From 82cecd48556e039590672244db0adcbebc91f9d4 Mon Sep 17 00:00:00 2001
From: Abhishek Ranjan <aranjans@google.com>
Date: Sun, 24 Nov 2024 18:32:19 +0530
Subject: [PATCH] update Retry backoff to follow changes made to gRFC A6

---
 examples/features/retry/README.md | 18 ++++++++++++++++++
 stream.go                         | 13 ++++++++-----
 2 files changed, 26 insertions(+), 5 deletions(-)

diff --git a/examples/features/retry/README.md b/examples/features/retry/README.md
index b725048a3862..5056a0279644 100644
--- a/examples/features/retry/README.md
+++ b/examples/features/retry/README.md
@@ -54,6 +54,24 @@ RetryableStatusCodes: Retry only when receiving these status codes.
         }`
 ```
 
+### Backoff Logic
+
+The backoff duration is calculated based on the following logic:
+
+1. **Initial Backoff**: The initial delay before the first retry.
+2. **Exponential Backoff**: The delay increases exponentially based on the number of retries, using the formula:
+   ```go
+   backoffDuration := InitialBackoff * time.Duration(math.Pow(BackoffMultiplier, float64(numRetries)))
+   ```
+3. **Max Backoff**: The calculated backoff duration is capped at `MaxBackoff` to prevent excessively long delays.
+4. **Jitter**: A random factor between 0.8 and 1.2 is applied to the backoff duration to avoid thundering herd problems. The final backoff duration is calculated as:
+   ```go
+   jitter := 0.8 + rand.Float64()*0.4 // Random value between 0.8 and 1.2
+   dur = time.Duration(float64(backoffDuration) * jitter)
+   ```
+
+This means that the backoff delay may be slightly lower than `InitialBackoff` or slightly higher than `MaxBackoff`, allowing for a more distributed retry behavior across clients.
+
 ### Providing the retry policy as a DialOption
 
 To use the above service config, pass it with `grpc.WithDefaultServiceConfig` to
diff --git a/stream.go b/stream.go
index 6d10d0ac8713..169087c7c708 100644
--- a/stream.go
+++ b/stream.go
@@ -707,12 +707,15 @@ func (a *csAttempt) shouldRetry(err error) (bool, error) {
 		dur = time.Millisecond * time.Duration(pushback)
 		cs.numRetriesSincePushback = 0
 	} else {
-		fact := math.Pow(rp.BackoffMultiplier, float64(cs.numRetriesSincePushback))
-		cur := float64(rp.InitialBackoff) * fact
-		if max := float64(rp.MaxBackoff); cur > max {
-			cur = max
+		// Calculate the backoff duration based on the exponential backoff formula.
+		backoffDuration := rp.InitialBackoff * time.Duration(math.Pow(rp.BackoffMultiplier, float64(cs.numRetriesSincePushback)))
+		if backoffDuration > rp.MaxBackoff {
+			backoffDuration = rp.MaxBackoff
 		}
-		dur = time.Duration(rand.Int64N(int64(cur)))
+
+		// Apply jitter by multiplying with a random factor between 0.8 and 1.2.
+		jitter := 0.8 + rand.Float64()*0.4
+		dur = time.Duration(float64(backoffDuration) * jitter)
 		cs.numRetriesSincePushback++
 	}