Skip to content

heavy full-suite: 77 real failures once the runner no longer OOMs #19

Closed

Description

@echobt

Summary

heavy / full-suite now runs to completion on blacksmith-16vcpu-ubuntu-2404 (PR #18, dispatch run 39420692371, 55 min 27 s) instead of being evicted at ~3 min. It fails on real tests: 77 failed, 2705 passed, 17 skipped, 17 errors in 3293 s.

Evidence

  • Before the eviction era, run 38997624076 (2026-10-17, hosted, ran to the end) failed with 75 failed / 17 errors on the same groups. These are pre-existing.
  • New since then: tests/test_publish_exclusions.py::test_current_repo_tree_scan_clean and ::test_bare_selected_copy_has_publisher_runtime_controls. They also fail locally on main 4a7166d (2 failed, 53 passed):
    SECRET tests/beacon/test_push.py:282 entropy H=4.42
    SECRET tests/datasets/test_pin.py:19 credential value under TOKEN
    
    The scanner flags test fixtures as secrets, so it is a test/scanner issue, not a leaked secret (to be confirmed by reading those lines).

Failure groups (count, test)

13 ERROR tests/challenge/test_authority_snapshot_v2.py::test_modified_package_rejects_before_restore
1 ERROR tests/challenge/test_authority_snapshot_v2.py::test_original_export_validates_and_bootstraps_actual_service
3 ERROR tests/challenge/test_carry_genesis_v2.py::test_trial_stages_pinned_complete_carry_genesis
3 FAILED tests/auditor/test_island_bisect_v2.py::test_honest_traces_when_real_island_hooks
3 FAILED tests/auditor/test_island_replay_v2.py::test_full_carry_and_window_equality
1 FAILED tests/auditor/test_island_replay_v2.py::test_verified_anchor_restart_corruption_and_od_warm_start
1 FAILED tests/challenge/test_capacity_outer_v2.py::test_actual_bounded_genesis_preparation_and_duplicate_debit
1 FAILED tests/challenge/test_capacity_outer_v2.py::test_actual_bounded_outer_child_matches_original_signed_replay
1 FAILED tests/challenge/test_service_capacity_v2.py::test_real_capacity_launch_failure_keeps_debit_and_prevents_restart_rerun
6 FAILED tests/layout/test_od_island_v2.py::test_od_bf16_same_layout_cpu
18 FAILED tests/layout/test_od_island_v2.py::test_od_objectives_layout_reference_and_hooks
4 FAILED tests/miner/test_capacity_runtime_v2.py::test_actual_small_kernel_attempt_and_cleanup
1 FAILED tests/miner/test_capacity_runtime_v2.py::test_real_runner_parent_death_retains_absolute_manager_timer
1 FAILED tests/miner/test_island_launch_v2.py::test_capacity_branch_uses_real_bounded_failure_before_publication
3 FAILED tests/miner/test_island_launch_v2.py::test_real_od_two_rank_worker_paths
1 FAILED tests/models/test_opendecision.py::test_check_manifest_param_count
1 FAILED tests/models/test_opendecision.py::test_eval_holdout_and_eval_cli
1 FAILED tests/models/test_opendecision.py::test_extend_params_and_warmstart_cli
1 FAILED tests/models/test_opendecision.py::test_stage_a_param_set_excludes_head
1 FAILED tests/protocol/test_network_v2.py::test_assign_unit_when_only_island_batch_is_divisible
24 FAILED tests/protocol/test_network_v2.py::test_od_general_n_when_v2
1 FAILED tests/relay/test_transport_v2.py::test_launch_safety_independent_systemd_units
1 FAILED tests/relay/test_transport_v2.py::test_provider_fix_pinned_kind_version_and_fail_closed
1 FAILED tests/test_launch_cuda_run.py::test_preflight_names_the_first_missing_input_and_spends_nothing
1 FAILED tests/test_publish_exclusions.py::test_bare_selected_copy_has_publisher_runtime_controls
1 FAILED tests/test_publish_exclusions.py::test_current_repo_tree_scan_clean

Likely causes per group, not verified one by one:

  • od fixtures (test_od_island_v2, test_network_v2::test_od_general_n_when_v2, test_opendecision, test_real_od_two_rank_worker_paths): od-tiny dims were removed in feat(od): activate od-base, remove od-tiny (light CI) #7 and the fixtures still use them.
  • Native systemd/kernel cases (test_actual_*, test_real_capacity_*, test_launch_safety_independent_systemd_units): they are deselected in per-PR CI, run in heavy/native in an isolated root env, and fail when run as the plain runner user.
  • test_provider_fix_pinned_kind_version_and_fail_closed: expects kind v0.31.0, finds v0.33.0 on the runner image.

No continue-on-error was added. heavy stays red until these are fixed or the suite is scoped.

Activity

  1. added 2 commits that reference this issue on Oct 20, 2026
  2. echobt commented on Oct 26, 2026

    @echobt
    ContributorAuthor

    Closing: the heavy workflow (native, full-suite, canary, od-suite island/layout-n4/layout-rest) is green on main 679eb5b (run 2026-10-26 03:11Z, 1h09) and on the nightly runs of 24/10 and 25/10; the 23/10 failure was a separate nightly. The 77 real failures this issue tracked were fixed by the heavy-green work. Reopen with a run link if a regression returns.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions