Change MetricsAggregatorService to publisher to support live metrics without shmem#306

Merged

nv-alicheng merged 33 commits into

mainfrom

prototype/alicheng-zmq-metrics

May 14, 2026

Commits on May 12, 2026

feat(metrics): add registry, samplers, and snapshot wire schema

nv-alicheng
and
claude
committed
feat(metrics): add MetricsPublisher and MetricsSnapshotSubscriber

nv-alicheng
and
claude
committed
refactor(metrics): wire pub/sub into aggregator, remove KVStore + mmap

nv-alicheng
and
claude
committed
refactor(load_generator): emit ERROR before COMPLETE for failed queries

nv-alicheng
and
claude
committed
docs(agents): update AGENTS.md for metrics pub/sub refactor

nv-alicheng
and
claude
committed
fix(metrics): address P0 review-council findings

nv-alicheng
and
claude
committed
test(metrics): rewrite skipped suites on registry/snapshot fixtures

nv-alicheng
and
claude
committed
test(templates): unblock TestTemplateIntegration without HF_TOKEN

nv-alicheng
and
claude
committed
docs(agents): add reference-hygiene rules + clean up violations

nv-alicheng
and
claude
committed
refactor(metrics_table): encapsulate in-flight task access

nv-alicheng
and
claude
committed
refactor(metrics): rename refresh_hz → publish_interval_s (seconds)

nv-alicheng
and
claude
committed
fix(report): numerically stable variance for integer-aggregate series

nv-alicheng
and
claude
committed
fix(metrics): drain / shutdown correctness for cancel + SIGTERM

nv-alicheng
and
claude
committed
fix(metrics): pre-check HDR `high >= 2*low` before HdrHistogram ctor

nv-alicheng
and
claude
committed
fix(metrics): guard publisher.start against double-STARTED

nv-alicheng
and
claude
committed
fix(metrics): structured logging around aggregator subprocess crash

nv-alicheng
and
claude
committed
docs(session): document publish-order invariant for ERROR-before-COMPLETE

nv-alicheng
and
claude
committed
feat(metrics): --drain-timeout flag, default bumped to 60s

nv-alicheng
and
claude
committed
feat(metrics): add SessionState.INITIALIZE for pre-START phase

nv-alicheng
and
claude
committed
test(registry): cover SeriesSampler internal boundaries

nv-alicheng
and
claude
committed
feat(metrics): add SessionState.INTERRUPTED for signal-handler shutdown

nv-alicheng
and
claude
committed
refactor(metrics): final snapshot = JSON file; pub/sub = TUI signal

nv-alicheng
and
claude
committed
fix(metrics): duplicate STARTED → log error + preserve session_start

nv-alicheng
and
claude
committed
fix(metrics): scrub NaN/Inf to None in snapshot_to_dict; allow_nan=False

nv-alicheng
and
claude
committed
fix(metrics): SIGTERM refresh duration; SIGINT no-op

nv-alicheng
and
claude
committed
test(execute): cover _load_final_snapshot_from_disk + Report fallback

nv-alicheng
and
claude
committed
fix(metrics): doc cleanup + contract enforcement (p50, output_dir)

nv-alicheng
and
claude
committed
fix(metrics): robustness — KeyboardInterrupt, finalize, .tmp cleanup

nv-alicheng
and
claude
committed
chore(tests): rename tests/datasets/ → tests/assets/datasets/

nv-alicheng
and
claude
committed
test(integration): use local char-tokenizer fixture, drop HF Hub dependency

nv-alicheng
and
claude
committed
test(integration): bump worker_initialization_timeout to 120s in CI

nv-alicheng
and
claude
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Change MetricsAggregatorService to publisher to support live metrics without shmem#306

Change MetricsAggregatorService to publisher to support live metrics without shmem#306
nv-alicheng merged 33 commits into
mainfrom
prototype/alicheng-zmq-metrics

Commits on May 12, 2026

feat(metrics): add registry, samplers, and snapshot wire schema

feat(metrics): add MetricsPublisher and MetricsSnapshotSubscriber

refactor(metrics): wire pub/sub into aggregator, remove KVStore + mmap

refactor(load_generator): emit ERROR before COMPLETE for failed queries

docs(agents): update AGENTS.md for metrics pub/sub refactor

fix(metrics): address P0 review-council findings

test(metrics): rewrite skipped suites on registry/snapshot fixtures

test(templates): unblock TestTemplateIntegration without HF_TOKEN

docs(agents): add reference-hygiene rules + clean up violations

refactor(metrics_table): encapsulate in-flight task access

refactor(metrics): rename refresh_hz → publish_interval_s (seconds)

fix(report): numerically stable variance for integer-aggregate series

fix(metrics): drain / shutdown correctness for cancel + SIGTERM

fix(metrics): pre-check HDR `high >= 2*low` before HdrHistogram ctor

fix(metrics): guard publisher.start against double-STARTED

fix(metrics): structured logging around aggregator subprocess crash

docs(session): document publish-order invariant for ERROR-before-COMPLETE

feat(metrics): --drain-timeout flag, default bumped to 60s

feat(metrics): add SessionState.INITIALIZE for pre-START phase

test(registry): cover SeriesSampler internal boundaries

feat(metrics): add SessionState.INTERRUPTED for signal-handler shutdown

refactor(metrics): final snapshot = JSON file; pub/sub = TUI signal

fix(metrics): duplicate STARTED → log error + preserve session_start

fix(metrics): scrub NaN/Inf to None in snapshot_to_dict; allow_nan=False

fix(metrics): SIGTERM refresh duration; SIGINT no-op

test(execute): cover _load_final_snapshot_from_disk + Report fallback

fix(metrics): doc cleanup + contract enforcement (p50, output_dir)

fix(metrics): robustness — KeyboardInterrupt, finalize, .tmp cleanup

chore(tests): rename tests/datasets/ → tests/assets/datasets/

test(integration): use local char-tokenizer fixture, drop HF Hub dependency

test(integration): bump worker_initialization_timeout to 120s in CI

Commits on May 13, 2026

fix(metrics): drain pending count, SIGTERM task GC, NaN display

style(tests): hoist lazy imports to top of test_report_builder