[mnkj0021]

Preserve real YARN worker identities in partial-log reports.

• NodeManager registration carries the reported host identity to container starts without leaking one worker's identity into another file.
• Cross-file inference is only used when a directory has one unambiguous registered host; ambiguous multi-worker layouts remain unresolved instead of borrowing the wrong capacity.
• ResourceManager assignment keeps the assigned host in the node inventory even when registration evidence is absent, while explicitly treating that node as unregistered/incomplete for final accounting.
• A later NodeManager registration without instance-type metadata preserves an instance type already learned from the ResourceManager.
• Existing incomplete-accounting/final-cost safeguards remain in place.
• Regression coverage includes NM-only identity, RM assignment without registration, split logs, co-located workers, later conflicting registration, instance-type preservation, and mapped-but-unregistered nodes.

Validation completed before the latest human-review follow-up:

• Python 3.12 full yarn-resource-cost unittest suite: 73/73 passed on the prior review-fix head
• py_compile: passed
• git diff --check: passed

The current head is b0920a5. The latest license-header and YARN Python workflows are action_required and need NVIDIA maintainer/vetter approval, so this description does not claim a fresh CI pass on the final review-fix head.

Fixes #276

[greptile-apps]

Confidence Score: 5/5

[Medium risk] Fixes worker identity tracking in resource cost logs.

The PR appears safe to merge; no new actionable issue or outstanding previous finding was identified.

Summary

The PR preserves reported worker identities in partial YARN logs and creates unknown-capacity nodes for RM allocations without registration evidence.

• Adds registration-evidence warnings and incomplete accounting for affected containers.
• Adds regressions for split logs, shared directories, and unregistered allocations.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  Logs[YARN logs] --> Registration[Registration host and capacity]
  Logs --> Allocation[RM allocation or NM start]
  Registration --> Reconcile[Reconcile container node]
  Allocation --> Reconcile
  Reconcile --> Report[Completeness and cost report]

Loading

_{Reviews (5) · Last reviewed commit: "test(yarn): cover registration and host-..."}

[gerashegalov]

Could we keep the host in the node inventory without treating this placeholder as registration evidence? With assignment maxima and a node-class map, this now produces a final cost even though no node registration was logged.

[mnkj0021]

Addressed on the current head b0920a5. RM-assignment placeholders remain in the node inventory, but final accounting now treats a used node without registration evidence as incomplete even if a node-class map makes it priceable. I added regression coverage for the mapped-but-unregistered case. Leaving this thread open for your confirmation.

[gerashegalov]

Could we track whether this file has reported a host instead of comparing node_id with the directory name? If the directory is worker-a, a later worker-b registration can relabel A’s container while leaving A’s capacity on it.

[mnkj0021]

Addressed on b0920a5. NodeManager parsing now tracks an explicit reported host per file rather than comparing the mutable node_id with the directory name, so a later registration for another worker cannot relabel an already-attributed container. Added a regression for that sequence. Leaving this open for your review.

[gerashegalov]

Could we preserve the existing instance type when this NM line doesn’t include one? Otherwise it clears a class already reported by the RM and can prevent pricing.

[mnkj0021]

Addressed on b0920a5. An NM registration now replaces instance_type only when that registration line actually reports one, so an RM-provided instance type is preserved otherwise. Added a focused regression. Leaving the thread open for confirmation.

[greptile-apps]

Want your agent to iterate on Greptile's feedback? Try greploops.