This meta issue tracks known issues with scaling clusters to large numbers of shards.

• Security

  • Improve Authorization performance in clusters with a large number of indices #67987
    • Authorization for Internal Requests like Stats or Rollover is Slow in Large Clusters #79632

• General

  • Push back on excessive requests for stats #51992
  • Add pagination to diagnostic APIs #87555
  • Make Cache More Memory Efficient #77546
  • Cluster Stats API Slows down Considerably for Larger Clusters #79563
  • Reduce merging in PersistedClusterStateService #79793
  • Make org.elasticsearch.action.admin.cluster.state.ClusterStateResponse Compress the Cluster State #79906
  • optimize getIndices in IndicesSegmentResponse #80064
  • Massive async shard fetch requests consume lots of heap memories on master node. #80694
  • Pending task batching can be a bottleneck #81626
  • Batch add-block-index-to-close, add-index-block and finalize-index-block tasks #81627
  • Stop unnecessary retries of shard-started tasks #81628
  • Batch up master tasks to create, mount, and delete snapshots #81846
  • Batch up failure-related ILM master tasks #81880
  • Stats actions should discard intermediate state on cancellation #82337
  • RestClusterGetSettingsAction Requests the Full Metadata from Master #82342
  • More Compact Serialization of Metadata #82608
  • Make TaskBatcher Less Lock-Heavy #82227
  • Speed up MappingStats Computation on Coordinating Node #82830
  • Add level=datastreams to Indices Stats #83049
  • (Re)Starting a Data Node Holding a Large Number of Indices can Take Minutes #83203
  • A Node Joining a Cluster with a Large State Receives the Full Uncompressed State in a ValidateJoinRequest #83204 -> Reduce resource needs of join validation #85380
  • RecoverySourceHandler#runWithGenericThreadPool caused deadlock #85839
  • Report stats related to new sizing guidance #86639
  • Make GetIndexAction cancellable #87681
  • Drop ClusterStateHealth#indices when unnecessary #90631
  • Improve scalability of BroadcastReplicationActions #92902
  • Computing IndicesQueryCache stats is O(N²) in shard count #97222
  • TransportBroadcastByNodeAction does O(#shards) work on transport worker thread #97914
  • Reduce usage of TransportMasterNodeReadAction #101805

• Snapshots + SLM

  • Less Verbose Serialization of Snapshot Failure in SLM Metadata #80942
  • Batch Snapshot Finalizations  #82824
  • Snapshot Deletion Could Run more Concurrently to Snapshot Creation #82853
  • Add Parameter to not Return Index Name Lists in GET Snapshots API #82937
  • Avoid capturing SnapshotsInProgress$Entry in queue #88707
  • Make SnapshotsInProgress diffable #88732
  • Make snapshot deletes less memory intensive by reordering repository metadata updates #89163
  • Snapshot creations have huge heap footprint after abrupt full-cluster restart #89952
  • Reduce the number of objects allocated by SLM when listing the snapshots to retain #99953

• Metrics

  • Add a Histogram of Transport Worker Time that is Spent per-Message #80428

• ILM + Allocation

  • Simple ILM Task Batching Implementation #78547
  • Optimize DataTierAllocationDecider Further #78235
  • ILM Can Create an Unlimited Number of Pending Clusterstate Updates on Slow Master Nodes #78246
  • Speed up DataTierAllocationDecider #78075
  • Implement DiffableStringMap's get and containsKey in terms of the wrapped innerMap #77965
  • Improve LifecycleExecutionState parsing. #77855
  • Reduce the number of times that LifecycleExecutionState is parsed when running a policy #77863
  • Speed up toXContent Collection Serialization in some Spots #78742
  • Store DataTier Preference directly on IndexMetadata #78668
  • Store Disk Threshold Ignore Setting in IndexMetadata #78672
  • Speed up Routing Nodes Priority Comparator #78609
  • Allow indices lookup to be built lazily #78745
  • Optimize XContent Object Parsers #78813
  • Find a way to Deduplicate Index Settings #78892
    • Implement setting deduplication via String interning #80493
  • MasterService#patchVersions is rather inefficient #77888
  • Replace RoutingTable#shardsWithState(...) with RoutingNodes#unassigned(...) #78931
  • Speedup computing cluster health #78969
  • IndexMetadataUpdater#applyChanges is rather inefficient #78980
  • Batch Cluster State Updates in Datastream Rollover #79782
  • Batch Index Settings Update Requests #79866
  • Make MasterService.patchVersions not Rebuild the Full CS #79860
  • Save some RoutingNodes Instantiations #79941
  • Cache DiscoveryNode#trimTier Result #80179
  • Rework ILM to not Require Inspecting all Indices on every Cluster State Update  #80407
  • Batch up failure-related ILM master tasks #81880
  • Faster ShardsLimitAllocationDecider #82251
  • Large ILM Task Batches are Executed too Slowly #82708
  • Make AllocationService#adaptAutoExpandReplicas Faster #83092
  • Speed up Building Indices Lookup in Metadata #83241
  • Speed up Name Collision Check in Metadata.Builder #83340
  • Make LIFECYCLE_NAME_SETTING a Field in IndexMetadata #83582
  • Use static empty store files metadata #84034
  • MetadataIndexAliasesService submits unbatched tasks at URGENT priority #89924
  • Improve sharing and diffability of IndexRoutingTable #94933
  • DataTiersUsageTransportAction is incredibly inefficient in large clusters #100230

• Search

  • Group shard request per node in the field capabilities API #74648
  • Group shard request per node in the can match phase #78164
  • Intern IndexFieldCapabilities Type String on Read #76405
  • Fix NumberFieldMapper Referencing its Own Builder #77131
  • Fix MatchOnlyTextFieldMapper Retaining a Reference to its Builder #77201
  • Fix TextFieldMapper Retaining a Reference to its Builder #77251
  • Search Responses to Many Shards use Excessive Amounts of Memory for OriginalIndices instances  #78314
  • Filter original indices in shard level request #78508
  • Merge field caps responses on each node? #82879
  • Updating index metadata does an expensive validation of the mapping even if unchanged #89309
  • Searches against a large number of unavailable shards result in very large responses #90622
  • Reduce copying when creating scroll/PIT ids #99219
  • Limit shard failures accumulated by searches #99220
  • Fork response-sending in OpenPointInTimeAction #99222

• Network

  • Sending Large Transport Messages Should be Optimized #82245
  • APIs like /_cluster/state Break for Large Clusters due to Response Size Limitations #79560
  • Render Mappings more Compact in GET /_cluster/state #83846
  • Add the Ability to Disable certain REST APIs via a Cluster Setting #84876
  • Track distribution of REST response sizes #84887
  • Make use of chunked REST response infrastructure in more APIs  #89838