Reduce write thread pool EWMA alpha (elastic#112206)

This helps reduce the impact of single long running tasks on the EWMA,
which is used for interpreting the queue contribution to ingestion load
in ingest autoscaling.

Author: pxsalehi

server/src/main/java/org/elasticsearch/threadpool/DefaultBuiltInExecutorBuilders.java

@@ -18,6 +18,7 @@
 import java.util.Map;
 
 import static java.util.Collections.unmodifiableMap;
+import static org.elasticsearch.threadpool.ThreadPool.indexAutoscalingEWMA;
 import static org.elasticsearch.threadpool.ThreadPool.searchAutoscalingEWMA;
 
 public class DefaultBuiltInExecutorBuilders implements BuiltInExecutorBuilders {
@@ -41,7 +42,7 @@ public Map<String, ExecutorBuilder> getBuilders(Settings settings, int allocated
                 ThreadPool.Names.WRITE,
                 allocatedProcessors,
                 10000,
-                new EsExecutors.TaskTrackingConfig(true, 0.1)
+                new EsExecutors.TaskTrackingConfig(true, indexAutoscalingEWMA)
             )
         );
         int searchOrGetThreadPoolSize = ThreadPool.searchOrGetThreadPoolSize(allocatedProcessors);
@@ -187,7 +188,7 @@ public Map<String, ExecutorBuilder> getBuilders(Settings settings, int allocated
                 ThreadPool.Names.SYSTEM_WRITE,
                 halfProcMaxAt5,
                 1000,
-                new EsExecutors.TaskTrackingConfig(true, 0.1)
+                new EsExecutors.TaskTrackingConfig(true, indexAutoscalingEWMA)
             )
         );
         result.put(
@@ -207,7 +208,7 @@ public Map<String, ExecutorBuilder> getBuilders(Settings settings, int allocated
                 ThreadPool.Names.SYSTEM_CRITICAL_WRITE,
                 halfProcMaxAt5,
                 1500,
-                new EsExecutors.TaskTrackingConfig(true, 0.1)
+                new EsExecutors.TaskTrackingConfig(true, indexAutoscalingEWMA)
             )
         );
         return unmodifiableMap(result);

server/src/main/java/org/elasticsearch/threadpool/ThreadPool.java

@@ -178,6 +178,13 @@ public static ThreadPoolType fromType(String type) {
 
     public static final double searchAutoscalingEWMA = 0.1;
 
+    // This value is chosen such that a sudden increase in the task durations would need to persist roughly for 120 samples
+    // for the EWMA value to be mostly representative of the increased task durations. Mostly representative means that the
+    // EWMA value is at least within 90% of the new increased task duration. This value also determines the impact of a single
+    // long-running task on the moving average and limits it roughly to 2% of the (long) task duration, e.g. if the current
+    // moving average is 100ms, and we get one task which takes 20s the new EWMA will be ~500ms.
+    public static final double indexAutoscalingEWMA = 0.02;
+
     private final Map<String, ExecutorHolder> executors;
 
     private final ThreadPoolInfo threadPoolInfo;