Skip to content

Sync BaseCodeExecutor.execute_code is called directly on the event loop in llm_flows/_code_execution — long executions freeze the whole server #6353

Closed

Description

@mhonda-blueish

** Please make sure you read the contribution guide and file the issues in the right place. **

Describe the bug

BaseCodeExecutor.execute_code() has a synchronous contract, and the LLM flow calls it directly on the running event loop:

  • src/google/adk/flows/llm_flows/_code_execution.py — _run_pre_processor (call site around L246)
  • src/google/adk/flows/llm_flows/_code_execution.py — _run_post_processor (call site around L356)

Both enclosing functions are async, so any executor whose execute_code performs blocking I/O freezes the entire event loop for the duration of the execution.

This is especially severe with AgentEngineSandboxCodeExecutor: a single sandboxes.execute_code call is a blocking HTTP request that regularly takes 2–3+ minutes for document-generation workloads. While it is in flight, the whole ADK FastAPI server (get_fast_api_app / api_server) is unresponsive:

  • /health-style endpoints and every other request stall (we measured 113–127 s of complete unresponsiveness that aligned exactly with in-flight execute_code calls),
  • on Kubernetes this fails liveness probes and the pod gets killed mid-task,
  • SSE streams of other sessions co-hosted on the same process stall as well,
  • in-process timers (e.g. per-step LLM timeouts of other agents) fire spuriously once the loop unfreezes.

To Reproduce

  1. Configure an LlmAgent with any BaseCodeExecutor whose execute_code blocks (e.g. AgentEngineSandboxCodeExecutor with a long-running snippet, or a toy executor with time.sleep(120)).
  2. Serve it with the ADK FastAPI app and trigger a code-execution turn via /run_sse.
  3. While the execution is in flight, request /health (or anything else) — the server does not respond until execute_code returns.

Expected behavior

Long-running code execution should not block the event loop / other requests served by the same process.

Proposed fix

Offload the call at both sites, keeping the executor contract unchanged:

code_execution_result = await asyncio.to_thread(
    code_executor.execute_code,
    invocation_context,
    CodeExecutionInput(...),
)

BuiltInCodeExecutor is not affected (it short-circuits earlier), and executor implementations need no changes. An alternative (larger) fix would be an async contract such as execute_code_async with a default to_thread wrapper around the sync method.

I'm happy to send a PR for the asyncio.to_thread variant if that direction is acceptable.

Desktop

  • OS: Linux (GKE) / macOS
  • Python version: 3.12 / 3.13
  • ADK version: 2.3.0 (call sites verified unchanged on current main)

Model Information

  • Not model-specific.

Activity

  1. added
    core[Component] This issue is related to the core interface and implementation
    on Jul 10, 2026
  2. ItsMacto commented on Jul 11, 2026

    @ItsMacto
    Contributor

    Hi, apologies for jumping ahead of the assignment here. I saw this and opened #6375 with a minimal fix (both execute_code call sites offloaded via asyncio.to_thread, matching the existing pattern in the artifact services, plus deterministic regression tests for both paths). @llalitkumarrr if you already have work in progress on this, happy to close mine or fold anything useful into yours.

  3. added a commit that references this issue on Jul 12, 2026
    585809c
  4. added a commit that references this issue on Aug 11, 2026
    892daac
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

core[Component] This issue is related to the core interface and implementation

Type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions