[improve][broker] Make maxBatchDeletedIndexToPersist configurable and document other related configs (apache#24392)

(cherry picked from commit d1bca65)

Author: lhotari

conf/broker.conf

@@ -1228,6 +1228,14 @@ managedLedgerCursorMaxEntriesPerLedger=50000
 # Max time before triggering a rollover on a cursor ledger
 managedLedgerCursorRolloverTimeInSeconds=14400
 
+# Maximum amount of memory used hold data read from storage (or from the cache).
+# This mechanism prevents the broker to have too many concurrent
+# reads from storage and fall into Out of Memory errors in case
+# of multiple concurrent reads to multiple concurrent consumers.
+# Set 0 in order to disable the feature.
+#
+managedLedgerMaxReadsInFlightSizeInMB=0
+
 # Max number of "acknowledgment holes" that are going to be persistently stored.
 # When acknowledging out of order, a consumer will leave holes that are supposed
 # to be quickly filled by acking all the messages. The information of which
@@ -1237,13 +1245,22 @@ managedLedgerCursorRolloverTimeInSeconds=14400
 # crashes.
 managedLedgerMaxUnackedRangesToPersist=10000
 
-# Maximum amount of memory used hold data read from storage (or from the cache).
-# This mechanism prevents the broker to have too many concurrent
-# reads from storage and fall into Out of Memory errors in case
-# of multiple concurrent reads to multiple concurrent consumers.
-# Set 0 in order to disable the feature.
-#
-managedLedgerMaxReadsInFlightSizeInMB=0
+# Maximum number of partially acknowledged batch messages per subscription that will have their batch
+# deleted indexes persisted. Batch deleted index state is handled when acknowledgmentAtBatchIndexLevelEnabled=true.
+# When this limit is exceeded, remaining batch message containing the batch deleted indexes will
+# only be tracked in memory. In case of broker restarts or load balancing events, the batch
+# deleted indexes will be cleared while redelivering the messages to consumers.
+managedLedgerMaxBatchDeletedIndexToPersist=10000
+
+# When storing acknowledgement state, choose a more compact serialization format that stores
+# individual acknowledgements as a bitmap which is serialized to an array of long values. NOTE: This setting requires
+# managedLedgerUnackedRangesOpenCacheSetEnabled=true to be effective.
+managedLedgerPersistIndividualAckAsLongArray=false
+
+# When set to true, a BitSet will be used to track acknowledged messages that come after the "mark delete position"
+# for each subscription. RoaringBitmap is used as a memory efficient BitSet implementation for the acknowledged
+# messages tracking. Unacknowledged ranges are the message ranges excluding the acknowledged messages.
+managedLedgerUnackedRangesOpenCacheSetEnabled=true
 
 # Max number of "acknowledgment holes" that can be stored in MetadataStore. If number of unack message range is higher
 # than this limit then broker will persist unacked ranges into bookkeeper to avoid additional data overhead into
@@ -1696,9 +1713,6 @@ narExtractionDirectory=
 # Maximum prefetch rounds for ledger reading for offloading
 managedLedgerOffloadPrefetchRounds=1
 
-# Use Open Range-Set to cache unacked messages
-managedLedgerUnackedRangesOpenCacheSetEnabled=true
-
 # For Amazon S3 ledger offload, AWS region
 s3ManagedLedgerOffloadRegion=
 

conf/standalone.conf

@@ -791,6 +791,23 @@ managedLedgerMaxSizePerLedgerMbytes=2048
 # crashes.
 managedLedgerMaxUnackedRangesToPersist=10000
 
+# Maximum number of partially acknowledged batch messages per subscription that will have their batch
+# deleted indexes persisted. Batch deleted index state is handled when acknowledgmentAtBatchIndexLevelEnabled=true.
+# When this limit is exceeded, remaining batch message containing the batch deleted indexes will
+# only be tracked in memory. In case of broker restarts or load balancing events, the batch
+# deleted indexes will be cleared while redelivering the messages to consumers.
+managedLedgerMaxBatchDeletedIndexToPersist=10000
+
+# When storing acknowledgement state, choose a more compact serialization format that stores
+# individual acknowledgements as a bitmap which is serialized to an array of long values. NOTE: This setting requires
+# managedLedgerUnackedRangesOpenCacheSetEnabled=true to be effective.
+managedLedgerPersistIndividualAckAsLongArray=false
+
+# When set to true, a BitSet will be used to track acknowledged messages that come after the "mark delete position"
+# for each subscription. RoaringBitmap is used as a memory efficient BitSet implementation for the acknowledged
+# messages tracking. Unacknowledged ranges are the message ranges excluding the acknowledged messages.
+managedLedgerUnackedRangesOpenCacheSetEnabled=true
+
 # Max number of "acknowledgment holes" that can be stored in MetadataStore. If number of unack message range is higher
 # than this limit then broker will persist unacked ranges into bookkeeper to avoid additional data overhead into
 # MetadataStore.
@@ -825,9 +842,6 @@ managedLedgerMinimumBacklogEntriesForCaching=1000
 # Maximum backlog entry difference to prevent caching entries that can't be reused.
 managedLedgerMaxBacklogBetweenCursorsForCaching=1000
 
-# Use Open Range-Set to cache unacked messages
-managedLedgerUnackedRangesOpenCacheSetEnabled=true
-
 # Managed ledger prometheus stats latency rollover seconds (default: 60s)
 managedLedgerPrometheusStatsLatencyRolloverSeconds=60
 
@@ -1296,3 +1310,4 @@ delayedDeliveryMaxNumBuckets=-1
 
 # Whether retain null-key message during topic compaction
 topicCompactionRetainNullKey=true
+

managed-ledger/src/main/java/org/apache/bookkeeper/mledger/ManagedLedgerConfig.java

@@ -487,6 +487,16 @@ public int getMaxBatchDeletedIndexToPersist() {
         return maxBatchDeletedIndexToPersist;
     }
 
+    /**
+     * Set max batch deleted index that will be persisted and recovered.
+     *
+     * @param maxBatchDeletedIndexToPersist
+     *            max batch deleted index that will be persisted and recovered.
+     */
+    public void setMaxBatchDeletedIndexToPersist(int maxBatchDeletedIndexToPersist) {
+        this.maxBatchDeletedIndexToPersist = maxBatchDeletedIndexToPersist;
+    }
+
     public boolean isPersistentUnackedRangesWithMultipleEntriesEnabled() {
         return persistentUnackedRangesWithMultipleEntriesEnabled;
     }

pulsar-broker-common/src/main/java/org/apache/pulsar/broker/ServiceConfiguration.java

@@ -2179,10 +2179,22 @@ The delayed message index time step(in seconds) in per bucket snapshot segment,
             + " will only be tracked in memory and messages will be redelivered in case of"
             + " crashes.")
     private int managedLedgerMaxUnackedRangesToPersist = 10000;
-    @FieldContext(
-            category = CATEGORY_STORAGE_ML,
-            doc = "Whether persist cursor ack stats as long arrays, which will compress the data and reduce GC rate")
+
+    @FieldContext(category = CATEGORY_STORAGE_ML,
+            doc = "Maximum number of partially acknowledged batch messages per subscription that will have their batch "
+                + "deleted indexes persisted. Batch deleted index state is handled when "
+                + "acknowledgmentAtBatchIndexLevelEnabled=true.\n\n"
+                + "When this limit is exceeded, remaining batch message containing the batch deleted indexes will "
+                + "only be tracked in memory. In case of broker restarts or load balancing events, the batch "
+                + "deleted indexes will be cleared while redelivering the messages to consumers.")
+    private int managedLedgerMaxBatchDeletedIndexToPersist = 10000;
+
+    @FieldContext(category = CATEGORY_STORAGE_ML,
+            doc = "When storing acknowledgement state, choose a more compact serialization format that stores"
+                    + " individual acknowledgements as a bitmap which is serialized to an array of long values.\n\n"
+                    + "NOTE: This setting requires managedLedgerUnackedRangesOpenCacheSetEnabled=true to be effective.")
     private boolean managedLedgerPersistIndividualAckAsLongArray = false;
+
     @FieldContext(
         category = CATEGORY_STORAGE_ML,
         doc = "If enabled, the maximum \"acknowledgment holes\" will not be limited and \"acknowledgment holes\" "
@@ -2205,8 +2217,10 @@ The delayed message index time step(in seconds) in per bucket snapshot segment,
     private int managedLedgerMaxUnackedRangesToPersistInMetadataStore = 1000;
     @FieldContext(
             category = CATEGORY_STORAGE_OFFLOADING,
-            doc = "Use Open Range-Set to cache unacked messages (it is memory efficient but it can take more cpu)"
-        )
+            doc = "When set to true, a BitSet will be used to track acknowledged messages that come after the \"mark "
+                    + "delete position\" for each subscription.\n\nRoaringBitmap is used as a memory efficient BitSet "
+                    + "implementation for the acknowledged messages tracking. Unacknowledged ranges are the message "
+                    + "ranges excluding the acknowledged messages.")
     private boolean managedLedgerUnackedRangesOpenCacheSetEnabled = true;
     @FieldContext(
         dynamic = true,

pulsar-broker/src/main/java/org/apache/pulsar/broker/service/BrokerService.java

@@ -2004,6 +2004,8 @@ public CompletableFuture<ManagedLedgerConfig> getManagedLedgerConfig(TopicName t
 
             managedLedgerConfig
                     .setMaxUnackedRangesToPersist(serviceConfig.getManagedLedgerMaxUnackedRangesToPersist());
+            managedLedgerConfig.setMaxBatchDeletedIndexToPersist(
+                    serviceConfig.getManagedLedgerMaxBatchDeletedIndexToPersist());
             managedLedgerConfig
                     .setPersistIndividualAckAsLongArray(serviceConfig.isManagedLedgerPersistIndividualAckAsLongArray());
             managedLedgerConfig.setPersistentUnackedRangesWithMultipleEntriesEnabled(