What ClickHouse/ClickHouse shipped

This week: fixed row loss in correlated subqueries with NULL outer values, arrayJoin under expression extraction, and dictGet constant folding. Also fixed WITH FILL bounds truncating, FULL JOIN over Enum keys, and a dictionary lazy-load hang.

This week's changes focus on correctness where rows could go missing. Correlated scalar and EXISTS subqueries now return rows produced for a NULL outer value, and arrayJoin and aggregation rows are no longer lost when common-expression extraction absorbs an OR operand or when dictionary lookup folding collapses a comparison to a constant. WITH FILL now rejects FROM and TO bounds that do not fit the ORDER BY column type instead of wrapping around silently. FULL JOIN over an Enum key now handles rows present only in the right table, and zero-sized nested aggregate states are bounded in one place for all combinators. A query that lazily loads a dictionary in a CPU workload no longer hangs. Iceberg readers also got fixes for pruning that skipped NULL partitions and for equality deletes droppin

This week: HTTP PUT uploads to table paths, Iceberg and Delta Lake reads now report total_rows_approx, query text hidden from the HTTP Request URI log line, and fixes for S3Queue RENAME, integer rounding at type limits, DEFAULT expressions through subcolumns, and Parquet read abo

Changes this week in ClickHouse: HTTP PUT uploads to existing table paths are supported as a first step toward accessing tables as files. Iceberg and Delta Lake reads now report total_rows_approx from metadata so clients can show progress. The HTTP Request URI log line and clickhouse.uri span no longer include query text or parameters. Fixes landed for RENAME of an S3Queue table whose startup failed, floor/ceil/round on integers with a negative N near type limits, DEFAULT expressions that read another omitted column through a subcolumn or ALIAS, server aborts on Parquet reads with no missing values, and definer identity for SQL SECURITY DEFINER views without an invoker. PromQL now evaluates operators on two constant numbers, so histogram_quantile with a fraction like 1./6. works.

This week: DISTINCT now runs on several nodes in Cascades and is priced; Iceberg manifests read on the IO pool; GDeflate codec on CPU; CAST from Variant and Dynamic fails only for alternatives present in data; fixes for Decimal-vs-float key analysis and arrayReduceInRanges bounds

This week in ClickHouse: the Cascades optimizer now runs DISTINCT on several nodes and prices the alternatives, and Iceberg manifest files are read on the IO thread pool with the separate decode pool removed. A GDeflate compression codec is available on the CPU. Casts from Variant and Dynamic now fail only for alternatives present in the data, which makes them easier to work with. Correctness fixes landed for key analysis when a Decimal key is compared with a float and for arrayReduceInRanges when range bounds do not fit in the array. Restarting a whole-schema MaterializedPostgreSQL database is fixed after a new PostgreSQL table appears, and S3 batch delete falls back to single-object deletes on stores that reject the batch request.

This week: UNIQUE KEY SELECTs now skip deleted rows, the columns cache reaches Parquet, external IN disk sets are faster, and parseDateTime rejects invalid ISO week 53 dates. Plus fixes for modulo, polygon dictionaries, and float-to-wide-integer conversions.

This week's changes to ClickHouse focus on correct results and cheaper reads. SELECT on UNIQUE KEY tables now skips rows removed by delete bitmaps, and the columns cache extends to Parquet files on local and object storage. The external IN disk set is optimized for common workloads, with spilling on by default. Fixes include modulo by a constant minimum signed divisor, Float32 and BFloat16 conversion to 128- and 256-bit integers, polygon dictionaries for small or far-apart polygons, parseDateTime rejecting invalid ISO week 53 dates, and outer JOIN failures with full_sorting_merge. The Hive metastore catalog now respects remote_url_allow_hosts, and the SSH-key handshake no longer discloses whether a user exists.

This week: bounded Native protocol collections, Parquet min/max from column chunk statistics, fixes for pointInPolygon with huge coordinates, SummingMergeTree FINAL zero rows, TTL merges without row TTL info, and skip indexes built from merged insert rows.

This week in the project: Native protocol collections read from a peer are now bounded, Parquet min and max can be answered from column chunk statistics, and CODEC(T64) works on a bucketed Map. Wrong results were fixed for pointInPolygon with huge coordinates, SELECT FINAL from SummingMergeTree, and TTL merges on parts without row TTL info. Skip indexes are now built from the merged rows when an INSERT merges its block, and several fixes were backported to the 26.8 and 26.9 branches.

New this week: experimental LATERAL JOIN for correlated subqueries, Overlay databases, FileLog glob patterns and reset command, dialect shortcuts, and vectorized least/greatest/clamp functions. Plus fixes for Distributed tables and Delta Lake.

ClickHouse updates this week include experimental LATERAL JOIN support for advanced SQL patterns and top-N-per-group queries, new Overlay databases that expose unions of source tables, and FileLog enhancements with glob patterns and reset capability. Performance improvements include vectorized least/greatest/clamp functions (3-5x faster) and dialect switching commands. Bug fixes address Distributed table edge cases, Delta Lake column mapping, and metadata handling.

ClickHouse now applies seccomp filtering at startup for improved containment. This week also brings performance optimizations for substring operations, array searches, and CSV reading, plus fixes for UTF8 case-insensitive search consistency.

ClickHouse continues improving security and performance. This week's updates include seccomp system call filtering at server startup for enhanced containment, performance optimizations for substring operations and array functions, and important fixes for UTF8-aware string operations and distributed query handling. These changes address both kernel interface containment and query execution consistency.

New in ClickHouse: removed legacy query analysis code, fixed Azure range validation, secured HDFS table cache access checks, and resolved query correctness issues with correlated subqueries and NaN handling.

This week's ClickHouse updates focus on stability and correctness: we completed the transition to the new analyzer by removing legacy query analysis code; fixed a critical security issue in HDFS database table caching that wasn't enforcing read source grants; improved Azure Blob Storage reliability with range offset validation; resolved query correctness bugs affecting correlated subqueries, parameterized views, and floating-point NaN handling; and merged query optimization settings for cleaner configuration.