Not a bug report — flagging a compatibility edge the 1.45.0 changelog entry
didn't call out, in case it's worth a line in the migration notes for anyone
who greps for it later.
If a bytes type is set as an attribute value in the SDK, it will no longer
be utf-8 decoded to a string, instead it will be passed along as is in
accordance with the OTEL spec.
Before 1.45.0, Span.set_attribute(key, some_bytes) silently coerced
some_bytes to str via utf-8 decode. After 1.45.0, the same call now
exports AnyValue.bytes_value on the wire. That's the documented, correct,
spec-compliant behaviour, and I'm not asking for it to change.
What I ran into: at least one popular instrumentation library
(langchain-ai/langsmith-sdk's OTel exporter — reported separately, langchain-ai/langsmith-sdk#3602)
has, across every version I checked, called set_attribute with a bytes
value for its main prompt/completion payload, apparently unintentionally,
relying on the old SDK behaviour to turn it back into readable text. Upgrading
only opentelemetry-sdk/opentelemetry-api (no code change, no langsmith
version change) silently turned every one of those attributes from a JSON
string into an opaque binary blob for anyone reading the exported traces —
no error, no warning, same span count, same attribute keys, just a different
wire type carrying the same bytes.
Nothing to fix here since the change is correct per spec, and it's the
instrumentation library's bug for passing bytes when it meant str. Just
noting it in case it's useful to link from the 1.45.0 changelog entry or the
migration guide, since this specific failure mode (payload silently becomes
unreadable, no exception anywhere) seems like exactly the kind of thing that's
hard for an instrumentation-library author to notice in their own tests if
they only assert on in/substring checks against the in-memory exporter
rather than the attribute's actual type.
Not a bug report — flagging a compatibility edge the 1.45.0 changelog entry
didn't call out, in case it's worth a line in the migration notes for anyone
who greps for it later.
Before 1.45.0,
Span.set_attribute(key, some_bytes)silently coercedsome_bytestostrvia utf-8 decode. After 1.45.0, the same call nowexports
AnyValue.bytes_valueon the wire. That's the documented, correct,spec-compliant behaviour, and I'm not asking for it to change.
What I ran into: at least one popular instrumentation library
(
langchain-ai/langsmith-sdk's OTel exporter — reported separately, langchain-ai/langsmith-sdk#3602)has, across every version I checked, called
set_attributewith abytesvalue for its main prompt/completion payload, apparently unintentionally,
relying on the old SDK behaviour to turn it back into readable text. Upgrading
only
opentelemetry-sdk/opentelemetry-api(no code change, no langsmithversion change) silently turned every one of those attributes from a JSON
string into an opaque binary blob for anyone reading the exported traces —
no error, no warning, same span count, same attribute keys, just a different
wire type carrying the same bytes.
Nothing to fix here since the change is correct per spec, and it's the
instrumentation library's bug for passing
byteswhen it meantstr. Justnoting it in case it's useful to link from the 1.45.0 changelog entry or the
migration guide, since this specific failure mode (payload silently becomes
unreadable, no exception anywhere) seems like exactly the kind of thing that's
hard for an instrumentation-library author to notice in their own tests if
they only assert on
in/substring checks against the in-memory exporterrather than the attribute's actual type.