TensorFlow v2: Graph does not appear on TensorBoard

A graph may not appear if TensorFlow could not trace any graph during the execution. For instance, below code does not use @tf.function and, because it is pure eager, there is no graph TensorFlow used.

# Missing @tf.function
def foo(x, y):
  return tf.linalg.cross(x, y)

tf.summary.trace_on()
foo([1,2], [3,4])
with writer.as_default():
  tf.summary.trace_export()

Other known issues

• When a function is invoked before trace, the graph does not show up

@tf.function
def foo(x, y):
  return tf.linalg.cross(x, y)

foo([1,2], [3,4]) # this is not traced.
tf.summary.trace_on()
foo([1,2], [3,4])
with writer.as_default():
  tf.summary.trace_export()

Start by reproducing the two Python snippets in the issue, focusing on the interactions between tf.summary.trace_on(), tf.summary.trace_export(), and @tf.function. Inspect the trace collection and export entry points named by the examples. Done means graphs appear in TensorBoard for the reported invocation orders, or the unsupported behavior is clearly documented.