Tornado instrumentation closes spans before on_finish is completed.

[pitabwire]

Ideally all trace spans in a request response cycle should be contained in one span. However because the tornado instrumentation library closes the span before calling the on_finish method. Orphaned spans from any instrumented code in on_finish method are created. Looking at the code a reordering would fix this situation.

Describe your environment

Python 3.8.5

requirements.txt
tornado==6.0.3

#tracing
opentelemetry-api==1.1.0
opentelemetry-sdk==1.1.0
opentelemetry-exporter-zipkin-json
opentelemetry-instrumentation
opentelemetry-instrumentation-dbapi
opentelemetry-instrumentation-tornado
opentelemetry-exporter-gcp-trace==1.0.0rc0

Steps to reproduce
Run the attached app below:

`

import tornado.ioloop
import tornado.web
from opentelemetry import trace
from opentelemetry.trace import SpanKind

from opentelemetry import trace
from opentelemetry.exporter.cloud_trace import CloudTraceSpanExporter
from opentelemetry.exporter.zipkin.json import ZipkinExporter
from opentelemetry.sdk.resources import SERVICE_NAME, Resource
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import (
ConsoleSpanExporter,
BatchSpanProcessor
)
from opentelemetry.instrumentation.tornado import TornadoInstrumentor

class MainHandler(tornado.web.RequestHandler):

def on_finish(self) -> None:
    with trace.get_tracer(__name__).start_as_current_span("On finish test", kind=SpanKind.CLIENT) as span:
        print("Testing on finish span")

    super(MainHandler, self).on_finish()

def get(self):
    self.write("Hello, world")

def init_tracing(is_deployment):
resource = Resource.create(attributes={SERVICE_NAME: "Profile Service"})
tracer_provider = TracerProvider(resource=resource)
trace.set_tracer_provider(tracer_provider)

if is_deployment:
    span_exporter = CloudTraceSpanExporter()
else:
    # span_exporter = ZipkinExporter()
    span_exporter = ConsoleSpanExporter()

span_processor = BatchSpanProcessor(span_exporter=span_exporter)
trace.get_tracer_provider().add_span_processor(span_processor)

TornadoInstrumentor().instrument()

def make_app():
return tornado.web.Application([
(r"/", MainHandler),
])

if name == "main":

init_tracing(is_deployment=False)
app = make_app()
app.listen(8888)
tornado.ioloop.IOLoop.current().start()

`

What is the expected behavior?
All traces in a request cycle should be contained in one span. However because the tornado instrumentation library closes the span before calling the on_finish method. This causes an orphaned span to be created.

What is the actual behavior?
Tornado instrumentation library should possibly run on_finish method before closing the main trace span

Additional context
I looked at the code and saw the affected part is here . Is there any reason to close the span before the parent on_finish method can be run?