Fix cardinality memory-usage considerations.

Default precision was computed based on the number of MULTI_BUCKET parents
instead of PER_BUCKET.

The ordinals-based execution mode was almost always used although ordinals
might have non-negligible memory usage compared to the counters.

Close #5452

src/main/java/org/elasticsearch/search/aggregations/metrics/cardinality/CardinalityAggregator.java

@@ -23,6 +23,7 @@ import com.google.common.base.Preconditions;
 import org.apache.lucene.index.AtomicReaderContext;
 import org.apache.lucene.util.BytesRef;
 import org.apache.lucene.util.FixedBitSet;
+import org.apache.lucene.util.RamUsageEstimator;
 import org.elasticsearch.ElasticsearchException;
 import org.elasticsearch.common.Nullable;
 import org.elasticsearch.common.lease.Releasable;
@@ -93,7 +94,10 @@ public class CardinalityAggregator extends MetricsAggregator.SingleValue {
         if (bytesValues instanceof BytesValues.WithOrdinals) {
             BytesValues.WithOrdinals values = (BytesValues.WithOrdinals) bytesValues;
             final long maxOrd = values.ordinals().getMaxOrd();
-            if (maxOrd <= reader.reader().maxDoc()) {
+            final long ordinalsMemoryUsage = OrdinalsCollector.memoryOverhead(maxOrd);
+            final long countsMemoryUsage = HyperLogLogPlusPlus.memoryUsage(precision);
+            // only use ordinals if they don't increase memory usage by more than 25%
+            if (ordinalsMemoryUsage < countsMemoryUsage / 4) {
                 return new OrdinalsCollector(counts, values, bigArrays);
             }
         }
@@ -195,6 +199,15 @@ public class CardinalityAggregator extends MetricsAggregator.SingleValue {
 
     private static class OrdinalsCollector implements Collector {
 
+        private static final long SHALLOW_FIXEDBITSET_SIZE = RamUsageEstimator.shallowSizeOfInstance(FixedBitSet.class);
+
+        /**
+         * Return an approximate memory overhead per bucket for this collector.
+         */
+        public static long memoryOverhead(long maxOrd) {
+            return RamUsageEstimator.NUM_BYTES_OBJECT_REF + SHALLOW_FIXEDBITSET_SIZE + (maxOrd + 7) / 8; // 1 bit per ord
+        }
+
         private final BigArrays bigArrays;
         private final BytesValues.WithOrdinals values;
         private final Ordinals.Docs ordinals;

src/main/java/org/elasticsearch/search/aggregations/metrics/cardinality/CardinalityAggregatorFactory.java

@@ -64,8 +64,8 @@ final class CardinalityAggregatorFactory extends ValueSourceAggregatorFactory<Va
     private int defaultPrecision(Aggregator parent) {
         int precision = HyperLogLogPlusPlus.DEFAULT_PRECISION;
         while (parent != null) {
-            if (parent.bucketAggregationMode() == BucketAggregationMode.MULTI_BUCKETS) {
+            if (parent.bucketAggregationMode() == BucketAggregationMode.PER_BUCKET) {
-                // if the parent is a multi-bucket aggregator, we substract 5 to the precision,
+                // if the parent is a per-bucket aggregator, we substract 5 to the precision,
                 // which will effectively divide the memory usage of each counter by 32
                 precision -= 5;
             }

src/main/java/org/elasticsearch/search/aggregations/metrics/cardinality/HyperLogLogPlusPlus.java

@@ -75,6 +75,13 @@ public final class HyperLogLogPlusPlus implements Releasable {
         return precision;
     }
 
+    /**
+     * Return the expected per-bucket memory usage for the given precision.
+     */
+    public static long memoryUsage(int precision) {
+        return 1L << precision;
+    }
+
     // these static tables come from the appendix of the paper
     private static final double[][] RAW_ESTIMATE_DATA = {
         // precision 4