fix: stop emptying sys.path while loading rxconfig (intermittent ModuleNotFoundError on `reflex run`) by Alek99 · Pull Request #7020 · reflex-dev/reflex

Alek99 · 2026-09-01T19:57:39Z

Summary

reflex run intermittently died at backend start with a spurious ModuleNotFoundError for an installed package (socketio, granian, reflex_components_sonner were all observed on the same machine, same venv). It happened on roughly half of dev-server starts while benchmarking a larger app.

Mechanism. reflex_base.config._load_config() cleared sys.path down to [cwd] for the duration of the rxconfig import and restored it afterwards. In reflex run the frontend thread's first get_config() (a fresh RegistrationContext) does that load while the main thread is still inside import reflex.app / the granian import. Python's import system walks sys.path by index, so any first-time import that overlaps the window fails.

Fix. Insert the project root at the front of sys.path once and leave it there (which reflex run already does for the app module via get_app_file()). Restoring the list afterwards is not safe either: removing an entry shifts indices under an in-progress import iterator in another thread. Measured with a stress script (one thread loading the config in a loop, main thread importing socketio 300 times):

loader	failed imports
main (`sys.path.clear()` + restore)	300 / 300
insert at 0 + restore afterwards	18 / 300
insert once, never remove (this PR)	0 / 300

The except Exception: retry with the original path fallback is gone with it: the original entries are never removed, so there is nothing to retry with.

Test

test_load_config_keeps_sys_path_intact_for_other_threads: a thread loads the config in a loop while the main thread imports a stdlib module 200 times. Fails on the old loader, passes on this one (mutation-checked).

🤖 Generated with Claude Code

`_load_config` cleared `sys.path` down to the project root for the duration of the rxconfig import and restored it afterwards. Python's import system walks `sys.path` by index, so any other thread importing during that window (in `reflex run`, the frontend thread loads the config while the main thread is still importing the backend) fails with ModuleNotFoundError for an unrelated, installed module (`socketio`, `granian`, `reflex_components_sonner` were all observed). The window is small but the backend import is long, so dev starts died intermittently with a spurious missing-module traceback. Insert the project root at the front of `sys.path` once and leave it there, which is what `reflex run` already does for the app module. Restoring the list afterwards is not safe either: removing an entry shifts indices under an in-progress import iterator in another thread (measured ~6% failures). Regression test: a thread loads the config in a loop while the main thread imports a stdlib module 200 times; fails on the old loader, passes now.

chatgpt-codex-connector · 2026-09-01T19:57:46Z

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

greptile-apps · 2026-09-01T20:00:27Z

Greptile Summary

The PR prevents concurrent configuration loading from temporarily emptying sys.path.

Keeps the current project root at the front of sys.path while preserving existing import paths.
Adds a concurrent regression test using a private temporary probe module.
Documents the intermittent ModuleNotFoundError fix in the release notes.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
packages/reflex-base/src/reflex_base/config.py	Replaces destructive `sys.path` clearing and restoration with an idempotent project-root insertion.
tests/units/test_config.py	Adds a concurrency regression test and now keeps the standard-library imports at module scope, resolving the prior review thread.
packages/reflex-base/news/7020.bugfix.md	Accurately records the intermittent import failure fixed by the configuration loader change.

_{Reviews (3): Last reviewed commit: "test: drop assertion made vacuous by the..." | Re-trigger Greptile}

codspeed-hq · 2026-09-01T20:01:06Z

Merging this PR will not alter performance

✅ 32 untouched benchmarks
⏩ 8 skipped benchmarks¹

_{Comparing alek/config-syspath-race (9070747) with main (812bb47)}

8 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

cubic-dev-ai

All reported issues were addressed

_{Reply with feedback, questions, or to request a fix.

Re-trigger cubic}

cubic-dev-ai

All reported issues were addressed across 3 files

_{Requires human review: Auto-approval blocked because this review re-detected 1 unresolved issue already reported by Cubic.

Re-trigger cubic}

masenf · 2026-09-01T20:30:58Z

overlaps with #6933, will decide how to reconcile

…e race test - config: insert the project root whenever it is not already sys.path[0] (cubic). With the previous `cwd not in sys.path` guard a later config load after a cwd change could resolve rxconfig from a stale project directory that was still ahead of it. Re-inserting only shifts entries back, which an in-flight import in another thread tolerates. - test: the probe was a shared stdlib module (colorsys), which rich also imports during config loading, so on Windows CI the loader thread's own import of it collided with the test's sys.modules.pop and died with KeyError inside importlib. Use a private module in a sibling temp directory appended to sys.path instead; nothing else imports it and it sits outside the project root, so the loader never evicts it. Still fails on main's loader (mutation-checked) and passes on this one. - test: importlib imported at module scope (greptile).

Alek99 · 2026-09-01T20:36:22Z

Re the overlap with #6933 (same bug, same window). Data that may help reconcile, from the stress script in this PR's description (one thread loading the config in a loop, main thread importing an installed module 300 times):

loader	failed imports
main: `sys.path.clear()` + restore	300 / 300
prepend cwd, remove afterwards (#6933's approach)	18 / 300
prepend once, never remove (this PR)	0 / 300

Removing the entry afterwards still races: PathFinder walks sys.path by index, so deleting index 0 under another thread's in-flight import makes it skip one entry. #6933 additionally does two things this PR does not (per-thread import recording for _config_module_deps, and the stale-rxconfig existence check), so the natural merge is #6933's recorder + existence check with the insert-once policy from here. Happy to rebase this onto #6933 or drop it, whichever is easier. The regression test here (test_load_config_keeps_sys_path_intact_for_other_threads) fails on both the main and the remove-afterwards loaders, so it can travel to either PR.

cubic-dev-ai

All reported issues were addressed across 2 files (changes from recent commits).

_{Reply with feedback, questions, or to request a fix.

Re-trigger cubic}

On Linux and Windows CI the probe module was evicted from sys.modules during its own exec (KeyError inside importlib), which only the config loader's dependency bookkeeping does; it does not reproduce locally. find_spec walks sys.path exactly like an import but registers nothing, so the loader cannot touch it. Still fails on main's loader (500 of 500 probes miss during the window) and passes on this one.

Ports the regression test from reflex-dev#7020. The existing test_load_config_keeps_sys_path_usable_for_other_threads gates on an Event with _get_config stubbed out, so it pins the invariant but never exercises the loader around a real rxconfig import. This one runs _load_config in a loop against an actual rxconfig.py while the main thread probes a private module at the end of sys.path with find_spec, which walks sys.path the way an import does without registering anything in sys.modules. Mutation-checked against main's loader (sys.path.clear() + restore): 500/500 probe lookups fail. On this branch's loader it passes, but not deterministically -- see the PR discussion for the measured residual rate. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012vT8EThKfwQuoPBjL2hZmo

masenf · 2026-09-01T22:42:50Z

taking #6933 instead

* fix(config): load rxconfig with cwd prepended instead of swapping out sys.path, racing concurrent first-time imports _load_config cleared sys.path down to the cwd for the duration of the rxconfig import, so any concurrent first-time import in another thread failed with ModuleNotFoundError (e.g. the lazy granian import when the backend starts while another thread loads the config). Prepending the cwd keeps the same resolution priority without blinding other threads. Dropping the clear also removes the except-retry fallback, which had been papering over a second bug: find_spec("rxconfig") answers from sys.modules, so a leftover module from another project directory faked the existence probe. Evict rxconfig from sys.modules before probing instead. * thanks greptile + add news fragment * fix: don't misattribute concurrent imports to rxconfig; remove exactly the inserted cwd entry * more reviews * Install the rxconfig import recorder only for the duration of the load * Keep the rxconfig import recorder installed instead of rebinding sys.meta_path * test: cover the rxconfig sys.path race with a real load loop Ports the regression test from #7020. The existing test_load_config_keeps_sys_path_usable_for_other_threads gates on an Event with _get_config stubbed out, so it pins the invariant but never exercises the loader around a real rxconfig import. This one runs _load_config in a loop against an actual rxconfig.py while the main thread probes a private module at the end of sys.path with find_spec, which walks sys.path the way an import does without registering anything in sys.modules. Mutation-checked against main's loader (sys.path.clear() + restore): 500/500 probe lookups fail. On this branch's loader it passes, but not deterministically -- see the PR discussion for the measured residual rate. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012vT8EThKfwQuoPBjL2hZmo * test: let the sys.path race test re-run instead of failing PRs Taking the prepended cwd entry back out is itself a sys.path shrink, so a probe walk that overlaps that one `del` can still skip an entry -- about 1 lookup in 5000, which flaked ~12% of runs. That residual is worth watching but is not actionable by whoever happens to open the next PR, so mark it flaky and let pytest-rerunfailures absorb it. reruns=3 puts a spurious failure at roughly 0.02% per run. Verified the reruns fire and that the test still fails all four attempts against a loader that clears sys.path outright, so a real regression is not masked. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012vT8EThKfwQuoPBjL2hZmo * tersify news fragment * fix(config): pass the load root into _get_config instead of re-reading cwd _get_config classified recorded imports as project-local by comparing them against Path.cwd() read *after* the rxconfig import. An rxconfig.py that changes the cwd (os.chdir, or anything it imports doing so) moved that root out from under the comparison, so its own project-local imports looked external: they were never recorded as deps, so the next load never evicted them and the following project inherited them from sys.modules. _load_config already resolves the root before prepending it to sys.path; hand that same Path down rather than resolving it twice. Removal of the sys.path entry was already immune to the drift, since it matches the inserted str by identity. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012vT8EThKfwQuoPBjL2hZmo refactor(config): fold _get_config into _load_config With the retry-on-failure path gone there is nothing left for the inner function to be a seam for: _load_config prepared sys.path, called _get_config exactly once, and tore sys.path back down. Merging them removes the indirection and the second place a project root had to be plumbed through. _load_config now takes project_root, defaulting to Path.cwd() resolved once up front, so a caller can load a config from a directory other than the cwd. Test changes follow the seam disappearing: - The two tests that stubbed _get_config to pause or perturb a load now drive a real rxconfig.py instead, which exercises the actual import path rather than a stub. Both were re-checked against the behavior they cover: by-value sys.path removal and clear()+restore each still fail them. - The event-pair setup those tests share with the concurrent-dep test is now a race_gate fixture rather than being rebuilt inline. - test_app.py patches _load_config, the seam that remains. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012vT8EThKfwQuoPBjL2hZmo refactor(config): keep the logic in _get_config, _load_config as entry point Same merge as the previous commit, inverted: _get_config holds the prepare/import/teardown body and takes the optional project_root, and _load_config is the no-argument entry point get_config and reload_config call. Keeping the _get_config name where the work happens leaves the 34 existing mock sites in test_app.py untouched, so the net delta is two files instead of three. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012vT8EThKfwQuoPBjL2hZmo deprecate _load_config in favor of _get_config get_config and reload_config now call _get_config directly, so nothing in the tree reaches _load_config any more. Since it may have callers outside the repo, keep the name working for now: it warns via console.deprecate (0.9.11, removal in 1.0) and delegates. A typing_extensions.deprecated stub under TYPE_CHECKING gives IDEs and type checkers the same signal without a second runtime warning. Test call sites move to _get_config with the loader. Two of them were patch targets (monkeypatch of the loader behind get_config, and the reload side_effect pair) that would have silently stopped intercepting had they kept pointing at the old name. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012vT8EThKfwQuoPBjL2hZmo * tersify deprecation news fragment Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012vT8EThKfwQuoPBjL2hZmo * _get_config: resolve absolute project_root path --------- Co-authored-by: Masen Furer <m_github@0x26.net> Co-authored-by: Claude <noreply@anthropic.com>

Alek99 requested a review from a team as a code owner September 1, 2026 19:57

news: 7020

9494be4

greptile-apps Bot reviewed Sep 1, 2026

View reviewed changes

Comment thread tests/units/test_config.py

cubic-dev-ai Bot reviewed Sep 1, 2026

View reviewed changes

Comment thread packages/reflex-base/src/reflex_base/config.py Outdated

cubic-dev-ai Bot reviewed Sep 1, 2026

View reviewed changes

Comment thread tests/units/test_config.py Outdated

Alek99 added 2 commits September 1, 2026 13:42

test: drop assertion made vacuous by the cleanup (cubic)

9070747

masenf closed this Sep 1, 2026

Navigation Menu

Uh oh!

fix: stop emptying sys.path while loading rxconfig (intermittent ModuleNotFoundError on `reflex run`) - #7020

fix: stop emptying sys.path while loading rxconfig (intermittent ModuleNotFoundError on `reflex run`)#7020
Alek99 wants to merge 5 commits into
mainfrom
alek/config-syspath-race

Alek99 commented Sep 1, 2026 •

edited by cubic-dev-ai Bot

Loading

Uh oh!

chatgpt-codex-connector Bot commented Sep 1, 2026

Uh oh!

greptile-apps Bot commented Sep 1, 2026 •

edited

Loading

Important Files Changed

Uh oh!

Uh oh!

codspeed-hq Bot commented Sep 1, 2026 •

edited

Loading

Uh oh!

cubic-dev-ai Bot left a comment •

edited

Loading

Uh oh!

Uh oh!

cubic-dev-ai Bot left a comment •

edited

Loading

Uh oh!

masenf commented Sep 1, 2026

Uh oh!

Alek99 commented Sep 1, 2026

Uh oh!

cubic-dev-ai Bot left a comment •

edited

Loading

Uh oh!

Uh oh!

masenf commented Sep 1, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

Alek99 commented Sep 1, 2026 • edited by cubic-dev-ai Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Test

Uh oh!

chatgpt-codex-connector Bot commented Sep 1, 2026

Uh oh!

greptile-apps Bot commented Sep 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Greptile Summary

Confidence Score: 5/5

Important Files Changed

Uh oh!

Uh oh!

codspeed-hq Bot commented Sep 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Merging this PR will not alter performance

Footnotes

Uh oh!

cubic-dev-ai Bot left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

cubic-dev-ai Bot left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

masenf commented Sep 1, 2026

Uh oh!

Alek99 commented Sep 1, 2026

Uh oh!

cubic-dev-ai Bot left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

masenf commented Sep 1, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Alek99 commented Sep 1, 2026 •

edited by cubic-dev-ai Bot

Loading

greptile-apps Bot commented Sep 1, 2026 •

edited

Loading

codspeed-hq Bot commented Sep 1, 2026 •

edited

Loading

cubic-dev-ai Bot left a comment •

edited

Loading

cubic-dev-ai Bot left a comment •

edited

Loading

cubic-dev-ai Bot left a comment •

edited

Loading