Performance [P3]: Index sandbox activity overlap validation

[berndverst]

Summary

build_sandbox_worker_profiles() validated activity ownership with a linear
next(...) scan over every previously accumulated activity, calling
activities_overlap() for each candidate. That made sandbox worker profile
construction O(n²) in the number of registered activities — startup-only
control-plane work, but it degrades badly for large generated or versioned
sandbox catalogs.

This replaces the scan with _ActivityOwnerIndex, a private index in
profile_builder.py that gives O(1) overlap lookups per activity:

• _by_name — keyed on the casefolded activity name, used when the incoming
  activity is unversioned (an unversioned activity overlaps all versions of
  the same name).
• _by_name_and_version — keyed on (casefolded name, version) where None
  represents an unversioned registration, used when the incoming activity has
  an explicit version (it overlaps the same version plus any unversioned
  registration of that name).

Semantics preserved exactly

The index mirrors activities_overlap() byte-for-byte, including its
normalization rules:

Case	Overlaps?
Same name (case-insensitive), one side unversioned	Yes
Same name, same explicit version	Yes
Same name, different explicit versions	No
Version comparison	Case-sensitive exact string match, as before

Registration order is tracked (_ActivityOwnerSlot retains the first owner plus
the first owner that differs from it) so the profile named in the error is the
same one a linear scan would have reported. The exception type (ValueError)
and message text are unchanged:

Sandbox activity '<activity>' is assigned to both worker profile '<a>' and '<b>'.

Profile ordering in the returned list, the same-profile overlap allowance, and
the truthiness check on the conflicting profile id are all unchanged.

activities_overlap() remains exported from helpers.py — no public API
signature changes, no renames, no removed symbols, no changed import paths.
This is an internal performance refactor with no observable behavior change,
so per repo convention there is no changelog entry.

Fixes #190

Verification

Measured improvement (synthetic catalogs, legacy scan vs. index):

Activities	Legacy	Indexed	Speedup
500	0.019 s	0.0007 s	26x
2000	0.300 s	0.0030 s	99x
8000	4.910 s	0.0133 s	369x

Legacy timings scale ~16x per 4x input (quadratic); indexed timings scale
linearly.

Tests — 4 new tests in tests/durabletask-azuremanaged/test_sandboxes_extension.py:

• ..._rejects_versioned_activity_overlapping_unversioned_owner — unversioned
  owner vs. versioned candidate, asserts the exact conflict message.
• ..._rejects_unversioned_activity_overlapping_versioned_owner — versioned
  owner vs. unversioned candidate, asserts the exact conflict message.
• ..._rejects_identical_activity_versions — same name + same explicit version,
  asserts the exact conflict message.
• test_activity_owner_index_matches_linear_overlap_scan — differential test
  that replays 48 registrations across 25 seeded shuffles and asserts the index
  agrees with the original activities_overlap() linear scan at every step.

The existing overlap test also gained an exact-message assertion. The
pre-existing same-name/different-version allowance test still passes unchanged.

The differential test was mutation-checked: a variant of the index that drops
its second-owner tracking is caught on the first seed, confirming the test has
teeth.

Commands run (session-local venv, so the shared interpreter was untouched):

• python -m pytest tests/durabletask-azuremanaged/test_sandboxes_extension.py tests/durabletask-azuremanaged/test_durabletask_grpc_interceptor.py -q → 46 passed (42 before this change + 4 new)
• python -m pytest tests/durabletask -q --ignore-glob="*_e2e.py" → 681 passed, 7 skipped
• flake8 . in durabletask-azuremanaged/ and tests/durabletask-azuremanaged/ (matching CI) → clean
• pyright with the repo's strict pyrightconfig.json → 0 diagnostics in profile_builder.py

Baseline was captured before the change; the only difference was a flaky
pre-existing core test (test_sync_client_recreate_cooldown_prevents_immediate_repeated_recreation)
that passed on re-run. Remaining tests/durabletask-azuremanaged modules require
a live Durable Task Scheduler endpoint (TASKHUB unset) and are out of scope here.

[Copilot]

Pull request overview

This PR optimizes sandbox worker profile construction in durabletask.azuremanaged by replacing a quadratic activity-overlap validation scan with a private indexed lookup structure, preserving existing overlap semantics and error reporting while improving cold-start control-plane performance for large activity catalogs.

Changes:

• Replaced the linear activities_overlap() scan in build_sandbox_worker_profiles() with _ActivityOwnerIndex for O(1) per-activity overlap lookups.
• Added _ActivityOwnerSlot tracking to preserve “first conflicting owner” ordering to keep the raised ValueError message identical to the legacy behavior.
• Expanded unit tests to assert exact overlap error messages and added a differential test to ensure the index matches the legacy linear scan across shuffled registrations.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
durabletask-azuremanaged/durabletask/azuremanaged/preview/sandboxes/profile_builder.py	Introduces _ActivityOwnerIndex/_ActivityOwnerSlot and switches overlap validation to indexed lookups while keeping conflict semantics and messaging stable.
tests/durabletask-azuremanaged/test_sandboxes_extension.py	Adds targeted overlap regression tests (including exact-message assertions) plus a differential test ensuring indexed behavior matches the legacy scan.

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated no new comments.

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated no new comments.