bytes_sent and bytes_dropped_writer telemetry count characters, not bytes #983

Describe the bug

datadog.dogstatsd.client.bytes_sent and bytes_dropped_writer count characters, not bytes. datadog/dogstatsd/base.py:1791 does self.bytes_sent += len(packet) on the str, while the bytes actually sent are encoded_packet — computed nine lines above at :1782 and handed to mysocket.send. Same at :1662, :1666 and :1734.

:1648 in the same file is byte-correct (len(packet_with_newline.encode(self.encoding))), from #927 "fix: report the right number of dropped bytes". That PR fixed bytes_dropped_queue and did not touch the other four sites.

To Reproduce

Point a client at a real UDP socket, send a metric with a multibyte tag, and compare the bytes received against the value the client then self-reports:

statsd.gauge("page.views", 1, tags=[u"city:montréal"])

datagram             : b'page.views:1|g|#city:montr\xc3\xa9al\n'
bytes actually sent  : 31
client reports       : 30

Over five cases (datadog/dogstatsd at fac2d5e, Python 3.13):

Expected behavior

The counters report the number of bytes. DataDog/datadog-go statsd/telemetry.go:73 documents TotalBytesSent as "the total number of bytes succesfully sent by the client", and feeds it to the same datadog.dogstatsd.client.bytes_sent metric at :290; being Go, its lengths are byte counts by construction.

Environment and Versions

• datadogpy master @ fac2d5e, datadog.__version__ 0.53.0
• Python 3.13.12, macOS

Additional context

The reason the suite is green: assert_equal_telemetry (tests/unit/dogstatsd/test_statsd.py:172) builds its default expectation as telemetry_metrics(bytes_sent=len(expected_payload)) — the same character count the implementation uses — and bytes_sent=len( appears 39 times in that file, never once with .encode(...). test_event_with_container_field is the clearest case: at :2567 it asserts the event header as _e{5,32} for a 19-character, 32-byte text, so the byte length is written into the same assertion, and then asserts bytes_sent=len(event3) seven lines later.

Found by checking datadogpy's DogStatsD wire handling field by field against datadog-go, not from a production incident. Written with Claude Code. Happy to open a PR.