Randomly select exponent

src/bgw/job_stat.c

@@ -199,8 +199,19 @@ calculate_next_start_on_failure(TimestampTz finish_time, int consecutive_failure
 	TimestampTz res = 0;
 	volatile bool res_set = false;
 	TimestampTz last_finish = finish_time;
-	float8 exponent = (consecutive_failures > MAX_FAILURES_MULTIPLIER ? MAX_FAILURES_MULTIPLIER :
+	/* the aim of backoff is to reduce conflicts in the case of jobs that fail to launch, 
+	 * as there are jobs competing for available background workers in that case. 
+	 * therefore we introduce some randomness in backoff calculation for that type of failure
+	 * by randomly selecting the exponent in the interval [0, consecutive_failures]
+	 * Otherwise as soon as a worker becomes available, they'll all try to grab it, but only 1 will anyway
+	 * so why should they all try at once? so jobs that experience failures will not only delay
+	 * their next_start but they will also delay it in a non-deterministic way
+	 * to reduce the probability of collisions
+	 */
+
+	int exponent = (consecutive_failures > MAX_FAILURES_MULTIPLIER ? MAX_FAILURES_MULTIPLIER :
 																		consecutive_failures);
+	int rand_exponent = random() % exponent + 1;
 	MemoryContext oldctx;
 
 	if (!IS_VALID_TIMESTAMP(finish_time))
@@ -228,7 +239,7 @@ calculate_next_start_on_failure(TimestampTz finish_time, int consecutive_failure
 
 		/* ival = retry_period ^ (consecutive_failures - 1)  */
 		/* arbitrarily choose 2 for exponential growth */
-		for (i = 0; i < exponent - 1; i++)
+		for (i = 0; i < (job == NULL) ? rand_exponent : exponent - 1; i++)
 		{
 			ival = DirectFunctionCall2(interval_mul, ival, Float8GetDatum(2));
 		}