Isolate Google ADK's deterministic providers on a workflow-private random stream - #1854
Merged
Merged
Conversation
ADK keeps its time, id, and random providers in contextvars.ContextVars. GoogleAdkPlugin set them in the worker's context, but workflow tasks run on the workflow task executor's threads, which start with an empty context, so ADK code inside a workflow read the defaults: wall-clock time and uuid.uuid4() for session, event, invocation, and function-call ids. Only debug mode, which runs activations inline, saw the deterministic values. Rebind each google.adk.platform ContextVar to one whose default is the Temporal provider so it is visible from every context, on Worker and Replayer alike. Also install the random provider ADK added in 2.8.0 and raise the google-adk floor to 2.8.0.
Use workflow.time() for the time provider, warn when installing replaces a provider set earlier in the calling context, and document that overrides must be made after the worker starts and that ADK id and random generation raise ReadOnlyContextError in read-only contexts. Tests assert provider identity.
Contributor
Author
|
Notes for review:
|
There was a problem hiding this comment.
🟡 Changes recommended
ADK’s public reset functions can restore nondeterministic standard-library providers inside workflows.
Once you've addressed the issues Copilot identified, you can request another Copilot review.
Pull request overview
Ensures Google ADK uses Temporal’s deterministic time, UUID, and random providers across worker and replay threads.
Changes:
- Installs thread-visible, idempotent ADK provider defaults.
- Adds worker, sandbox, replay, fallback, and override tests.
- Requires Google ADK 2.8.0 and documents compatibility implications.
File summaries
| File | Description |
|---|---|
temporalio/contrib/google_adk_agents/_plugin.py |
Implements deterministic provider installation. |
tests/contrib/google_adk_agents/test_adk_platform_providers.py |
Tests provider behavior across execution contexts. |
temporalio/contrib/google_adk_agents/README.md |
Documents provider semantics. |
pyproject.toml |
Raises the Google ADK minimum version. |
uv.lock |
Locks Google ADK 2.8.0 and dependencies. |
CHANGELOG.md |
Records the fix and replay compatibility warning. |
Review details
- Files reviewed: 5/6 changed files
- Comments generated: 1
- Review effort level: Balanced
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
patbeqo
requested changes
Sep 11, 2026
reset_*_provider() restores the module's _default_* binding, so install now rebinds that too; a set-then-reset cycle lands back on the deterministic providers instead of wall clock and stdlib random. ADK ids and randoms now come from a workflow.new_random() cached on the workflow instance (as the opentelemetry and langsmith integrations do) so ADK draws never shift the sequence user code sees from workflow.random(), with an explicit read-only guard so query handlers cannot advance the cached stream.
patbeqo
requested changes
Sep 15, 2026
# Conflicts: # uv.lock
Per review: query handlers and update validators now get a nondeterministic fallback random instead of ReadOnlyContextError - their results are never replayed, and the guard only has to keep the cached private stream untouched (a replay test proves it stays untouched). workflow.uuid4() accepts an optional keyword-only random argument so the uuid-from-generator derivation lives in one place; the ADK id provider and langsmith's _uuid_from_random now delegate to it. Changelog entries rehomed under Unreleased after the 1.33.0 cut and reworded for the private-stream design.
patbeqo
requested changes
Sep 15, 2026
Query handlers and update validators now get time.time() from the ADK time provider: their results are never replayed, and workflow.time() would hand them the last activation's timestamp, stale by however long the workflow has been parked. Covered by a query in the query-during-run test.
Per review: the keyword random shadowed the module-level random() function inside uuid4, which forced the body to inline the runtime call. With the rename the body simply calls random() again, and callers read as workflow.uuid4(rng=...).
patbeqo
approved these changes
Sep 16, 2026
Per review: the read-only and outside-workflow fallbacks now return a fresh random.Random() instead of module-level singletons. Nothing needs to persist there - read-only results are never replayed, and ADK's guidance to return an existing instance only matters for a seeded generator whose sequence must continue; an unseeded one draws fresh OS entropy either way, and ADK's only get_random() caller uses the value immediately for retry jitter.
…workflow-threads Resolve temporalio/contrib/google_adk_agents/_plugin.py in favor of this branch's provider installation (module-level providers, the locked _install_provider that also rebinds ADK's _default_* bindings, read-only handling, and the workflow-private random stream), which supersedes the three-argument _install_provider and the workflow.random()-sharing providers added by #1675; keep #1675's optional model SDK passthrough. Reconcile the CHANGELOG to one entry for the google-adk>=2.8.0 bump and drop #1675's note about workflow.random()/workflow.uuid4() sequences shifting across the upgrade, which no longer applies now that ADK draws from a private stream. Reword three #1675 test comments to match.
…workflow-threads Resolve _plugin.py by keeping both the OpenTelemetry passthrough bullet from #1899 and this branch's providers bullet, take main's uv.lock (pyproject is unchanged by this branch), and rebuild the CHANGELOG from main's text: the union merge had moved this branch's notes into the released 1.34.0 section. The Unreleased entries now describe this branch relative to 1.34.0, which already shipped #1675's provider installation, including a breaking-change note for workflows started on 1.34.0.
DABH
changed the title
is_read_only() is also true while a dynamic workflow's dynamic_config runs, and that hook is replayed, so a wall-clock fallback there was not replay safe. The time provider now returns workflow.time() in every in-workflow context; a query sees the activation timestamp instead of the wall clock, which is harmless because nothing a query computes is persisted. Ids and randoms keep the fresh-entropy fallback in read-only contexts, where no deterministic source is available (workflow.random() raises and new_random() needs a reseed callback that read-only mode forbids); the comments now say so and name dynamic_config as the one replayed read-only hook. Also stop claiming uuid4(rng=...) touches no workflow state: it advances the supplied generator, which a caller could make the shared one.
DABH
enabled auto-merge (squash)
October 1, 2026 05:47
tconley1428
reviewed
Oct 1, 2026
With an explicit generator uuid4 reads no workflow state, so it does not belong under workflow.*. The plugin now builds the UUID from its private stream with the same construction, and contrib.langsmith goes back to its own one-line construction.
brianstrauch
requested changes
Oct 1, 2026
brianstrauch
left a comment
brianstrauch
left a comment
Member
There was a problem hiding this comment.
Requesting changes for two reproduced issues in the ADK RNG provider: failures during workflow construction or with slotted classes, and nondeterministic IDs from replayed dynamic_config callbacks.
The existing ADK suite (88 passed, 5 skipped) and poe lint passed on this head. Focused real-worker reproductions exposed the cases described below.
…-only draws to queries and validators workflow.instance() is None during __init__, and a class with __slots__ cannot take a cache attribute, so storing the private stream on the workflow object failed in both cases. Key it by the SDK's per-run runtime object in a WeakKeyDictionary instead. The entropy fallback for read-only code also covered a dynamic workflow's dynamic_config, which is replayed; ids drawn there differed on replay. Only query handlers and update validators are never replayed, so only they get fresh entropy. Other read-only contexts raise ReadOnlyContextError, as workflow.random() does. The runtime exposes the internal distinction as workflow_in_query_or_validator().
workflow.new_random() seeds from the same value as workflow.random(), so an unnamed private stream starts out identical to the workflow's own: the Nth ADK id equaled the Nth workflow.uuid4(). new_random() now takes an optional name that is mixed into the seed (and into every reseed), and the plugin uses one. Without a name nothing changes.
brianstrauch
approved these changes
Oct 2, 2026
…enerator does Drop the workflow_in_query_or_validator runtime accessor and the ReadOnlyContextError for replayed read-only callbacks: minting ADK ids inside dynamic_config or a patch activation callback is not worth runtime plumbing. Read-only code gets a fresh unseeded generator and the private stream is never touched there, matching contrib.opentelemetry.
tconley1428
approved these changes
Oct 5, 2026
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 was changed
1.34.0 (via #1675) made
GoogleAdkPlugininstall ADK's time, uuid, and random providers asContextVardefaults, so they finally apply inside workflow tasks. This PR finishes that work:workflow.new_random("temporalio.contrib.google_adk_agents")per run, kept in aWeakKeyDictionarykeyed by the SDK's per-run runtime object so it works during__init__and with__slots__classes) instead ofworkflow.random().new_random()gains an optionalnamemixed into the seed: an unnamed instance starts out identical toworkflow.random(), which would have made the Nth ADK id equal the Nthworkflow.uuid4()(Brian'sRequestInputexample). Without a name nothing changes. ADK's draw count no longer shifts the sequences user code sees fromworkflow.random()/workflow.uuid4(), including acrossgoogle-adkupgrades that change how many ids ADK generates. The plugin builds its v4 UUIDs from that stream itself, with the same construction asworkflow.uuid4(); there are no public API changes.workflow.random()raises andnew_random()needs a reseed callback read-only mode forbids. In 1.34.0 an ADK id or random draw inside a query fails it withWhile in read-only function, action attempted: random._install_provideralso rebinds ADK's_default_*bindings, soreset_*_provider()restores the Temporal providers instead of the standard-library ones (in 1.34.0 a reset silently returns ADK to wall-clock time). Installation takes a lock, is idempotent, and warns when it replaces an override set before the worker started.💥 Breaking change
Workflows started on 1.34.0 that generated ADK ids or jitter — e.g. one parked on a HITL response whose recorded function-call id was minted from the shared stream — may not replay deterministically across this upgrade. 1.34.0 carried the same caveat for the previous transition; the module is experimental and 1.34.0 shipped on 2026-09-30, so the window of affected workflows is as small as it will ever be. Drain such workflows or use worker versioning. Logged under Breaking Changes.
Testing
tests/contrib/google_adk_agents/test_adk_platform_providers.py: 14 tests over sandboxed and unsandboxed workers plus replay — provider visibility from fresh threads, read-only draws, reset semantics, idempotence, the override warning, outside-workflow fallbacks, and ids drawn in__init__and in a__slots__workflow. Run against 1.34.0's plugin, 8 of them fail on behavior (the read-only query failure above, wall-clock time afterreset_time_provider(), non-idempotent re-install, no warning on an orphaned override), 3 fail only because they reference this PR's module-level helpers, and 1 passes. Full ADK and LangSmith suites and the lint gate pass on the merged tree.