, parsed[key] still holds A (the continue never updates it), so A is flagged a second time. Fix: pop A from parsed immediately on the first duplicate detection instead of deferring to the post-loop cleanup.

parser.py (parse_power_scrape) — A prometheus text-format parse error (ValueError from text_string_to_metric_families) is reported as Reason.ENDPOINT_HTTP_ERROR. That reason code means "HTTP transport failed" — operators will chase network/port configuration instead of the exporter's output format. A separate PARSE_ERROR (or similar) reason code is needed.

samples.py (_has_non_monotonic_device) — The non-monotonic check uses <= instead of <. Equal consecutive timestamps — possible at 1 Hz on a 1-second-resolution clock — satisfy the condition and produce a false TIMESTAMP_NON_MONOTONIC, marking valid artifacts as invalid.

Plausible Bugs

parser.py (parse_power_scrape) — The except ValueError catch on list(text_string_to_metric_families(text)) is too narrow. Non-ValueError exceptions from internal parser state (e.g. IndexError, KeyError) or a future prometheus_client version escape uncaught to the collector thread and crash it rather than recording a reason code and continuing.

samples.py (SampleWriter.__init__) — If writerow or flush raises after open() succeeds (e.g. disk full), self._handle is never closed. The class has no __del__, __enter__/__exit__, or except clause in __init__ to handle this. Under repeated retries the process could hit EMFILE.

contract.py (atomic_write_json) — If os.fdopen raises after mkstemp succeeds (e.g. at the EMFILE boundary), the raw fd is leaked. The except BaseException block unlinks the temp path but never calls os.close(fd).

Minor

windows.py (_check_coverage) — if device.gpu_uuids else "" in the gap key f-string is dead code. The len(device.gpu_uuids) != 1 guard above guarantees exactly one UUID at that point. Should be removed to avoid confusion if the guard is ever refactored.

manifest.py (mark_terminal) — The docstring says "Freeze lifecycle state" but mark_terminal itself has no guard against a second call. In practice this is safe because stop_and_finalize (the only caller path) has its own if self._outcome is not None: return guard — but the protection lives one level up, not in the method that claims to freeze. Worth either adding a guard to mark_terminal directly or updating the docstring to reflect where the invariant is actually enforced.

Integration Gap (series-level)

None of the four PRs in this series (#287–#291) touch the vendored benchmark_serving.py at src/srtctl/benchmarks/scripts/sa-bench/benchmark_serving.py. The MeasurementWindow stamping integration exists only on the dev branch. As-is, any power-enabled run on main after this series lands will produce MEASUREMENT_WINDOW reason codes and be publication-invalid — the production evidence (8/8 valid windows) came from that dev branch integration, not anything shipping here.

Is there a tracking issue for the stamping integration? And is there a plan to block required mode publication until it lands?

[FrankD412]

Oh -- I hadn't seen the note about vendor integration being a follow-on where eventually it seems like srt-slurm will pull from a central location (from #290). I think that means the integration gap critique is moot.

[edwingao28]

Oh -- I hadn't seen the note about vendor integration being a follow-on where eventually it seems like srt-slurm will pull from a central location (from #290). I think that means the integration gap critique is moot.

Thanks for confirming. Yes, that is the intended split: this series lands the artifact contract, collector, measurement-window plumbing, and validator, while the benchmark-side adapter will follow the migration to the centrally sourced benchmark implementation

[edwingao28]

Code Review

Confirmed Bugs

windows.py (_scan) — When 3+ window files share the same (benchmark_type, concurrency) key, the first file (A) ends up in artifact_errors twice. When B arrives, both A and B are correctly flagged as MEASUREMENT_WINDOW_DUPLICATE. When C arrives, parsed[key] still holds A (the continue never updates it), so A is flagged a second time. Fix: pop A from parsed immediately on the first duplicate detection instead of deferring to the post-loop cleanup.

parser.py (parse_power_scrape) — A prometheus text-format parse error (ValueError from text_string_to_metric_families) is reported as Reason.ENDPOINT_HTTP_ERROR. That reason code means "HTTP transport failed" — operators will chase network/port configuration instead of the exporter's output format. A separate PARSE_ERROR (or similar) reason code is needed.

samples.py (_has_non_monotonic_device) — The non-monotonic check uses <= instead of <. Equal consecutive timestamps — possible at 1 Hz on a 1-second-resolution clock — satisfy the condition and produce a false TIMESTAMP_NON_MONOTONIC, marking valid artifacts as invalid.

Plausible Bugs

parser.py (parse_power_scrape) — The except ValueError catch on list(text_string_to_metric_families(text)) is too narrow. Non-ValueError exceptions from internal parser state (e.g. IndexError, KeyError) or a future prometheus_client version escape uncaught to the collector thread and crash it rather than recording a reason code and continuing.

samples.py (SampleWriter.__init__) — If writerow or flush raises after open() succeeds (e.g. disk full), self._handle is never closed. The class has no __del__, __enter__/__exit__, or except clause in __init__ to handle this. Under repeated retries the process could hit EMFILE.

contract.py (atomic_write_json) — If os.fdopen raises after mkstemp succeeds (e.g. at the EMFILE boundary), the raw fd is leaked. The except BaseException block unlinks the temp path but never calls os.close(fd).

Minor

windows.py (_check_coverage) — if device.gpu_uuids else "" in the gap key f-string is dead code. The len(device.gpu_uuids) != 1 guard above guarantees exactly one UUID at that point. Should be removed to avoid confusion if the guard is ever refactored.

manifest.py (mark_terminal) — The docstring says "Freeze lifecycle state" but mark_terminal itself has no guard against a second call. In practice this is safe because stop_and_finalize (the only caller path) has its own if self._outcome is not None: return guard — but the protection lives one level up, not in the method that claims to freeze. Worth either adding a guard to mark_terminal directly or updating the docstring to reflect where the invariant is actually enforced.

Integration Gap (series-level)

None of the four PRs in this series (#287–#291) touch the vendored benchmark_serving.py at src/srtctl/benchmarks/scripts/sa-bench/benchmark_serving.py. The MeasurementWindow stamping integration exists only on the dev branch. As-is, any power-enabled run on main after this series lands will produce MEASUREMENT_WINDOW reason codes and be publication-invalid — the production evidence (8/8 valid windows) came from that dev branch integration, not anything shipping here.

Is there a tracking issue for the stamping integration? And is there a plan to block required mode publication until it lands?

Thanks for the detailed review @FrankD412 . I addressed all eight code findings in fca5e2b:

[FrankD412]

Code Review (second pass)

Correctness Bugs

windows.py (convert_running_windows) — _scan guards against a symlinked windows/ directory via _stays_below, which resolves the path and verifies it stays within the power artifact tree. convert_running_windows has no equivalent guard — windows_dir.is_dir() follows symlinks by default, so if windows/ is a symlink to an arbitrary directory, convert_running_windows will atomically overwrite any .json files found there. In a SLURM environment where the benchmark child runs as the same user, a malicious or misconfigured child could replace windows/ with a symlink before teardown. Fix: add the same _stays_below check at the top of convert_running_windows, mirroring _scan.

parser.py (parse_power_scrape) — saw_power_sample is set to True before the MIG instance filter, so a node where every GPU is in MIG mode never gets POWER_METRIC_MISSING:

saw_power_sample = True          # set here, before MIG check
if "GPU_I_ID" in sample.labels or "GPU_I_PROFILE" in sample.labels:
    reasons.append(Reason.MIG_INSTANCE_UNSUPPORTED)
    continue                     # sample discarded

Every scrape cycle returns readings=() with only MIG_INSTANCE_UNSUPPORTED. From the manifest's perspective this is indistinguishable from a transient case where a few MIG instances appeared — there's no signal that this node will structurally never contribute whole-GPU power readings. Fix: move saw_power_sample = True to after the MIG check, or use a separate saw_non_mig_power_sample flag, so that an all-MIG node returns both MIG_INSTANCE_UNSUPPORTED and POWER_METRIC_MISSING.

topology.py (resolve_roles, resolve_het_groups) — Both functions use len(distinct) != 1 to detect a device assigned to conflicting roles/het-groups. This correctly catches len == 2+ (genuine conflict), but also fires on len == 0 (empty assignments), which is a misconfiguration, not a conflict. Since validate_devices calls both, an ExpectedDevice with assignments=() produces both CONFLICTING_WORKER_ROLES and CONFLICTING_HET_GROUPS simultaneously — a two-code error that directs debugging toward a scheduling conflict that doesn't exist. Fix:

if len(distinct) == 0:
    return {}, (Reason.NO_DEVICE_ASSIGNMENT,)  # or equivalent
if len(distinct) != 1:
    return {}, (Reason.CONFLICTING_WORKER_ROLES,)

manifest.py (PowerManifest) — mark_terminal's double-call guard (if self.status in TERMINAL_STATUSES: raise RuntimeError) is bypassed by direct field assignment. PowerManifest is not frozen, and the orchestrator already writes manifest.status = STATUS_RUNNING directly. Any code path that directly sets manifest.status to a non-terminal value after mark_terminal has committed resets the guard, allowing a second mark_terminal call to silently overwrite stopped_at_unix and publication_valid.

Fragility

windows.py (_bracketing_sequence) — The function relies on times being sorted ascending: before[-1] is the closest sample at or before start, and after[0] is the closest at or after end — but only if the sequence is ordered. The type signature is Sequence[float] with no ordering contract. derive_observed_devices currently provides tuple(sorted(times)) which satisfies this, but there's no assertion or documented invariant. If the call site changes to preserve insertion/scrape-sequence order, gap calculations silently produce wrong results with no error.

samples.py (SampleWriter) — The file handle is opened in __init__ but the class has no __enter__/__exit__. Resource safety depends entirely on every caller explicitly calling close(). A context manager would enforce this at the type level and prevent handle leaks on early exit between construction and the finally block.

Test Coverage

tests/test_power_artifacts.py — convert_running_windows is not imported in the test file and has no happy-path test. The only related test exercises the OSError path. A regression in the conversion logic (wrong key written, wrong status checked, silent atomic-write failure) would go undetected. Per CLAUDE.md: "when we make a new significant feature change, we should always add a new test."

Simplification

windows.py — _is_strict_int (the isinstance(value, int) and not isinstance(value, bool) bool-exclusion pattern) duplicates logic already present in contract.py's is_finite_number. Should be consolidated into contract.py and imported.

windows.py — _stays_below is is_safe_relative_subpath's filesystem-resolving companion and belongs in contract.py alongside it, rather than as a private utility in windows.py. Any future module needing symlink-safe path resolution would otherwise have to either import from windows.py (wrong coupling direction) or copy the function.

[edwingao28]

Code Review (second pass)

Correctness Bugs

windows.py (convert_running_windows) — _scan guards against a symlinked windows/ directory via _stays_below, which resolves the path and verifies it stays within the power artifact tree. convert_running_windows has no equivalent guard — windows_dir.is_dir() follows symlinks by default, so if windows/ is a symlink to an arbitrary directory, convert_running_windows will atomically overwrite any .json files found there. In a SLURM environment where the benchmark child runs as the same user, a malicious or misconfigured child could replace windows/ with a symlink before teardown. Fix: add the same _stays_below check at the top of convert_running_windows, mirroring _scan.

parser.py (parse_power_scrape) — saw_power_sample is set to True before the MIG instance filter, so a node where every GPU is in MIG mode never gets POWER_METRIC_MISSING:

saw_power_sample = True          # set here, before MIG check
if "GPU_I_ID" in sample.labels or "GPU_I_PROFILE" in sample.labels:
    reasons.append(Reason.MIG_INSTANCE_UNSUPPORTED)
    continue                     # sample discarded

Every scrape cycle returns readings=() with only MIG_INSTANCE_UNSUPPORTED. From the manifest's perspective this is indistinguishable from a transient case where a few MIG instances appeared — there's no signal that this node will structurally never contribute whole-GPU power readings. Fix: move saw_power_sample = True to after the MIG check, or use a separate saw_non_mig_power_sample flag, so that an all-MIG node returns both MIG_INSTANCE_UNSUPPORTED and POWER_METRIC_MISSING.

topology.py (resolve_roles, resolve_het_groups) — Both functions use len(distinct) != 1 to detect a device assigned to conflicting roles/het-groups. This correctly catches len == 2+ (genuine conflict), but also fires on len == 0 (empty assignments), which is a misconfiguration, not a conflict. Since validate_devices calls both, an ExpectedDevice with assignments=() produces both CONFLICTING_WORKER_ROLES and CONFLICTING_HET_GROUPS simultaneously — a two-code error that directs debugging toward a scheduling conflict that doesn't exist. Fix:

if len(distinct) == 0:
    return {}, (Reason.NO_DEVICE_ASSIGNMENT,)  # or equivalent
if len(distinct) != 1:
    return {}, (Reason.CONFLICTING_WORKER_ROLES,)

manifest.py (PowerManifest) — mark_terminal's double-call guard (if self.status in TERMINAL_STATUSES: raise RuntimeError) is bypassed by direct field assignment. PowerManifest is not frozen, and the orchestrator already writes manifest.status = STATUS_RUNNING directly. Any code path that directly sets manifest.status to a non-terminal value after mark_terminal has committed resets the guard, allowing a second mark_terminal call to silently overwrite stopped_at_unix and publication_valid.

Fragility

windows.py (_bracketing_sequence) — The function relies on times being sorted ascending: before[-1] is the closest sample at or before start, and after[0] is the closest at or after end — but only if the sequence is ordered. The type signature is Sequence[float] with no ordering contract. derive_observed_devices currently provides tuple(sorted(times)) which satisfies this, but there's no assertion or documented invariant. If the call site changes to preserve insertion/scrape-sequence order, gap calculations silently produce wrong results with no error.

samples.py (SampleWriter) — The file handle is opened in __init__ but the class has no __enter__/__exit__. Resource safety depends entirely on every caller explicitly calling close(). A context manager would enforce this at the type level and prevent handle leaks on early exit between construction and the finally block.

Test Coverage

tests/test_power_artifacts.py — convert_running_windows is not imported in the test file and has no happy-path test. The only related test exercises the OSError path. A regression in the conversion logic (wrong key written, wrong status checked, silent atomic-write failure) would go undetected. Per CLAUDE.md: "when we make a new significant feature change, we should always add a new test."

Simplification

windows.py — _is_strict_int (the isinstance(value, int) and not isinstance(value, bool) bool-exclusion pattern) duplicates logic already present in contract.py's is_finite_number. Should be consolidated into contract.py and imported.

windows.py — _stays_below is is_safe_relative_subpath's filesystem-resolving companion and belongs in contract.py alongside it, rather than as a private utility in windows.py. Any future module needing symlink-safe path resolution would otherwise have to either import from windows.py (wrong coupling direction) or copy the function.

Code Review (second pass)

Correctness Bugs

windows.py (convert_running_windows) — _scan guards against a symlinked windows/ directory via _stays_below, which resolves the path and verifies it stays within the power artifact tree. convert_running_windows has no equivalent guard — windows_dir.is_dir() follows symlinks by default, so if windows/ is a symlink to an arbitrary directory, convert_running_windows will atomically overwrite any .json files found there. In a SLURM environment where the benchmark child runs as the same user, a malicious or misconfigured child could replace windows/ with a symlink before teardown. Fix: add the same _stays_below check at the top of convert_running_windows, mirroring _scan.

parser.py (parse_power_scrape) — saw_power_sample is set to True before the MIG instance filter, so a node where every GPU is in MIG mode never gets POWER_METRIC_MISSING:

saw_power_sample = True          # set here, before MIG check
if "GPU_I_ID" in sample.labels or "GPU_I_PROFILE" in sample.labels:
    reasons.append(Reason.MIG_INSTANCE_UNSUPPORTED)
    continue                     # sample discarded

Every scrape cycle returns readings=() with only MIG_INSTANCE_UNSUPPORTED. From the manifest's perspective this is indistinguishable from a transient case where a few MIG instances appeared — there's no signal that this node will structurally never contribute whole-GPU power readings. Fix: move saw_power_sample = True to after the MIG check, or use a separate saw_non_mig_power_sample flag, so that an all-MIG node returns both MIG_INSTANCE_UNSUPPORTED and POWER_METRIC_MISSING.

topology.py (resolve_roles, resolve_het_groups) — Both functions use len(distinct) != 1 to detect a device assigned to conflicting roles/het-groups. This correctly catches len == 2+ (genuine conflict), but also fires on len == 0 (empty assignments), which is a misconfiguration, not a conflict. Since validate_devices calls both, an ExpectedDevice with assignments=() produces both CONFLICTING_WORKER_ROLES and CONFLICTING_HET_GROUPS simultaneously — a two-code error that directs debugging toward a scheduling conflict that doesn't exist. Fix:

if len(distinct) == 0:
    return {}, (Reason.NO_DEVICE_ASSIGNMENT,)  # or equivalent
if len(distinct) != 1:
    return {}, (Reason.CONFLICTING_WORKER_ROLES,)

manifest.py (PowerManifest) — mark_terminal's double-call guard (if self.status in TERMINAL_STATUSES: raise RuntimeError) is bypassed by direct field assignment. PowerManifest is not frozen, and the orchestrator already writes manifest.status = STATUS_RUNNING directly. Any code path that directly sets manifest.status to a non-terminal value after mark_terminal has committed resets the guard, allowing a second mark_terminal call to silently overwrite stopped_at_unix and publication_valid.

Fragility

windows.py (_bracketing_sequence) — The function relies on times being sorted ascending: before[-1] is the closest sample at or before start, and after[0] is the closest at or after end — but only if the sequence is ordered. The type signature is Sequence[float] with no ordering contract. derive_observed_devices currently provides tuple(sorted(times)) which satisfies this, but there's no assertion or documented invariant. If the call site changes to preserve insertion/scrape-sequence order, gap calculations silently produce wrong results with no error.

samples.py (SampleWriter) — The file handle is opened in __init__ but the class has no __enter__/__exit__. Resource safety depends entirely on every caller explicitly calling close(). A context manager would enforce this at the type level and prevent handle leaks on early exit between construction and the finally block.

Test Coverage

tests/test_power_artifacts.py — convert_running_windows is not imported in the test file and has no happy-path test. The only related test exercises the OSError path. A regression in the conversion logic (wrong key written, wrong status checked, silent atomic-write failure) would go undetected. Per CLAUDE.md: "when we make a new significant feature change, we should always add a new test."

Simplification

windows.py — _is_strict_int (the isinstance(value, int) and not isinstance(value, bool) bool-exclusion pattern) duplicates logic already present in contract.py's is_finite_number. Should be consolidated into contract.py and imported.

windows.py — _stays_below is is_safe_relative_subpath's filesystem-resolving companion and belongs in contract.py alongside it, rather than as a private utility in windows.py. Any future module needing symlink-safe path resolution would otherwise have to either import from windows.py (wrong coupling direction) or copy the function.

thanks for the review @FrankD412
addressed all review and fixed in 67dc632

1. convert_running_windows symlink escape — fixed. Added root confinement and regression coverage proving external JSON is not modified.

2. MIG semantics — defined and pinned. POWER_METRIC_MISSING means the scrape emitted no DCGM_FI_DEV_POWER_USAGE. A MIG-labeled sample emitted the metric but is unsupported, so it reports only MIG_INSTANCE_UNSUPPORTED. If the session receives no usable samples for an expected GPU, topology validation reports EXPECTED_DEVICE_MISSING. The exact parser result is now pinned by a regression assertion.

3. Empty ExpectedDevice.assignments — fixed at construction. ExpectedDevice.__post_init__ now rejects the invalid state instead of adding a runtime reason code for something the production builder cannot emit.

4. Manifest terminal guard — fixed. Terminal commitment is now independent of the mutable status field, so status reassignment cannot enable a second terminal transition.

5. Unordered bracketing input — fixed. _bracketing_sequence sorts defensively, with public-path regression coverage.

6. SampleWriter context manager — no change because PowerTelemetrySession is the sole production owner and closes the writer under its lock. A lexical with block cannot span the session’s cross-thread lifecycle and would not enforce ownership.

7. Conversion happy-path coverage — fixed. convert_running_windows was already imported and exercised by the OSError test; the missing successful running → interrupted case is now covered.

8. _is_strict_int extraction — no change because it is not equivalent to is_finite_number, which accepts finite floats such as 1.5. The integer-only predicate currently has one consumer.

9. _stays_below relocation — no change because Its current callers are both in windows.py; moving it would add a public abstraction without a cross-module consumer.

Verification: 69 artifact tests passed; Ruff, format, ty, and git diff --check passed.

Please let me know if the fix looks good to you, thanks!

[codecov-commenter]

Codecov Report

❌ Patch coverage is 92.38965% with 50 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (main@9d8d92b). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
src/srtctl/core/power/windows.py	83.74%	33 Missing ⚠️
src/srtctl/core/power/samples.py	94.20%	8 Missing ⚠️
src/srtctl/core/power/contract.py	96.77%	3 Missing ⚠️
src/srtctl/core/power/manifest.py	96.25%	3 Missing ⚠️
src/srtctl/core/power/parser.py	96.92%	2 Missing ⚠️
src/srtctl/core/power/topology.py	98.71%	1 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #288   +/-   ##
=======================================
  Coverage        ?   69.97%           
=======================================
  Files           ?       76           
  Lines           ?    10068           
  Branches        ?        0           
=======================================
  Hits            ?     7045           
  Misses          ?     3023           
  Partials        ?        0           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.