mirror of
https://github.com/ClickHouse/ClickHouse.git
synced 2024-11-26 17:41:59 +00:00
7.9 KiB
7.9 KiB
ClickHouse release 1.1.54284
- This is bugfix release for previous 1.1.54282 release. It fixes ZooKeeper nodes leak in
parts/
directory.
ClickHouse release 1.1.54282
This is a bugfix release. The following bugs were fixed:
DB::Exception: Assertion violation: !_path.empty()
error when inserting into a Distributed table.- Error when parsing inserted data in RowBinary format if the data begins with ';' character.
- Errors during runtime compilation of certain aggregate functions (e.g.
groupArray()
).
ClickHouse release 1.1.54276
New features:
- You can use an optional WITH clause in a SELECT query. Example query:
WITH 1+1 AS a SELECT a, a*a
- INSERT can be performed synchronously in a Distributed table: OK is returned only after all the data is saved on all the shards. This is activated by the setting insert_distributed_sync=1.
- Added the UUID data type for working with 16-byte identifiers.
- Added aliases of CHAR, FLOAT and other types for compatibility with the Tableau.
- Added the functions toYYYYMM, toYYYYMMDD, and toYYYYMMDDhhmmss for converting time into numbers.
- You can use IP addresses (together with the hostname) to identify servers for clustered DDL queries.
- Added support for non-constant arguments and negative offsets in the function
substring(str, pos, len).
- Added the max_size parameter for the
groupArray(max_size)(column)
aggregate function, and optimized its performance.
Major changes:
- Improved security: all server files are created with 0640 permissions.
- Improved error messages for queries with invalid syntax.
- Significantly reduced memory consumption and improved performance when merging large sections of MergeTree data.
- Significantly increased the performance of data merges for the ReplacingMergeTree engine.
- Improved performance for asynchronous inserts from a Distributed table by batching multiple source inserts. To enable this functionality, use the setting distributed_directory_monitor_batch_inserts=1.
Backward incompatible changes:
- Changed the binary format of aggregate states of
groupArray(array_column)
functions for arrays.
Complete list of changes:
- Added the
output_format_json_quote_denormals
setting, which enables outputting nan and inf values in JSON format. - Optimized thread allocation when reading from a Distributed table.
- Settings can be modified in readonly mode if the value doesn't change.
- Added the ability to read fractional granules of the MergeTree engine in order to meet restrictions on the block size specified in the preferred_block_size_bytes setting. The purpose is to reduce the consumption of RAM and increase cache locality when processing queries from tables with large columns.
- Efficient use of indexes that contain expressions like
toStartOfHour(x)
for conditions liketoStartOfHour(x) op сonstexpr.
- Added new settings for MergeTree engines (the merge_tree section in config.xml):
- replicated_deduplication_window_seconds sets the size of deduplication window in seconds for Replicated tables.
- cleanup_delay_period sets how often to start cleanup to remove outdated data.
- replicated_can_become_leader can prevent a replica from becoming the leader (and assigning merges).
- Accelerated cleanup to remove outdated data from ZooKeeper.
- Multiple improvements and fixes for clustered DDL queries. Of particular interest is the new setting distributed_ddl_task_timeout, which limits the time to wait for a response from the servers in the cluster.
- Improved display of stack traces in the server logs.
- Added the "none" value for the compression method.
- You can use multiple dictionaries_config sections in config.xml.
- It is possible to connect to MySQL through a socket in the file system.
- The
system.parts
table has a new column with information about the size of marks, in bytes.
Bug fixes:
- Distributed tables using a Merge table now work correctly for a SELECT query with a condition on the _table field.
- Fixed a rare race condition in ReplicatedMergeTree when checking data parts.
- Fixed possible freezing on "leader election" when starting a server.
- The max_replica_delay_for_distributed_queries setting was ignored when using a local replica of the data source. This has been fixed.
- Fixed incorrect behavior of
ALTER TABLE CLEAR COLUMN IN PARTITION
when attempting to clean a non-existing column. - Fixed an exception in the multiIf function when using empty arrays or strings.
- Fixed excessive memory allocations when deserializing Native format.
- Fixed incorrect auto-update of Trie dictionaries.
- Fixed an exception when running queries with a GROUP BY clause from a Merge table when using SAMPLE.
- Fixed a crash of GROUP BY when using distributed_aggregation_memory_efficient=1.
- Now you can specify the database.table in the right side of IN and JOIN.
- Too many threads were used for parallel aggregation. This has been fixed.
- Fixed how the "if" function works with FixedString arguments.
- SELECT worked incorrectly from a Distributed table for shards with a weight of 0. This has been fixed.
- Crashes no longer occur when running
CREATE VIEW IF EXISTS.
- Fixed incorrect behavior when input_format_skip_unknown_fields=1 is set and there are negative numbers.
- Fixed an infinite loop in the
dictGetHierarchy()
function if there is some invalid data in the dictionary. - Fixed
Syntax error: unexpected (...)
errors when running distributed queries with subqueries in an IN or JOIN clause and Merge tables. - Fixed the incorrect interpretation of a SELECT query from Dictionary tables.
- Fixed the "Cannot mremap" error when using arrays in IN and JOIN clauses with more than 2 billion elements.
- Fixed the failover for dictionaries with MySQL as the source.
Improved workflow for developing and assembling ClickHouse:
- Builds can be assembled in Arcadia.
- You can use gcc 7 to compile ClickHouse.
- Parallel builds using ccache+distcc are faster now.
ClickHouse release 1.1.54245
New features:
- Distributed DDL (for example,
CREATE TABLE ON CLUSTER
). - The replicated request
ALTER TABLE CLEAR COLUMN IN PARTITION.
- The engine for Dictionary tables (access to dictionary data in the form of a table).
- Dictionary database engine (this type of database automatically has Dictionary tables available for all the connected external dictionaries).
- You can check for updates to the dictionary by sending a request to the source.
- Qualified column names
- Quoting identifiers using double quotation marks.
- Sessions in the HTTP interface.
- The OPTIMIZE query for a Replicated table can can run not only on the leader.
Backward incompatible changes:
- Removed SET GLOBAL.
Minor changes:
- If an alert is triggered, the full stack trace is printed into the log.
- Relaxed the verification of the number of damaged or extra data parts at startup (there were too many false positives).
Bug fixes:
- Fixed a bad connection "sticking" when inserting into a Distributed table.
- GLOBAL IN now works for a query from a Merge table that looks at a Distributed table.
- The incorrect number of cores was detected on a Google Compute Engine virtual machine. This has been fixed.
- Changes in how an executable source of cached external dictionaries works.
- Fixed the comparison of strings containing null characters.
- Fixed the comparison of Float32 primary key fields with constants.
- Previously, an incorrect estimate of the size of a field could lead to overly large allocations. This has been fixed.
- Fixed a crash when querying a Nullable column added to a table using ALTER.
- Fixed a crash when sorting by a Nullable column, if the number of rows is less than LIMIT.
- Fixed an ORDER BY subquery consisting of only constant values.
- Previously, a Replicated table could remain in the invalid state after a failed DROP TABLE.
- Aliases for scalar subqueries with empty results are no longer lost.
- Now a query that used compilation does not fail with an error if the .so file gets damaged.