From 61226309980ae1ab488930b2c73024c3ca7eb6a6 Mon Sep 17 00:00:00 2001
From: Masha Basmanova <mbasmanova@fb.com>
Date: Tue, 12 Dec 2023 12:45:03 -0800
Subject: [PATCH] Fix OutputBuffer::isOverutilized (#7985)

Summary:
OutputBuffer::isOverutilized() should return true if it is half-full. It used to
return true only when full preventing effective scaling of consumers. The
buffer maintains is-full state only for a short duration of time. Since
producers are blocked, one fetch from one consumer resets the state to
not-null. Thus the coordinator that's monitoring output buffers may never see
buffer getting over-utilized and mitigate by scaling the number of consumers.
This caused some queries to be very slow as they used only 4 threads for
TableWrite.

Pull Request resolved: https://github.com/facebookincubator/velox/pull/7985

Reviewed By: Yuhta

Differential Revision: D52073744

Pulled By: mbasmanova

fbshipit-source-id: 246e065d885ce52601acd4cdcba5e708c5628da5
---
 velox/exec/OutputBuffer.cpp | 2 +-
 velox/exec/OutputBuffer.h   | 5 +++--
 2 files changed, 4 insertions(+), 3 deletions(-)

diff --git a/velox/exec/OutputBuffer.cpp b/velox/exec/OutputBuffer.cpp
index a14fd6313031..7b9bdec21e8a 100644
--- a/velox/exec/OutputBuffer.cpp
+++ b/velox/exec/OutputBuffer.cpp
@@ -637,7 +637,7 @@ double OutputBuffer::getUtilization() const {
 }
 
 bool OutputBuffer::isOverutilized() const {
-  return (totalSize_ > maxSize_) && !atEnd_;
+  return (totalSize_ > (0.5 * maxSize_)) && !atEnd_;
 }
 
 } // namespace facebook::velox::exec
diff --git a/velox/exec/OutputBuffer.h b/velox/exec/OutputBuffer.h
index c7a1eaedeaed..0b6cb5b66198 100644
--- a/velox/exec/OutputBuffer.h
+++ b/velox/exec/OutputBuffer.h
@@ -187,8 +187,9 @@ class OutputBuffer {
   // Gets the memory utilization ratio in this output buffer.
   double getUtilization() const;
 
-  // Indicates if this output buffer is over-utilized and thus blocks its
-  // producers.
+  // Indicates if this output buffer is over-utilized, i.e. at least half full,
+  // and will start blocking producers soon. This is used to dynamically scale
+  // the number of consumers, for example, increase number of TableWriter tasks.
   bool isOverutilized() const;
 
  private: