Logs: array attribute values are silently dropped (and diverge from sentry-javascript)

Inizia in sentry_sdk/utils.py, in format_attribute intorno alla riga 2125, e traccia come vengono serializzati gli attributi del logger prima dell’ingestion. Confronta il comportamento di Python con il comportamento di fallback descritto di @sentry/core serializeAttributes. Il lavoro è completato quando i valori non primitivi non scompaiono più silenziosamente e il comportamento di serializzazione o diagnostica scelto è coperto dai controlli esistenti pertinenti.

Environment

sentry-sdk 2.61.0, Python 3.14, logs enabled via enable_logs=True.

What happened

Passing a list-valued attribute to sentry_sdk.logger.* results in the attribute silently disappearing from the Logs UI — no value, no warning, no debug log.

sentry_sdk.logger.info(
    "workflow_engine processed event",
    attributes={
        "workflow_ids": [101, 202, 303],   # list[int]  -> dropped (blank in UI)
        "rule_names": ["high_cpu", "oom"],  # list[str]  -> dropped (blank in UI)
        "organization_id": 42,              # int        -> stored OK
        "mixed_list": [1, "two", 3.0],      # mixed      -> stored as the string "[1, 'two', 3.0]"
        "nested_dict": {"a": [1, 2]},       # dict       -> stored as the string "{'a': [1, 2]}"
    },
)

Querying the log back (EAP logs dataset), the homogeneous scalar lists come back null, while the scalars and the stringified mixed/dict values are present.

Root cause

format_attribute (sentry_sdk/utils.py:2125) preserves a homogeneous scalar list/tuple as a real array, but falls back to safe_repr (a string) for mixed lists and objects:

elif isinstance(val, list):
    ty = type(val[0])
    if ty in (str, int, float, bool) and all(type(v) is ty for v in val):
        return copy.deepcopy(val)   # sent as a real array
...
return safe_repr(val)               # everything else -> string

Sentry's log storage (EAP) only retains scalar attribute types (string / number / boolean), so the real arrays are dropped on ingestion. The net effect is counterintuitive: homogeneous lists vanish, but mixed lists/objects survive (as strings).

Divergence from sentry-javascript

@sentry/core (v10.50) does the opposite for logs — serializeAttributes(logAttributes, true) runs with fallback=true, so every non-primitive (arrays and objects) is JSON.stringify'd into a string attribute and shows up. Its attributes.js even has an explicit comment that it intentionally does not send real arrays yet to avoid a future breaking change. So the same logging API gives different results across SDKs.

Suggested fix

Make Python consistent with JS for now: stringify non-primitive attribute values (instead of sending homogeneous arrays that get dropped), or at minimum emit a debug log / drop-record when an attribute value can't be represented, so the loss isn't silent.

A docs note has been added in getsentry/sentry-docs#17921.

🤖 Generated with Claude Code