One NaN or inf value silently drops the whole batch of logs, spans or metrics #8066

How do you use Sentry?

Sentry Saas (sentry.io)

Version

2.72.0 (latest on PyPI at the time of writing; also on master 0f7acd5)

Steps to Reproduce

Self-contained, only needs pip install sentry-sdk. It starts a local DSN sink on 127.0.0.1 and
counts the log records it receives. A plain stdlib logging call with a NaN argument is enough.

# Repro: one NaN attribute value drops the whole batch of logs
import json, logging, sys, threading
from http.server import BaseHTTPRequestHandler, HTTPServer

import sentry_sdk

received = []

class Sink(BaseHTTPRequestHandler):
    def log_message(self, *a): pass
    def do_POST(self):
        body = self.rfile.read(int(self.headers["Content-Length"]))
        import gzip
        if self.headers.get("Content-Encoding") == "gzip": body = gzip.decompress(body)
        lines = body.split(b"\n")
        for ln in lines[1:]:
            try:
                d = json.loads(ln)
            except Exception:
                continue
            if "items" in d: received.append(len(d["items"]))
        self.send_response(200); self.send_header("Content-Length", "2"); self.end_headers(); self.wfile.write(b"{}")

srv = HTTPServer(("127.0.0.1", 0), Sink)
threading.Thread(target=srv.serve_forever, daemon=True).start()

value = float("nan") if sys.argv[1] == "nan" else 0.5
sentry_sdk.init(dsn="http://k@127.0.0.1:%d/1" % srv.server_port, enable_logs=True)

for i in range(10):
    sentry_sdk.logger.info("innocent log %d" % i)
log = logging.getLogger("app"); log.setLevel(logging.INFO)
log.info("loss=%s", value)                                 # plain stdlib logging call, NaN as a %s argument
sentry_sdk.flush()
print(sys.argv[1], "-> log records delivered:", sum(received), "(sent 11)")

Run python repro.py ok and python repro.py nan.

Expected Result

The 10 innocent logs are delivered. At most the one record with the non-finite value is dropped
(or the value is sent as a string), and the drop is recorded in a client report.

Actual Result

ok  -> log records delivered: 11 (sent 11)
nan -> log records delivered: 0 (sent 11)

All 11 logs are lost, with no client report; only debug=True shows the error. The same happens
with span streaming: one span with a NaN attribute loses the whole trace (6 of 6 spans in my
test), and with buffered metrics: one sentry_sdk.metrics.gauge("loss", float("nan")) next to
10 counters, and 0 of 11 metrics arrive (11 of 11 with 0.5).

Cause: json_dumps in sentry_sdk/utils.py:125 uses allow_nan=False. format_attribute
(utils.py:2159) keeps non-finite floats as they are, and serialization only happens later in
the transport worker, so the ValueError loses the whole envelope. The event path already
handles this in serializer.py:286.

Possible fix: turn non-finite floats into strings (or drop that one attribute) in
format_attribute, and record a client report if an item still fails to serialize. Happy to
open a PR if that sounds right.