Tracer doesn't flush implicitly on shutdown.

[feldentm-SAP]

Description

Currently, shutdown of a tracer requires this wrapper:

// Shutdown will flush any remaining spans and shut down the exporter.
	return tracerProvider.Shutdown
	return func(ctx context.Context) error {
		// apparently, an explicit force flush is required to be sure that buffers are flushed
		tracerProvider.ForceFlush(ctx)
		return tracerProvider.Shutdown(ctx)
	}

Our naive assumption was that Shutdown would also ForceFlush implicitly.
We classify this as a bug, because the result is that even with proper calls to Shutdown, buffered spans will be lost.

Environment

• OS: linux
• Architecture: any
• Go Version: go 1.25.1
• opentelemetry-go version: go.opentelemetry.io/otel/sdk v1.38.0

Steps To Reproduce

Trying to trace a kubernetes job with defer.End() on a span in main had a 100% success rate for both the issue and the fix.

Expected behavior

Shutdown should call ForceFlush but continue in case of errors, because on EOL for a job, the best you can do is try to flush everything and close all connections.

[dmathieu]

Are you using a custom span exporter?
Shutdown() on the tracer provider calls the same method on each span processor, which calls it on each span exporter.

https://pkg.go.dev/go.opentelemetry.io/otel/sdk/trace#SpanExporter

// Shutdown notifies the exporter of a pending halt to operations. The
// exporter is expected to perform any cleanup or synchronization it
// requires while honoring all timeouts and cancellations contained in
// the passed context.

So calling shutdown on the span exporter should have a behavior similar to ForceFlush, which is to send everything it needs to send before stopping itself.

Calling ForceFlush, and then Shutdown is not safe, as there could be new data emitted between the two calls, and if Shutdown doesn't ensure everything is sent, it may lose data.

[feldentm-SAP]

The span exporter is "go.opentelemetry.io/otel/exporters/otlp/otlptrace/otlptracegrpc".

The observation regarding safety is correct. However, our expectation is that if you end a span in the same thread/go routine that will also call Shutdown afterwards, the span should always be sent.

[dmathieu]

That's what the batch span processor does.
On shutdown, it flushes the queue and then shuts down the exporter and itself.
https://github.com/open-telemetry/opentelemetry-go/blob/sdk/v1.38.0/sdk/trace/batch_span_processor.go#L224-L250