Verify support for Python 3.15 - #2

Codecov Report

❌ Patch coverage is 96.66667% with 1 line in your changes missing coverage. Please review.

📢 Thoughts on this report? Let us know!

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

_{Comment @coderabbitai help to get the list of available commands.}

CI failures: stale base, not the 3.15 change

The branch is on 0a27f63, 7 commits behind main. Both jobs fail because of that:

• test (3.15): the old pyarrow>=18,<26 pin resolves to 25.0.1, which only ships wheels up to cp314, so uv builds it from source and CMake can't find Arrow C++. main now allows <27, and pyarrow 26.0.0 has cp315 wheels.
• test (3.10): with polars 2.0.0, a timestamp past year 9999 raises OverflowError: date value out of range instead of a Polars panic. main updated both avro tests for this in a9a0b53.

Rebasing on main should clear both.

The next blocker on 3.15 is duckdb: 1.5.6 has no cp315 wheel yet, so the job will try a full source build. Keep 3.14 in the matrix and add 3.15 next to it until duckdb ships 3.15 wheels?

Thank you. Yes, let's wait for duckdb having compatible binary wheel packages published. Thank you for creating a ticket about it already, @eavanvalkenburg.

• Publish CPython 3.15 wheels duckdb/duckdb-python#659

CI run #39690369174 reveals that validation on Python 3.15 does not succeed yet. However, we haven't been able identify the root cause: The process gets cancelled with no obvious error message or other hints about it.

@hampsterx or @coderabbitai: Can you find any clues why this goes south? I have re-ran the CI workflow three times already, so I'd consider the job to be "deterministically failing" -- but why?

#!/bin/bash
set -uo pipefail
echo '=== Run metadata ==='
gh api repos/dlt-contrib/dlt-filesystem/actions/runs/39690369174 --jq '{id,head_sha,path,event,status,conclusion,run_attempt,created_at,updated_at,html_url}'
echo '=== All jobs ==='
gh api --paginate 'repos/dlt-contrib/dlt-filesystem/actions/runs/39690369174/jobs?filter=all&per_page=100' --jq '.jobs[] | {id,name,status,conclusion,started_at,completed_at,steps,html_url}'
echo '=== Run log tail ==='
gh run view 39690369174 --repo dlt-contrib/dlt-filesystem --log 2>&1 | tail -220
echo '=== Workflow file discovery ==='
fd --hidden -t f -e yml -e yaml . .github

#!/bin/bash
set -uo pipefail
echo '=== Workflow outline ==='
ast-grep outline .github/workflows/tests.yml
echo '=== Workflow content ==='
cat -n .github/workflows/tests.yml
for job in 119534795387 119536561745 119542386305; do
  echo "=== Job $job log tail ==="
  gh api "repos/dlt-contrib/dlt-filesystem/actions/jobs/$job/logs" | tail -150
  echo "=== Job $job check annotations ==="
  check_url=$(gh api "repos/dlt-contrib/dlt-filesystem/actions/jobs/$job" --jq .check_run_url)
  if [ -n "$check_url" ] && [ "$check_url" != null ]; then
    gh api "$check_url" --jq '{conclusion,output}'
    gh api "$check_url/annotations" --jq '.[] | {path,start_line,annotation_level,message}'
  fi
done
echo '=== Test configuration discovery ==='
fd --hidden -t f '^(pyproject\.toml|Makefile|pytest\.ini|tox\.ini|conftest\.py)$'

#!/bin/bash
set -uo pipefail
for job in 119534795387 119536561745 119542386305; do
  echo "=== Job $job log tail ==="
  gh api --allow-escape-sequences "repos/dlt-contrib/dlt-filesystem/actions/jobs/$job/logs" | tail -110
done
echo '=== Project configuration outline ==='
ast-grep outline pyproject.toml
echo '=== Project configuration ==='
cat -n pyproject.toml
echo '=== Test fixture outline ==='
ast-grep outline tests/conftest.py

#!/bin/bash
set -uo pipefail
echo '=== URI test outline ==='
ast-grep outline tests/test_util.py
echo '=== URI test code ==='
rg -n -C 30 'def test_requote_uri' tests/test_util.py
echo '=== URI implementation discovery ==='
rg -n -C 5 --type py 'def requote_uri|import .*requote_uri|from .* import.*requote_uri' .
echo '=== Latest job log selected evidence ==='
gh api --allow-escape-sequences repos/dlt-contrib/dlt-filesystem/actions/jobs/119542386305/logs | python -c '
import sys,re
lines=sys.stdin.read().splitlines()
pending={}
for line in lines:
    if re.search(r"Python|requests==|urllib3==|coverage==|pytest.*==|Runner Image|Image:|Version:|Poe =>|worker|shutdown|Memory|memory|disk",line):
        print(line)
    result=re.search(r"\[gw\d+\].*?\b(PASSED|FAILED|SKIPPED|ERROR|XFAIL|XPASS)\s+(.*)",line)
    if result:
        pending.pop(result.group(2).strip(),None)
    else:
        test=re.search(r"\b(tests/\S+::.*)",line)
        if test:
            pending[test.group(1).strip()]=line
print("=== Test starts without matching result (may include formatting mismatches) ===")
for line in pending.values(): print(line)
'
echo '=== Python 3.10 equivalent test output ==='
gh api --allow-escape-sequences repos/dlt-contrib/dlt-filesystem/actions/jobs/119534795016/logs | grep -E 'sig=%80|Python 3|requests==|coverage=='

#!/bin/bash
set -uo pipefail
echo '=== YAML test outline ==='
ast-grep outline tests/format/test_yaml.py
echo '=== Recursive alias test ==='
rg -n -C 25 'def test_recursive_alias_raises_cleanly' tests/format/test_yaml.py
echo '=== YAML codec discovery ==='
fd -t f -i 'yaml'
rg -n -C 5 --type py 'def .*yaml|safe_load|SafeLoader|RecursionError|recursive.alias' src tests/format/test_yaml.py
for job in 119534795387 119536561745 119542386305 119534795016; do
  echo "=== Job $job YAML evidence ==="
  gh api --allow-escape-sequences "repos/dlt-contrib/dlt-filesystem/actions/jobs/$job/logs" | grep -Ei 'pyyaml==|test_recursive_alias|test_yaml.py|shutdown signal' | tail -35
done

#!/bin/bash
set -uo pipefail
ast-grep outline src/dlt_filesystem/source/format/iterable_codec.py
sed -n '1,115p' src/dlt_filesystem/source/format/iterable_codec.py
sed -n '175,225p' src/dlt_filesystem/source/format/iterable_codec.py
rg -n -C 8 'normaliz|def read_via_iterable' src/dlt_filesystem/source/format/iterable_codec.py | tail -145

@amotl Yes. All three attempts stall at the same recursive YAML test. The last test shown in the log is misleading because pytest runs two workers.

Evidence

In each Python 3.15 attempt, worker gw0 starts this test but never reports a result:

tests/format/test_yaml.py::test_recursive_alias_raises_cleanly

The same test passes in approximately one millisecond on Python 3.10.

Worker gw1 continues through other tests until progress reaches 86%. The runner then reports:

The runner has received a shutdown signal. This can happen when the runner service is stopped, or a manually started runner is canceled.
The operation was canceled.

See the latest failed job.

Likely trigger

At Line 157 of tests/format/test_yaml.py, the test loads a self-referential YAML structure:

a: &a [*a]

The test expects normalization to terminate with RecursionError. In src/dlt_filesystem/source/format/iterable_codec.py, _normalize_row() passes this cyclic structure to dlt.common.utils.map_nested_values_in_place() at Line 438.

This makes cyclic normalization the primary suspect, not DuckDB installation or the final URI test. Installation and standalone imports succeed.

The logs do not prove why the runner shuts down. Memory exhaustion is a possible explanation, but there is no OOM diagnostic in the retrieved logs. A Python 3.15 or coverage interaction also remains unverified.

Next diagnostic check

In the Python 3.15 environment, run only this test without parallel workers. Use an external timeout to bound the run:

timeout --kill-after=5s 60s \
  python -m pytest -o addopts='' -vv \
  -o faulthandler_timeout=15 \
  tests/format/test_yaml.py::test_recursive_alias_raises_cleanly

Then repeat with coverage enabled:

timeout --kill-after=5s 60s \
  python -m pytest -o addopts='' -vv --cov \
  -o faulthandler_timeout=15 \
  tests/format/test_yaml.py::test_recursive_alias_raises_cleanly

These checks should capture a stack trace before shutdown and distinguish a normalization problem from a coverage interaction. I could inspect the logs and code, but I could not execute this reproduction here.

If the isolated test confirms the cycle traversal, the robust fix is explicit cycle detection before normalization. It should not depend on the interpreter eventually raising RecursionError.

You are interacting with an AI system.

Observation

Indeed when executed on Python 3.15, the process goes into an endless loop or infinite recursion, hogging one CPU core at 100%, impossible to be killed by a regular CTRL+C. The log output shows:

tests/format/test_yaml.py::test_recursive_alias_raises_cleanly Timeout (0:00:15)!
Thread 0x00007ff856088340 (most recent call first):
  File "/Users/amo/dev/dlt-contrib/dlt-filesystem/.venv315/lib/python3.15/site-packages/dlt/common/utils.py", line 378 in map_nested_values_in_place
  File "/Users/amo/dev/dlt-contrib/dlt-filesystem/.venv315/lib/python3.15/site-packages/dlt/common/utils.py", line 378 in map_nested_values_in_place
  File "/Users/amo/dev/dlt-contrib/dlt-filesystem/.venv315/lib/python3.15/site-packages/dlt/common/utils.py", line 378 in map_nested_values_in_place
[...]

.... with the main error location message ... map_nested_values_in_place repeated exactly 100 times.

Recap

The software test reads like this.

def test_recursive_alias_raises_cleanly(tmp_path):
    """A self-recursive alias makes a cyclic structure that safe_load accepts but that cannot be
    flattened into a finite record; normalization hits Python's recursion limit and raises a
    bounded RecursionError rather than hanging or emitting corrupt data. Pathological input,
    pinned here as an unsupported-but-safe boundary (not a code-execution or data-leak hole)."""
    path = write_yaml(tmp_path / "cyclic.yaml", "a: &a [*a]\n")
    with pytest.raises(RecursionError):
        list(read_yaml(iter([FileItemStub(path)]), chunksize=10))  # ty: ignore[invalid-argument-type]

Q&A

Is that new behaviour intended? Any idea how to mitigate the problem? Should it be reported to the Python core team as a regression compared to Python 3.14?

/cc @coderabbitai, @hampsterx

#!/bin/bash
set -uo pipefail
ast-grep outline tests/format/test_yaml.py
rg -n -C 12 'def test_recursive_alias_raises_cleanly' tests/format/test_yaml.py
ast-grep outline src/dlt_filesystem/source/format/iterable_codec.py
rg -n -C 15 'map_nested_values_in_place|def _normalize_row|def read_yaml' src/dlt_filesystem/source/format/iterable_codec.py
fd -t f '^(pyproject\.toml|uv\.lock)$'
rg -n -C 4 'name = "dlt"|dlt[>=<~\[]|requires-python' pyproject.toml uv.lock

#!/bin/bash
set -uo pipefail
for issue in 148260 131798; do
  echo "=== CPython issue $issue ==="
  gh api "repos/python/cpython/issues/$issue" --jq '{title,body,state,html_url}'
done
echo '=== dlt traversal implementation ==='
curl -fsSL https://raw.githubusercontent.com/dlt-hub/dlt/devel/dlt/common/utils.py | python -c '
import sys
s=sys.stdin.read().splitlines()
for i,line in enumerate(s):
    if line.startswith("def map_nested_values_in_place("):
        end=next((j for j in range(i+1,len(s)) if s[j].startswith("def ")), min(i+140,len(s)))
        for j in range(max(0,i-2),end): print(f"{j+1}: {s[j]}")
'

@amotl Thanks for the detailed reproduction. This is not expected behavior for ordinary Python recursion. Python 3.15 still documents RecursionError when execution exceeds the recursion limit. Your result warrants investigation as a possible CPython regression. It does not yet identify the interpreter defect. Python documentation

What the traceback establishes

The hang occurs during recursive traversal in map_nested_values_in_place(), not during YAML parsing.

I inspected the upstream implementation. It recursively visits dictionaries, lists, and tuples without cycle detection. It does not catch RecursionError or KeyboardInterrupt. Its recursive calls use *args and **kwargs; preserve that call form when reducing the reproduction.

The 100 repeated frames do not mean that recursion stopped at depth 100. faulthandler limits its output to 100 frames per thread. The actual stack can be deeper. Documentation

Isolate the possible CPython regression

Try this standalone reproduction on both Python 3.14 and 3.15. It removes YAML, dlt, pytest, and coverage while preserving the relevant recursive call form:

import faulthandler
import sys

def walk(value, *args, **kwargs):
    for index, child in enumerate(value):
        if isinstance(child, list):
            value[index] = walk(child, *args, **kwargs)
    return value

print(sys.version, flush=True)
print("recursion limit:", sys.getrecursionlimit(), flush=True)

cycle = []
cycle.append(cycle)

# Exit after dumping the stack, even if normal interruption fails.
# exit=True skips cleanup; use this only in this diagnostic process.
faulthandler.dump_traceback_later(10, exit=True)
try:
    walk(cycle)
except RecursionError:
    print("RecursionError raised as expected", flush=True)
finally:
    faulthandler.cancel_dump_traceback_later()

I have not executed this reproduction.

If it hangs on 3.15 but raises RecursionError on 3.14, report that reduced case to the CPython team. Include:

• Full sys.version output and interpreter installation source.
• Operating system and architecture.
• Whether the build uses free threading or JIT.
• Results with the same script on both versions.
• Whether PYTHON_JIT=0 changes the result, if the build supports JIT.

Disabling JIT is a diagnostic comparison, not a confirmed workaround. I did not find a verified matching CPython issue.

Mitigation in this package

Reject cycles explicitly before calling map_nested_values_in_place() at Line 438 of src/dlt_filesystem/source/format/iterable_codec.py.

Use an iterative traversal with container identities on the active traversal path. Reject a container when its identity already occurs on that path. Do not reject every repeated identity: two fields can legitimately reference the same non-cyclic YAML anchor.

For compatibility with the current test, the guard could raise an explicit RecursionError("Cyclic container in decoded record"). A dedicated exception would provide a clearer contract but would require updating the test.

This removes reliance on the interpreter eventually stopping cyclic traversal. Increasing the recursion limit would not fix the cycle.

The test’s “unsupported-but-safe” description also needs qualification: this reproduction demonstrates an availability risk, although it does not demonstrate code execution or data disclosure.

Learnt from: amotl
URL: https://github.com/dlt-contrib/dlt-filesystem/pull/2

Timestamp: 2026-10-21T23:53:57.491Z
Learning: In dlt-contrib/dlt-filesystem, the maintainer reproduced a Python 3.15 hang in tests/format/test_yaml.py::test_recursive_alias_raises_cleanly for YAML input "a: &a [*a]". The process consumed one CPU core and did not respond to normal Ctrl+C. A faulthandler dump showed repeated recursive calls at Line 378 of dlt.common.utils.map_nested_values_in_place. This confirms a normalization hang in that environment, but does not establish the specific CPython defect or the cause of GitHub runner shutdown.

_{You are interacting with an AI system.}

@hampsterx: Is your clanker able to verify this on Python 3.14 vs. Python 3.15 and derive a corresponding outcome?

@amotl Confirmed locally on Linux x86_64 at 0f7cff8, using dlt 1.31.0 and PyYAML 6.0.3:

• Python 3.14.9: the recursive YAML test passes, with and without coverage.
• Python 3.15.1: it hangs in map_nested_values_in_place() in both configurations.

A dependency-free reproduction also confirms the difference: with the recursion limit set to 1,000, recursive calls forwarding *args/**kwargs reach 3,000 calls on 3.15.1. On 3.14.9 they raise RecursionError. JIT was disabled.

This matches CPython #155151: the specialized CALL_EX_PY path lacks a recursion-limit check. Fix #156346 is open, so a duplicate report isn’t needed.

I suggest retaining 3.14 in the matrix and rejecting cyclic input explicitly before enabling 3.15.

Excellent. Thanks for discovering the corresponding upstream ticket.

I suggest retaining 3.14 in the matrix and rejecting cyclic input explicitly before enabling 3.15.

Thank you. Feel free to submit a relevant patch, optimally stacked on this one.

• Thank you for submitting fix: reject cyclic YAML and CBOR input before normalization #31. 💯