Altinity Stable® Build for ClickHouse® 25.8

Backward Incompatible Changes

• Function geoToH3() now accepts arguments in the order (lat, lon, res). To keep the legacy order (lon, lat, res), enable geotoh3_lon_lat_input_order = true. #78852.

• Removed server setting format_alter_commands_with_parentheses. Note: ALTER commands may not work correctly in mixed-version replication setups (25.4+ with replicas older than 24.2) until all replicas are upgraded. #79970.

• Default expressions now require backticks around identifiers containing dots. #83162.

• Heredoc (dollar-quoted) tags are now restricted to word characters only to avoid parsing ambiguity. #84846.

• Table engine setting allow_dynamic_metadata_for_data_lakes is now enabled by default. #85044.

• Parquet output now writes Enum values as BYTE_ARRAY with ENUM logical type by default. #84169.

• JSON formats no longer quote 64-bit integers by default. #74079.

• DeltaLake storage now uses the delta-kernel-rs implementation by default. #79541.

• Removed legacy index types annoy and usearch. If present, they must be dropped. #79802.

• Experimental index full_text was renamed to gin, then to text. #79024 #80855.

• Materialized views are now validated against the target table when allow_materialized_view_with_bad_select is disabled. #74481.

• SummingMergeTree validation now skips aggregation for columns used in partition or sort keys. #78022.

• Fixed dateTrunc behavior with negative date/datetime values. #77622.

• Removed legacy MongoDB integration. Setting use_legacy_mongodb_integration is now obsolete. #77895.

• Added filesystem cache setting allow_dynamic_cache_resize (default false). Older versions allowed dynamic resizing without an explicit setting. #79148.

• URL reads now apply enable_url_encoding consistently across redirect chains. Default value of enable_url_encoding is now false. #79563 #80088.

• Server settings backup_threads and restore_threads are now enforced to be non-zero. #80224.

• bitNot() for String now returns a zero-terminated string. #80791.

• extractKeyValuePairs now supports a 5th argument: unexpected_quoting_character_strategy.

• countMatches now continues counting after empty matches. To keep legacy behavior, enable count_matches_stop_at_empty_match. #81676.

• BACKUP queries, merges, and mutations now use both server-wide throttlers and dedicated throttlers simultaneously. #81753.

• Creating tables without insertable columns is now forbidden. #81835.

• Cluster functions with archives can now process archives on multiple nodes using cluster_function_process_archive_on_multiple_nodes (default true). Set it to false for compatibility when upgrading to 25.7+ from earlier versions. #82355.

• SYSTEM RESTART REPLICAS now only affects databases where the user has SHOW TABLES permission. #83321.

• JSON input now infers Array(Dynamic) instead of unnamed Tuple for mixed-type arrays. To keep legacy behavior, disable input_format_json_infer_array_of_dynamic_from_array_of_different_types. #80859.

• S3 latency metrics were moved to histograms. #82305.

• Lazy materialization is now enabled only when the analyzer is enabled. #83791.

• Default concurrent_threads_scheduler changed from round_robin to fair_round_robin. #84747.

Click any tab in the tab bar above for more information.

SQL / Main features

• (Beta) Lightweight updates for MergeTree-family tables via patch parts, using UPDATE ... SET ... WHERE .... Lightweight deletes are implemented via lightweight updates and can be enabled with lightweight_delete_mode = 'lightweight_update' (requires enable_block_number_column, enable_block_offset_column, allow_experimental_lightweight_update). #82004 #85952

  • Added table setting max_uncompressed_bytes_in_patches to limit total uncompressed patch-part size and avoid SELECT slowdowns / misuse. #85641

• On-the-fly lightweight deletes are supported with lightweight_deletes_sync = 0 and apply_mutations_on_fly = 1. #79281

• (Beta) Correlated subqueries (EXISTS, IN, scalar) are enabled by default. #85107

• Approximate vector search with vector similarity indexes is now GA (see also vector_search_filter_strategy). #85888

• CPU & resource management: workload scheduler (CREATE WORKLOAD, CREATE RESOURCE) is available (see https://clickhouse.com/docs/operations/workload-scheduling).

• Added new data types Time and Time64 ([H]HH:MM:SS[.fractional]) with casts and related functions; compatibility settings for legacy toTime. #81217

• Added CoalescingMergeTree engine. #79344

• Data Lakes & catalogs are promoted to Beta (see below).

• FINAL query improvements:

  • Fixed cases where FINAL could be lost for Distributed tables. #78428
  • Settings use_skip_indexes_if_final and use_skip_indexes_if_final_exact_mode now default to true, enabling skip indexes for FINAL queries (set use_skip_indexes_if_final_exact_mode = false to keep legacy approximate behavior). #81331

• Added NumericIndexedVector (bit-sliced + Roaring-compressed numeric vectors) and 20+ related functions (groupNumericIndexedVectorState, numericIndexedVector*) for building/analysis/arithmetic. Improves joins, filters, and aggregations on sparse data and can significantly reduce storage. #74193

• Major areas with active development and notable improvements:

  • JSON data type
  • Database=Replicated
  • Parallel replicas
  • S3Queue
  • Keeper
  • Parquet format

Click any tab in the tab bar above for more information.

Functions

• SimpleAggregateFunction now supports anyLastRespectNulls and anyRespectNulls. #80219

• Added functions divideOrNull, moduloOrNull, intDivOrNull, positiveModuloOrNull to return NULL when the right argument is zero. #78276.

• Users can now do case-insensitive JSON key lookups using JSONExtractCaseInsensitive (and other variants of JSONExtract). #83770.

• Text index improvements:

  • Explicit parameters are now supported via key-value pairs (tokenizer required; optional max_rows_per_postings_list, ngram_size). #80262
  • Text index supports string tokenizer. #81752
  • tokens() supports string tokenizer. #80195
  • hasAll() can use tokenbf_v1 / ngrambf_v1 full-text skipping indices. #77662
  • Added mapContainsKey, mapContainsValue, mapContainsValueLike, mapExtractValueLike functions and bloomfilter-index support for filtering map values. #78171 (UnamedRus).

• Added experimental functions searchAny and searchAll which are general purpose tools to search text indexes. #80641.

• Added setting cast_string_to_date_time_mode to control DateTime parsing mode when casting from String. #80210

• Added dateTimeToUUIDv7(DateTime) conversion function. #84319

• Added UInt128 ↔ IPv6 conversion support, enabling arithmetic/bitwise operations (e.g. bitAnd) with conversion back to IPv6. #76928

• Array functions: arrayLevenshteinDistance, arrayLevenshteinDistanceWeighted, arraySimilarity.

• String functions: stringBytesUniq, stringBytesEntropy; SparseGrams, SparseGramsUTF8, SparseGramsHashes (n-grams extraction with additional longer/unique chunks).

• DateTime functions:

  • Added toInterval(value, unit) conversion function. #78723
  • Added nowInBlock64(scale) function. #84178

• Base32 / Bech32:

  • Added base32Encode, base32Decode.
  • Added bech32Encode, bech32Decode. #80239

• Geo functions:

  • Added MultiPolygon support for pointInPolygon. #79773
  • Added WKB parsing functions (e.g. readWKBPolygon, etc.). #80139
  • Added WKB encoding via wkb(...).
  • Added polygonsIntersectCartesian and polygonsIntersectSpherical.

• Color space conversion functions: colorOKLCHToSRGB, colorSRGBToOKLCH. #83679

• Time series / PromQL helpers:

  • Added timeSeriesDerivToGrid and timeSeriesPredictLinearToGrid aggregate functions. #84328
  • Added timeSeries* helper functions for resampling to a time grid and computing PromQL-like delta, rate, idelta, irate. #80590
  • Added helper functions: timeSeriesInstantDeltaToGrid, timeSeriesInstantRateToGrid, timeSeriesRateToGrid, timeSeriesResampleToGridWithStaleness, timeSeriesDeltaToGrid, timeSeriesLastTwoSamples.

• Added financial functions: financialInternalRateOfReturnExtended (XIRR), financialInternalRateOfReturn (IRR), financialNetPresentValueExtended (XNPV), financialNetPresentValue (NPV). #81599

• SQL compatibility: added lag and lead window functions. #82108

• Introspection:

  • added getServerSetting() and getMergeTreeSetting() functions. #78439
  • functions to unpack and compare MergeTree parts (mergeTreePartInfo, isPartCoveredBy). #80573.

Click any tab in the tab bar above for more information.

MergeTree-related

• Allow filtering query-selected parts by the disk they reside on. #80650

• filterPartsByQueryConditionCache now respects merge_tree_min_{rows,bytes}_for_seek to align with other index-based filtering methods. #80312

• Secondary index performance improvements:

  • Speed up secondary indexes by evaluating expressions on multiple granules at once. #64109
  • Process indexes in increasing order of size (prioritizing minmax and vector indexes, then smaller indexes). #84094
  • Process higher-granularity min-max indexes first. #83798

• Added MergeTree virtual columns:

  • _part_granule_offset: zero-based granule/mark index of each row within its part. #82341
  • _part_starting_offset: cumulative row count of all preceding parts, computed at query time and preserved through execution (even after pruning). #79417

• Projections: allow using _part_offset in normal projections (first step towards projection indexes). #78429

• Added setting max_merge_delayed_streams_for_parallel_write to control how many columns merges can flush in parallel (analogous to max_insert_delayed_streams_for_parallel_write). Helps reduce memory usage for vertical merges (notably for wide tables such as system.metric_log dumped to S3).

Click any tab in the tab bar above for more information.

Performance

• Added query plan optimization for lazy column materialization (read columns only after ORDER BY/LIMIT), controlled by query_plan_optimize_lazy_materialization. #55518

• Filtering by multiple projections.

• JOIN optimizations:

  • Merge equality conditions from filter step into JOIN condition to use them as hash keys. #78877
  • Added parallel_hash_join_threshold to fall back to hash algorithm when the right table is small. #76185
  • Added min_joined_block_size_rows (default 65409) to squash small JOIN blocks (when supported by algorithm). #81886
  • Added query-plan optimizations: merge WHERE into ON when possible (query_plan_merge_filter_into_join_condition), and shard joins by PK ranges for similarly ordered large tables (query_plan_join_shard_by_pk_ranges=1).
  • Low-level hash JOIN optimizations.

• User-defined functions: UDFs can be marked as deterministic in XML; query cache now checks UDF determinism and caches results when applicable. #77769

• Bloom filter index optimizations.

• Distributed query optimizations:

  • Added enable_parallel_blocks_marshalling to offload (de)compression and (de)serialization of network blocks into pipeline threads, improving performance of large distributed queries. #78694
  • Added enable_producing_buckets_out_of_order_in_aggregation (enabled by default) to send aggregation buckets out-of-order during memory-efficient aggregation, improving performance when some buckets merge slower (may slightly increase memory usage). #80179

• Compressed logs and profile events in the native protocol to reduce bandwidth usage and improve responsiveness on large clusters. #82535

• Avoid substituting table functions with *-cluster variants when a query contains JOINs or subqueries. #84335

• Improved cache locality using rendezvous hashing. #82511

• Fixed incorrect use of distributed_depth for cluster-function detection (could lead to data duplication); use client_info.collaborate_with_initiator instead. #85734

Click any tab in the tab bar above for more information.

Codecs / compression

• Added default_compression_codec for MergeTree tables, used when column codecs are not explicitly specified in CREATE TABLE. #66394

• DoubleDelta codec can now be applied only to numeric columns (no longer allowed for FixedString). #84383

• Changed default index granularity for text indexes to 64. #82162

Click any tab in the tab bar above for more information.

Other small SQL things

• Added TRUNCATE ALL TABLES FROM db LIKE '%pattern%'.

• DESCRIBE SELECT ... is now supported in addition to DESCRIBE (SELECT ...). #82947

• Added support for using external disks for temporary data.

• Fixed snapshots for all tables via enable_shared_storage_snapshot_in_query.

• Virtual column _table is now available everywhere.

Click any tab in the tab bar above for more information.

JSON data type

• JSON data type is production-ready (see https://jsonbench.com/). Dynamic and Variant data types are also production-ready. #77785

• Improved performance of reading JSON subcolumns from shared MergeTree data by introducing new shared-data serializations. #83777

• Compact parts: store marks per substream to allow reading individual subcolumns. The new format is controlled by write_marks_for_substreams_in_compact_parts (old format is still supported for reads). #77940

• write_marks_for_substreams_in_compact_parts is now enabled by default, significantly improving subcolumn reads from newly created Compact parts. Note: servers older than 25.5 cannot read these new Compact parts (disable the setting to preserve downgrade compatibility). #84171

• Added MergeTree setting merge_max_dynamic_subcolumns_in_wide_part to limit the number of dynamic subcolumns in Wide parts after merges. #87646

• Marked settings allow_experimental_variant/dynamic/json and enable_variant/dynamic/json as obsolete; Variant, Dynamic, and JSON are now enabled unconditionally. #85934

• MergeTree shared-data serialization is controlled by object_shared_data_serialization_version and object_shared_data_serialization_version_for_zero_level_parts. Default will remain map for the next few releases, then switch to advanced (zero-level parts: map or map_with_buckets). Note: changing shared-data serialization is supported only for object_serialization_version = v3 (currently default is v2, planned to switch to v3 in future releases).

• Fixed potential thread pool starvation when reading JSON column samples with many subcolumns by respecting merge_tree_use_prefixes_deserialization_thread_pool. #91208

Click any tab in the tab bar above for more information.

ObjectStore - S3/Azure/GCS

• Added support for custom IAM roles via S3 extra_credentials(role_arn = ...). #84011

• Added setting s3_slow_all_threads_after_network_error. #80035

• Azure config: add missing partition_columns_in_data_file. #85373

• Azure Blob Storage: replace Curl HTTP client with Poco client (enabled by default). Added multiple Azure client settings mirroring S3, introduced aggressive connect timeouts for both Azure and S3, improved Azure profile events/metrics introspection, and significantly improved cold-query latencies. Revert to Curl via azure_sdk_use_native_client=false. #83294

• S3Queue/AzureQueue: improved resilience to ZooKeeper connection loss without potential duplicates. Requires enabling use_persistent_processing_nodes (modifiable via ALTER TABLE MODIFY SETTING). #85995

• azureBlobStorage table engine: cache and reuse managed identity auth tokens to reduce throttling. #79860

• Permissions: azureBlobStorage, deltaLakeAzure, and icebergAzure now properly validate AZURE permissions; all *-Cluster variants validate permissions via their non-clustered counterparts; icebergLocal and deltaLakeLocal now require FILE permissions. #84938

• Azure Blob Storage: added extra_credentials authentication with client_id and tenant_id. #84235

• Object storage auth: added AWS S3 authentication with explicitly provided IAM role; added OAuth for GCS; synchronized object storage connection parameter serialization interfaces (previously Cloud-only, now open-sourced). #84011

• gcs function now requires GRANT READ ON S3 permission. #83503

• Object removals: collect all removed objects and execute a single object storage remove operation. #85316

Click any tab in the tab bar above for more information.

Datalakes & catalogs

• Data Lake Catalogs (Unity, REST, Glue, Hive Metastore) are promoted from Experimental to Beta. #85848
• Added _data_lake_snapshot_version to Data Lake table engines. #84659
• Implemented file pruning in Data Lake engines by virtual columns. #84520
• Added show_data_lake_catalogs_in_system_tables to control exposing Data Lake tables in system.tables. #85411

Iceberg

• Writes & catalogs: CREATE TABLE can create a new Iceberg table; INSERT into Iceberg is supported (controlled by allow_experimental_insert_into_iceberg), incl. compatibility for PyIceberg; REST and Glue catalogs are supported for writes; DROP TABLE removes from REST/Glue catalogs and deletes metadata. #83983 #82692 #84466 #84684 #84136 #85395

• Partitioning & pruning: Hive-style partitioned writes; partition columns are no longer virtual; bucket partitioning helpers icebergHash / icebergBucketTransform; partition pruning is enabled by default. #76802 #79262

• Time travel & metadata resolution: time travel via iceberg_timestamp_ms / iceberg_snapshot_id; snapshot version is not cached between SELECTs anymore (fixes time-travel issues); REST catalog metadata resolution fixes; version-hint.text support (read via iceberg_enable_version_hint, write support added). system.iceberg_history (availability/behavior to verify). #85038 #85531 #78594 #85130

• Metadata caching/observability & scaling: Iceberg metadata files cache via use_iceberg_metadata_files_cache; async iteration of Iceberg objects without materializing per-file object lists; system.iceberg_metadata_log records metadata files used during SELECT. #85369 #86152

• Schema evolution & column IDs: schema evolution support incl. complex types (plus add/drop/modify for simple types); read data files by field/column IDs to support renamed columns. #85769 #83653

• Statistics & optimizations: write more Iceberg stats (column sizes, lower/upper bounds) for pruning; pruning by bounds; faster count() without filters. #85746 #78242 #78090

• Deletes & maintenance: equality deletes and position deletes; ALTER DELETE for merge-on-read; reduced RAM usage for large position-delete files (keep only last Parquet row group in RAM); OPTIMIZE TABLE merges position-delete files into data files to reduce Parquet count/size; Spark compatibility fix for position-delete files. #85843 #83094 #80237 #85549 #85329 #85250 #85762

• Iceberg metadata format & write fixes: compressed metadata.json via iceberg_metadata_compression_method (all ClickHouse compression methods); fixes for complex-type writes; update metadata timestamp in writes. #85196 #85330 #85711

• Catalog-specific: Glue catalog supports reading after schema evolution and TimestampTZ; Hive Metastore catalog supported for Iceberg. #82301 #83132 #77677

• Object storage Hive integration fixes (LowCardinality, column fill order, empty format LOGICAL_ERROR, partition-column-only edge cases, schema assertions, partial parallel replicas fix, deterministic containers). #85538

• Since 25.7, ClickHouse works with Databricks-managed Iceberg tables.

Delta Lake

• Delta-kernel-based implementation enabled by default; local storage support; deltaLakeLocal table function for filesystem-mounted tables; deltaAzure for Azure; time travel via delta_lake_snapshot_version; experimental writes gate allow_experimental_delta_lake_writes + write support for DeltaLake. #79541 #79416 #79781 #85295 #86180 #85564

MergeTree as a Data Lake

• Readonly MergeTree tables support refresh for the one-writer / many-readers pattern. #76467

• Custom disks for databases.

Click any tab in the tab bar above for more information.

Integrations

• Experimental: Ytsaurus table engine and table function. #77606

• Arrow Flight: ClickHouse can expose an Arrow Flight RPC endpoint via arrowflight_port and can also query Arrow Flight servers using the arrowflight() table function (SELECT * FROM arrowflight('host:port', 'dataset_name')). #74184

• Kafka:

  • Added inline SASL credentials support (kafka_sasl_username, kafka_sasl_password).
  • Fixed cases where a materialized view could start too late (after the Kafka table that streams into it). #72123
  • Implemented Keeper-based rebalance logic for StorageKafka2 with permanent/temporary partition locks and dynamic redistribution across replicas. #78726
  • Added a system table to store erroneous incoming messages from Kafka-like engines. #68873

• PostgreSQL protocol:

  • Added COPY command support. #74344
  • Added SCRAM-SHA256 support and updated Postgres wire authentication.
  • Use FixedString for PostgreSQL CHARACTER, CHAR, and BPCHAR. #77304

• Force secure connections for mysql_port and postgresql_port. #82962

• MySQL protocol: implemented information schema tables to support C# MySQL clients. #84397

• PromQL: added basic PromQL dialect support (dialect='promql') for TimeSeries tables (promql_table_name). Supports rate, delta, increase, and prometheusQuery() wrapper; no operators and no HTTP API yet. #75036

Click any tab in the tab bar above for more information.

S3Queue

• Added per-table overrides for MV insert settings: min_insert_block_size_rows_for_materialized_views and min_insert_block_size_bytes_for_materialized_views (override profile-level settings). #83971

• Improved S3Queue/AzureQueue performance by enabling parallel INSERT pipeline via parallel_inserts=true (scales close to linearly with processing_threads_num). #77671

• Added _time virtual column to S3Queue. #78926

• Macros (e.g. {uuid}) are now supported in keeper_path for S3Queue. #82463

• Added server setting s3queue_disable_streaming (runtime changeable, no restart required) to disable streaming for S3Queue tables. #82515

Click any tab in the tab bar above for more information.

Formats

• Added Hash output format, which computes a single hash value over all rows and columns of the result (useful as a lightweight result fingerprint when data transfer is a bottleneck). Example: SELECT arrayJoin(['abc', 'def']), 42 FORMAT Hash returns e5f9e676db098fdb9530d2059d8c23ef. #84607

• Parquet:

  • Added experimental Parquet Reader v3 (new faster implementation with page-level filter pushdown and PREWHERE), enabled via input_format_parquet_use_native_reader_v3. #82789
  • Geo types: WKB-encoded geometry data is remapped to corresponding ClickHouse types.
  • Added JSON support: write ClickHouse JSON columns to Parquet and read Parquet JSON columns as ClickHouse JSON. #79649
  • Enum values are now written as BYTE_ARRAY with ENUM logical type by default. #84169
  • Added ParquetReadRowGroups and ParquetPrunedRowGroups counters.

• ORC: added settings for compression block size and changed the default from 64KB to 256KB (Spark/Hive compatible). #80602

• Added format_schema_source setting to define the source of format_schema. #80874

Click any tab in the tab bar above for more information.

Operations

• Database engines Atomic and Ordinary now support disk setting to store table metadata files on a specified disk. #80546

• Settings constraints now support explicitly disallowed values. #78499

• Tweaked jemalloc configuration for improved performance. #81807

• Memory tracking improvements:

  • External library allocations are now accounted by ClickHouse memory tracker (may increase reported memory usage / trigger MEMORY_LIMIT_EXCEEDED). #84082
  • Added MemoryResidentWithoutPageCache metric for more accurate memory usage reporting when userspace page cache is enabled. #81233
  • Memory tracker can be corrected using cgroup data via memory_worker_correct_memory_tracker (when cgroups are available and memory_worker_use_cgroup is enabled). #83981
  • Improved cgroup v2 detection on systems with both v1 and v2 enabled. #78566

• Backups: allow backing up PostgreSQL, MySQL, and DataLake databases (definition only, without data). #79982

• ODBC/JDBC: allow using named collections and unify parameter names. #80334 #83410

• Views created by ephemeral users now store a copy of the user and are not invalidated after the user is deleted. #84763

• AWS ECS credentials: added token expiration so it can be reloaded. #82422

• Added bind_host setting in clusters configuration to force distributed connections through a specific network. #74741

• Logging: introduced async logging and enabled it by default (<logger><async>false</async></logger> disables it). Added tunable limits and introspection, per-channel JSON logging via logger.formatting.channel, and fixed send_logs_source_regexp after refactoring. #82516 #80125 #85105 #86331 #85797

• Added schema_type setting for system.metric_log with schema options wide, transposed, and transposed_with_wide_view (note: transposed_with_wide_view does not support subsecond resolution; event_time_microseconds remains as alias for compatibility). #78412

• Set remote_filesystem_read_method default to threadpool. #77368

• Allow specifying http_response_headers in any http_handlers. #79975

• system.zookeeper now exposes auxiliary_zookeepers data. #80146

Click any tab in the tab bar above for more information.

CLI

• clickhouse-local supports persistent database storage via --path (e.g. ch --path .).

• clickhouse-local supports implicit tables (queries can omit FROM).

• Bundled <code>chdig</code> (top-like TUI for ClickHouse) as part of ClickHouse. #79666

• clickhouse-client: added AI-powered SQL generation via ?? prefix (OpenAI/Anthropic, automatic schema discovery). Supports zero-config API key discovery via OPENAI_API_KEY / ANTHROPIC_API_KEY. #83314 #83787

• clickhouse-benchmark: added --max-concurrency option with gradual ramp-up of parallel queries. #85623

Click any tab in the tab bar above for more information.

Web UI

• Added a landing page listing embedded web tools (opened automatically for browser-like user agents). #81129

• Improved Web UI scrollbars appearance. #82869

• clickhouse-server with embedded configuration now supports Web UI access by responding to HTTP OPTIONS requests. #82870*

Click any tab in the tab bar above for more information.

Security

• CI/CD can now use FIPS-permissive mode. #82415 #83418

• Fixed passing external roles in interserver queries. #79099

• Parameterized CREATE USER query.

• External authentication: match forward_headers case-insensitively. #84737

• Added setting access_control_improvements.enable_user_name_access_type to enable/disable precise user/role grants (introduced in #72246). Should be disabled for compatibility with replicas older than 25.1. #79842

• Added new source access types READ and WRITE, deprecating previous source-related access types (e.g. GRANT READ, WRITE ON S3 TO user). Controlled by access_control_improvements.enable_read_write_grants (disabled by default). #73659

• Azure/FILE permissions: azureBlobStorage, deltaLakeAzure, and icebergAzure now properly validate AZURE permissions; *-Cluster variants validate against their non-clustered counterparts; icebergLocal and deltaLakeLocal now require FILE permissions. #84938

• Added new syntax for scoped S3 grants: GRANT READ ON S3('s3://foo/.*') TO user. #84503

• Mask credentials in deltaLakeAzure, deltaLakeCluster, icebergS3Cluster, and icebergAzureCluster. #85889

• Query masking rules can now throw LOGICAL_ERROR on match to detect credential leaks in logs. #78094

Click any tab in the tab bar above for more information.

Network

• Drop TCP connections after a configured number of queries or after a time threshold. #81472

• HTTP buffers: embed proxy configuration using builder-based setup. #77693

• Server bandwidth throttling: allow reloading max_local_read_bandwidth_for_server and max_local_write_bandwidth_for_server without server restart. #82083

Click any tab in the tab bar above for more information.

Analyzer

• Improved resolution of compound identifiers (notably better ARRAY JOIN compatibility). To keep the old behavior, enable analyzer_compatibility_allow_compound_identifiers_in_unflatten_nested. #85492

• Distributed queries: experimental query plan distribution via serialize_query_plan / process_query_plan_packet.

Click any tab in the tab bar above for more information.

Database=Replicated

• Added logs_to_keep setting for Replicated databases to control how many log entries are kept in ZooKeeper (lower values reduce ZNodes; higher values allow replicas to catch up after longer downtime). #84183

• Added database_replicated settings to define default values for DatabaseReplicatedSettings when they are not explicitly specified in CREATE DATABASE ... ENGINE=Replicated. #85127

• Enabled backoff logic for replicated tasks to reduce CPU/memory usage and log volume. Added settings max_postpone_time_for_failed_replicated_fetches_ms, max_postpone_time_for_failed_replicated_merges_ms, and max_postpone_time_for_failed_replicated_tasks_ms. #74576

Click any tab in the tab bar above for more information.

Parallel replicas

• Added projection support for parallel replicas (controlled by parallel_replicas_support_projection, requires parallel_replicas_local_plan). #82807

• Parallel replicas no longer wait for slow unused replicas once all read tasks are already assigned to other replicas. #80199

• Distributed INSERT SELECT into replicated MergeTree now uses parallel replicas more efficiently by selecting different data on different nodes and inserting independently. #78041

Click any tab in the tab bar above for more information.

Keeper

• Added support for setting arbitrary watches in Keeper multi queries. #84964

• Added clickhouse-keeper-utils CLI tool to dump/analyze Keeper snapshots and changelogs (including extracting log ranges). #81677

• keeper-utils: added configurable limit for dumping top subtrees. #85198

• Added get_acl command to KeeperClient. #84641

• Keeper performance improvements:

  • Reduced contention on storage lock. #84732
  • Improved RocksDB initial loading. #83390
  • Move changelog files between disks in a background thread to avoid blocking Keeper (e.g. when moving to S3 disk). #82485
  • Added limits for log entry cache size via latest_logs_cache_entry_count_threshold and commit_logs_cache_entry_count_threshold. #84877

• ACL improvements:

  • Added keeper_server.cleanup_old_and_ignore_new_acl to clear existing ACLs and ignore new ACL requests (should remain enabled until a new snapshot is created). #82496
  • Added support for specific permissions in world:anyone ACL. #82755

Click any tab in the tab bar above for more information.

Introspection

• Added pressure metrics to async metrics. #80779

• Added a dimensional metric for ConcurrentBoundedQueue size, labelled by queue type and queue id. #84675

• CGroup-level and OS-level metrics are now reported together: cgroup metrics use CGroup<Metric> names, OS/procfs metrics use OS<Metric>. #84317

• Added query_id to system.errors. #76581

• Added last_error_message, last_error_trace, and query_id to system.error_log. #79836

• system.functions now shows the ClickHouse version where each function was introduced. #79839

• Added system.zookeeper_connection_log with historical ZooKeeper connection information. #79494

• Added support for partially aggregated metrics. #85328

• Added system.completions system table. #83833

• Refreshes of refreshable materialized views are now logged in system.query_log. #71333

• Added system.codecs system table to introspect available codecs. #81600

• Added async metrics for server TCP sockets. #80188

• Added mergeTreeProjection table function to read projection parts directly for introspection. #78429

Click any tab in the tab bar above for more information.