From 64228fd838a694bc2a33747c379feca9358ad490 Mon Sep 17 00:00:00 2001
From: Penporn Koanantakool <38085909+penpornk@users.noreply.github.com>
Date: Wed, 1 Sep 2021 18:21:19 -0700
Subject: [PATCH] Update 20210513-pluggable-profiler-for-tensorflow.md

Clarify TensorFlow's Start()/Stop() call behavior.
---
 rfcs/20210513-pluggable-profiler-for-tensorflow.md | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/rfcs/20210513-pluggable-profiler-for-tensorflow.md b/rfcs/20210513-pluggable-profiler-for-tensorflow.md
index 56104e110..70708b45f 100644
--- a/rfcs/20210513-pluggable-profiler-for-tensorflow.md
+++ b/rfcs/20210513-pluggable-profiler-for-tensorflow.md
@@ -390,6 +390,9 @@ Repeated profiling requests should not increase resource utilization over time.
 * No memory corruption: profiling should not corrupt memory. e.g. due to dangling pointers.
 * No deadlocks: any synchronization necessary to start/stop profiling should not block any application (TF) thread for a long time.
 
+### Assumptions
+Profiler libraries can assume that each time TensorFlow calls a `Start()`, it will call a `Stop()` next, i.e., no multiple consecutive `Start()` calls.
+
 ## **Alternatives Considered**
 
 ## **Performance Implications**