[dancingactor]

Description

This PR skips placement group resource parsing when updating existing available resources.

Since this updating logic is going to be used in global scheduling (selecting a node within the cluster to assign a task) and local allocation in the upcoming PR #65871, this critical path needs to be optimized.

Performance Test

The optimization is 1.7×–24× faster on the alloc/free paths:

Scenario	Before	After	Speedup
CPU alloc+free	4252 ns	2448 ns	1.74×
CPU alloc failure	2261 ns	2229 ns	~1.0×
GPU alloc+free (full)	4283 ns	2357 ns	1.82×
GPU alloc+free (fractional)	4294 ns	2412 ns	1.78×
GPU alloc failure	2214 ns	2306 ns	~1.0×
Free only	980 ns	41 ns	24.0×
Subtract	1016 ns	101 ns	10.1×

Related PR

Related to #65871

[gemini-code-assist]

Code Review

This pull request refactors the resource allocation logic in NodeResourceInstanceSet::TryAllocate to optimize lookups and handle implicit resources. However, the implementation incorrectly assumes that mixed resource demands (such as 1.5 GPUs) are invalid, splitting the allocation into mutually exclusive integer and fractional branches. This introduces a critical regression where mixed requests always fail, which also led to modifying several unit tests to use integer demands instead of fractional ones. To resolve this, the allocation logic should be updated to allow mixed demands by falling through to fractional allocation after allocating full instances, and the corresponding unit tests should be reverted to their original state.

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{Reviewed by Cursor Bugbot for commit e6d5e7e. Configure here.}

[Yicheng-Lu-llll]

I suspect the major time we spend is regex(ParsePgFormattedResource) not vector copying. could you do some experiments to very that? if regex is the major cost, If, we could just skip the parse when the key is already there.

place and undoing on failure feels fragile.

[dancingactor]

After verifying, your suspicion makes perfect sense. The biggest cost is regex(ParsePgFormattedResource) not vector copying, so I've gone with the Set()-only fix rather than the in-place-mutation approach. It recovers ~90–95% of the total speedup and requires no fragile undo-on-failure logic

Scenario	Before (master)	Set() fix only	in-place only	in-place + Set() fix
CPU alloc+free	4252.42	2448.00	2264.03	2302.67
CPU alloc failure	2260.87	2229.40	2085.74	2109.22
GPU alloc+free (full)	4282.83	2356.93	2242.38	2290.81
GPU alloc+free (frac)	4293.57	2412.36	2240.00	2329.91
GPU alloc failure	2213.86	2305.68	2190.77	2136.17
Free only	980.00	40.80	5.00	4.90
Subtract	1015.98	100.76	35.54	34.20

[github-actions]

This pull request has been automatically marked as stale because it has not had
any activity for 14 days. It will be closed in another 14 days if no further activity occurs.
Thank you for your contributions.

You can always ask for help on our discussion forum or Ray's public slack channel.

If you'd like to keep this open, just leave any comment, and the stale label will be removed.

[Yicheng-Lu-llll]

Thanks! left some nits, would be great to also update the pr body.

[dancingactor]

Updated the PR description. Thanks for review!