SWE-agent takes a GitHub issue and tries to automatically fix it, using your LM of choice. It can also be employed for offensive cybersecurity or competitive coding challenges. [NeurIPS 2024]
fix(run): use Fix: threading.local() in SaveApplyPatchHook to prevent race condition in multi-worker batch runs (fixes #1284) (#1366)
Shivam Tiwari78c1a06fdf4a
diff --git a/sweagent/run/hooks/apply_patch.py b/sweagent/run/hooks/apply_patch.pyindex f3a982d8..d199ec1f 100644--- a/sweagent/run/hooks/apply_patch.py+++ b/sweagent/run/hooks/apply_patch.py@@ -1,4 +1,5 @@import subprocess+import threadingfrom pathlib import Pathimport rich@@ -21,27 +22,30 @@ class SaveApplyPatchHook(RunHook):self.logger = get_logger("swea-save_apply_patch", emoji="⚡️")self._apply_patch_locally = apply_patch_locallyself._show_success_message = show_success_message+ # Thread-local storage so that concurrent workers in run-batch do not+ # overwrite each other's per-instance state (_env, _problem_statement).+ self._local = threading.local()def on_init(self, *, run):self._output_dir = Path(run.output_dir)def on_instance_start(self, *, index: int, env: SWEEnv, problem_statement: ProblemStatementConfig):- self._env = env- self._problem_statement = problem_statement+ self._local.env = env+ self._local.problem_statement = problem_statementdef on_instance_completed(self, *, result: AgentRunResult):- instance_id = self._problem_statement.id+ instance_id = self._local.problem_statement.idpatch_path = self._save_patch(instance_id, result.info)if patch_path:if not self._apply_patch_locally:returnif not _is_promising_patch(result.info):return- if self._env.repo is None:+ if self._local.env.repo is None:return- if not isinstance(self._env.repo, LocalRepoConfig):+ if not isinstance(self._local.env.repo, LocalRepoConfig):return- local_dir = Path(self._env.repo.path)+ local_dir = Path(self._local.env.repo.path)self._apply_patch(patch_path, local_dir)@staticmethoddiff --git a/tests/test_run_hooks.py b/tests/test_run_hooks.pyindex 0ca8d022..a1f25ed2 100644--- a/tests/test_run_hooks.py+++ b/tests/test_run_hooks.py@@ -1,8 +1,12 @@import os+import threading+from concurrent.futures import ThreadPoolExecutor+from unittest.mock import MagicMockimport pytest-from sweagent.agent.problem_statement import GithubIssue+from sweagent.agent.problem_statement import GithubIssue, TextProblemStatement+from sweagent.run.hooks.apply_patch import SaveApplyPatchHookfrom sweagent.run.hooks.open_pr import OpenPRConfig, OpenPRHookfrom sweagent.types import AgentRunResult@@ -73,3 +77,53 @@ def test_should_open_pr_success_has_pr_override(open_pr_hook_init_for_sop, agenthook._problem_statement = GithubIssue(github_url="https://github.com/swe-agent/test-repo/issues/19")hook._config.skip_if_commits_reference_issue = Falseassert hook.should_open_pr(agent_run_result)+++def test_save_apply_patch_hook_concurrent_workers_save_to_correct_dirs(tmp_path):+ """Regression test for #1284: concurrent workers must not overwrite each+ other's per-instance state (_problem_statement, _env) in SaveApplyPatchHook.++ Before the fix, a single shared hook instance stored _problem_statement as a+ plain instance attribute. When two workers both called on_instance_start()+ before either finished, the second write overwrote the first, causing both+ workers to save their patch into the same (wrong) output directory.++ The fix uses threading.local() so every worker thread sees its own copy.+ """+ hook = SaveApplyPatchHook(show_success_message=False)+ hook._output_dir = tmp_path++ # Barrier ensures both workers have called on_instance_start before either+ # proceeds to on_instance_completed, making the race window deterministic.+ barrier = threading.Barrier(2)++ def worker(instance_id: str, patch_content: str) -> None:+ ps = TextProblemStatement(text=f"Issue for {instance_id}", id=instance_id)+ env = MagicMock()+ env.repo = None # skip apply_patch_locally path++ hook.on_instance_start(index=0, env=env, problem_statement=ps)++ # Hold here until both threads have written their problem_statement so+ # that the race window is guaranteed to be open.+ barrier.wait()++ result = AgentRunResult(+ info={"submission": patch_content, "exit_status": "submitted"},+ trajectory=[],+ )+ hook.on_instance_completed(result=result)++ with ThreadPoolExecutor(max_workers=2) as pool:+ fa = pool.submit(worker, "instance-A", "patch A content")+ fb = pool.submit(worker, "instance-B", "patch B content")+ fa.result()+ fb.result()++ patch_a = tmp_path / "instance-A" / "instance-A.patch"+ patch_b = tmp_path / "instance-B" / "instance-B.patch"++ assert patch_a.exists(), "Patch for instance-A was not saved to its own directory"+ assert patch_b.exists(), "Patch for instance-B was not saved to its own directory"+ assert patch_a.read_text() == "patch A content"+ assert patch_b.read_text() == "patch B content"