Repository navigation
feat(telemetry): report concurrency limit state (limit, active, queued, timeouts) - #5280
Merged
Merged
Conversation
Contributor
Author
|
/ok to test 245a8ce |
marcromeyn
enabled auto-merge
October 29, 2026 10:23
svcnvidia-nemo-ci
added
area:core
Contributor
|
🌿 Preview your docs: https://nvidia-preview-pr-5280-20e449f9c8e2.docs.buildwithfern.com/nemo/gym |
svcnvidia-nemo-ci
left a comment
svcnvidia-nemo-ci
left a comment
Contributor
There was a problem hiding this comment.
SHIP
LGTM — no concerns found.
Reviewed all 11 changed files and directly affected paths, including semaphore handoff and cancellation, admission-timeout handling, owner cleanup, metric aggregation and observer lifecycle, and documentation consistency. The double-release handoff regression is covered by the current guard and test.
Static review only; tests were not executed in this read-only environment.
Note: formal approval is not available for this request, so this review is posted as a comment only.
Review generated by codex. This was a full review. Request a full review with /review reset=true.
github-actions
Bot
added
the
sla:review-overdue
marcromeyn
force-pushed
the
romeyn/concurrency-state-main
branch
from
November 2, 2026 13:52
245a8ce to
7bc0817
Compare
Contributor
Author
|
/ok to test 7bc0817 |
marcromeyn
force-pushed
the
romeyn/concurrency-state-main
branch
from
November 3, 2026 10:21
7bc0817 to
044d19a
Compare
Contributor
Author
|
/ok to test 044d19a |
…d, timeouts) A queue-wait duration shows that work waited, not why. A telemetry dashboard needs the state of each concurrency limit, "limit 100, active 100, queued 75", to show that the limit is full and work is piling up behind it. Add MeteredSemaphore, an asyncio.BoundedSemaphore that counts its slots in use, its waiters, and its failed admissions with plain integers. The process observer exports them through observable instruments read at export: gym.concurrency.limit, .active, .available and .queue_depth, plus gym.concurrency.admission_timeouts_total and .admission_cancelled_total. Points carry only bounded attributes: component, limit name and, for servers, the server name. A component without a configured limit reports no series instead of a limit of 0, and the timeout counter exists only for limits that can time out. The admission timeout moves into the semaphore so a timeout is counted as a timeout rather than as a cancellation, as asyncio.wait_for would make it look. Wire it into the limits that exist today with the clearest owner: the rollout driver (num_samples_in_parallel), openai_model and servers built on it (max_concurrent_requests, upstream_pool_timeout_seconds), the environment-server episode admission (max_concurrent_episodes, queue_timeout_seconds) and math_with_judge's library verifier. ``` ``` driver [rollout_driver, rollouts] limit=1 active=1 available=0 queue_depth=2 driver [rollout_driver, rollouts] limit=1 active=1 available=0 queue_depth=1 driver [rollout_driver, rollouts] limit=1 active=1 available=0 queue_depth=0 driver [rollout_driver, rollouts] limit=1 active=0 available=1 queue_depth=0 policy_model [model_server, upstream_requests, policy_model] limit=1 active=1 queue_depth=0 admission_timeouts_total=0 ``` ``` driver [rollout_driver, rollouts] limit=3 active=3 available=0 queue_depth=0 policy_model [model_server, upstream_requests, policy_model] limit=1 active=1 available=0 queue_depth=2 policy_model [model_server, upstream_requests, policy_model] limit=1 active=1 available=0 queue_depth=1 policy_model [model_server, upstream_requests, policy_model] limit=1 active=0 available=1 queue_depth=0 ``` ``` policy_model [model_server, upstream_requests, policy_model] limit=1 active=1 queue_depth=2 admission_timeouts_total=0 policy_model [model_server, upstream_requests, policy_model] limit=1 active=0 queue_depth=0 admission_timeouts_total=2 Signed-off-by: Marc Romeyn <mromeijn@nvidia.com>
marcromeyn
force-pushed
the
romeyn/concurrency-state-main
branch
from
November 4, 2026 11:05
044d19a to
20e449f
Compare
Contributor
Author
|
/ok to test 20e449f |
iguehring-nvidia
approved these changes
Nov 4, 2026
github-actions
Bot
removed
the
sla:review-overdue
Contributor
|
🔄 Merge queue validation started! You can track the progress here: https://github.com/NVIDIA-NeMo/Gym/actions/runs/41876809199 |
This branch was previously deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What and why
A queue-wait duration shows that work waited, not why. To show "limit 100, active 100, queued 75", the telemetry dashboards need the state of each concurrency limit.
This PR adds
MeteredSemaphore(nemo_gym/telemetry/concurrency.py), a drop-inasyncio.BoundedSemaphorethat counts slots in use, waiters, and failed admissions with plain integers. The process observer exports them as observable instruments, read at export, so acquiring a slot makes no OpenTelemetry call:gym.concurrency.limit/.active/.availablegym.concurrency.queue_depthgym.concurrency.admission_timeouts_totalgym.concurrency.admission_cancelled_totalAttributes are bounded: component,
limit_name(one server can own several limits), and the server name. Missing stays missing: a component without a limit reports no series, not a limit of 0. The admission timeout moves into the semaphore, so a timeout is counted as a timeout, not as the cancellationasyncio.wait_forwould make it look like.Wired into the limits with the clearest owner:
limit_namerolloutsnum_samples_in_parallelopenai_model(andlitellm_model)upstream_requestsmax_concurrent_requestsupstream_pool_timeout_secondsBaseEnvironmentServerepisodesmax_concurrent_episodesqueue_timeout_secondsmath_with_judgelibrary_verifierlibrary_verifier_max_concurrencyThe metrics page gains a "Concurrency Limit Metrics" section. Other agent, model, resources and sandbox limits follow as one constructor change each.
Relationship to #2935: this answers the same dashboard question and reuses its instrument names and module path, so its panels keep working. It is built on
mainbecause #2935's branch has no common ancestor with it, sets synchronous gauges on every acquire, and would countwait_fortimeouts as cancellations. #2935's queue-wait histogram fits as one call inMeteredSemaphore.acquireonce it is rebased.Validation
telemetryextra: telemetry, environment-server,openai_model,litellm_model, rollout collection, reverification and docs-link tests: 1494 passed.release()before that waiter resumes would pass the base class's bound check;release()raises when no slot is held, and its regression test fails without that check.pre-commiton the changed files;math_with_judgeserver tests (130 passed, before the rebase, which only touched an import line).Rollout evidence
gym env startwithopenai_model(gpt-6-luna), console metrics exporter andmax_concurrent_requests=1, thengym eval run --no-serve --limit 3 --concurrency 3: the model limit read limit 1, active 1, queue depth 2 → 1 → 0, and the driver limit active 3 draining to 0; 3/3 rollouts. Before the rebase, a run withupstream_pool_timeout_seconds=0.5counted the two queued calls as 2 admission timeouts.Compatibility
No configuration changes and new metrics only. The admission timeouts raise the same
TimeoutErrorand return the same failure responses. Releasing a slot that is not held now raisesValueError; no call site does that.Linear: FEP-1844 (part of FEP-1829 / GYM-1170).