ClickHouse applies this setting when the query contains the product of distributed tables, i.e. when the query for a distributed table contains a non-GLOBAL subquery for the distributed table.
If `enable_optimize_predicate_expression = 1`, then the execution time of these queries is equal, because ClickHouse applies `WHERE` to the subquery when processing it.
If `enable_optimize_predicate_expression = 0`, then the execution time of the second query is much longer, because the `WHERE` clause applies to all the data after the subquery finishes.
If `force_index_by_date=1`, ClickHouse checks whether the query has a date key condition that can be used for restricting data ranges. If there is no suitable condition, it throws an exception. However, it does not check whether the condition actually reduces the amount of data to read. For example, the condition `Date != ' 2000-01-01 '` is acceptable even when it matches all the data in the table (i.e., running the query requires a full scan). For more information about ranges of data in MergeTree tables, see "[MergeTree](../../operations/table_engines/mergetree.md)".
If `force_primary_key=1`, ClickHouse checks to see if the query has a primary key condition that can be used for restricting data ranges. If there is no suitable condition, it throws an exception. However, it does not check whether the condition actually reduces the amount of data to read. For more information about data ranges in MergeTree tables, see "[MergeTree](../../operations/table_engines/mergetree.md)".
Enables or disables checksum verification when decompressing the HTTP POST data from the client. Used only for ClickHouse native compression format (not used with `gzip` or `deflate`).
Sets the maximum number of acceptable errors when reading from text formats (CSV, TSV, etc.).
The default value is 0.
Always pair it with `input_format_allow_errors_ratio`. To skip errors, both settings must be greater than 0.
If an error occurred while reading rows but the error counter is still less than `input_format_allow_errors_num`, ClickHouse ignores the row and moves on to the next one.
Sets the maximum percentage of errors allowed when reading from text formats (CSV, TSV, etc.).
The percentage of errors is set as a floating-point number between 0 and 1.
The default value is 0.
Always pair it with `input_format_allow_errors_num`. To skip errors, both settings must be greater than 0.
If an error occurred while reading rows but the error counter is still less than `input_format_allow_errors_ratio`, ClickHouse ignores the row and moves on to the next one.
If `input_format_allow_errors_ratio` is exceeded, ClickHouse throws an exception.
Enables or disables the full SQL parser if the fast stream parser can't parse the data. This setting is used only for the [Values](../../interfaces/formats.md#data-format-values) format at the data insertion. For more information about syntax parsing, see the [Syntax](../../query_language/syntax.md) section.
In this case, you can use an SQL expression as a value, but data insertion is much slower this way. If you insert only formatted data, then ClickHouse behaves as if the setting value is 0.
When executing the `INSERT` query, the ClickHouse client prepares data and sends it to the server for writing. The client gets the table structure from the server when preparing the data. In some cases, the client needs more information than the server sends by default. Turn on the extended data exchange with `input_format_defaults_for_omitted_fields = 1`.
When the extended data exchange is enabled, the server sends the additional metadata along with the table structure. The composition of the metadata depends on the operation.
Operations where you may need the extended data exchange enabled:
When writing data, ClickHouse throws an exception if input data contain columns that do not exist in the target table. If skipping is enabled, ClickHouse doesn't insert extra data and doesn't throw an exception.
To improve insert performance, we recommend disabling this check if you are sure that the column order of the input data is the same as in the target table.
-`ALL` — If the right table has several matching rows, the data is multiplied by the number of these rows. This is the normal `JOIN` behavior from standard SQL.
-`ANY` — If the right table has several matching rows, only the first one found is joined. If the right table has only one matching row, the results of `ANY` and `ALL` are the same.
Sets the type of [JOIN](../../query_language/select.md) behavior. When merging tables, empty cells may appear. ClickHouse fills them differently based on this setting.
- 1 — `JOIN` behaves the same way as in standard SQL. The type of the corresponding field is converted to [Nullable](../../data_types/nullable.md#data_type-nullable), and empty cells are filled with [NULL](../../query_language/syntax.md).
In ClickHouse, data is processed by blocks (sets of column parts). The internal processing cycles for a single block are efficient enough, but there are noticeable expenditures on each block. The `max_block_size` setting is a recommendation for what size of block (in number of rows) to load from tables. The block size shouldn't be too small, so that the expenditures on each block are still noticeable, but not too large, so that the query with LIMIT that is completed after the first block is processed quickly. The goal is to avoid consuming too much memory when extracting a large number of columns in multiple threads, and to preserve at least some cache locality.
Blocks the size of `max_block_size` are not always loaded from the table. If it is obvious that less data needs to be retrieved, a smaller block is processed.
## preferred_block_size_bytes
Used for the same purpose as `max_block_size`, but it sets the recommended block size in bytes by adapting it to the number of rows in the block.
However, the block size cannot be more than `max_block_size` rows.
ClickHouse uses multiple threads when reading from [MergeTree*](../table_engines/mergetree.md) tables. This setting turns on/off the uniform distribution of reading tasks over the working threads. The algorithm of the uniform distribution aims to make execution time for all the threads approximately equal in a `SELECT` query.
If the number of rows to be read from a file of a [MergeTree*](../table_engines/mergetree.md) table exceeds `merge_tree_min_rows_for_concurrent_read` then ClickHouse tries to perform a concurrent reading from this file on several threads.
If the distance between two data blocks to be read in one file is less than `merge_tree_min_rows_for_seek` rows, then ClickHouse does not seek through the file, but reads the data sequentially.
When searching data, ClickHouse checks the data marks in the index file. If ClickHouse finds that required keys are in some range, it divides this range into `merge_tree_coarse_index_granularity` subranges and searches the required keys there recursively.
If ClickHouse should read more than `merge_tree_max_rows_to_use_cache` rows in one query, it does not use the cash of uncompressed blocks. The [uncompressed_cache_size](../server_settings/settings.md#server-settings-uncompressed_cache_size) server setting defines the size of the cache of uncompressed blocks.
ClickHouse uses this setting when reading data from tables. If the total storage volume of all the data to be read exceeds `min_bytes_to_use_direct_io` bytes, then ClickHouse reads the data from the storage disk with the `O_DIRECT` option.
Queries sent to ClickHouse with this setup are logged according to the rules in the [query_log](../server_settings/settings.md#server_settings-query-log) server configuration parameter.
The size of blocks to form for insertion into a table.
This setting only applies in cases when the server forms the blocks.
For example, for an INSERT via the HTTP interface, the server parses the data format and forms blocks of the specified size.
But when using clickhouse-client, the client parses the data itself, and the 'max_insert_block_size' setting on the server doesn't affect the size of the inserted blocks.
The setting also doesn't have a purpose when using INSERT SELECT, since data is inserted using the same blocks that are formed after SELECT.
The default is slightly more than `max_block_size`. The reason for this is because certain table engines (`*MergeTree`) form a data part on the disk for each inserted block, which is a fairly large entity. Similarly, `*MergeTree` tables sort data during insertion, and a large enough block size allows sorting more data in RAM.
The maximum number of query processing threads, excluding threads for retrieving data from remote servers (see the 'max_distributed_connections' parameter).
For example, when reading from a table, if it is possible to evaluate expressions with functions, filter with WHERE and pre-aggregate for GROUP BY in parallel using at least 'max_threads' number of threads, then 'max_threads' are used.
If less than one SELECT query is normally run on a server at a time, set this parameter to a value slightly less than the actual number of processor cores.
For queries that are completed quickly because of a LIMIT, you can set a lower 'max_threads'. For example, if the necessary number of entries are located in every block and max_threads = 8, then 8 blocks are retrieved, although it would have been enough to read just one.
The smaller the `max_threads` value, the less memory is consumed.
## max_compress_block_size
The maximum size of blocks of uncompressed data before compressing for writing to a table. By default, 1,048,576 (1 MiB). If the size is reduced, the compression rate is significantly reduced, the compression and decompression speed increases slightly due to cache locality, and memory consumption is reduced. There usually isn't any reason to change this setting.
For [MergeTree](../../operations/table_engines/mergetree.md)" tables. In order to reduce latency when processing queries, a block is compressed when writing the next mark if its size is at least 'min_compress_block_size'. By default, 65,536.
The actual size of the block, if the uncompressed data is less than 'max_compress_block_size', is no less than this value and no less than the volume of data for one mark.
Let's look at an example. Assume that 'index_granularity' was set to 8192 during table creation.
We are writing a UInt32-type column (4 bytes per value). When writing 8192 rows, the total will be 32 KB of data. Since min_compress_block_size = 65,536, a compressed block will be formed for every two marks.
We are writing a URL column with the String type (average size of 60 bytes per value). When writing 8192 rows, the average will be slightly less than 500 KB of data. Since this is more than 65,536, a compressed block will be formed for each mark. In this case, when reading data from the disk in the range of a single mark, extra data won't be decompressed.
There usually isn't any reason to change this setting.
The maximum part of a query that can be taken to RAM for parsing with the SQL parser.
The INSERT query also contains data for INSERT that is processed by a separate stream parser (that consumes O(1) RAM), which is not included in this restriction.
The maximum number of simultaneous connections with remote servers for distributed processing of a single query to a single Distributed table. We recommend setting a value no less than the number of servers in the cluster.
The following parameters are only used when creating Distributed tables (and when launching a server), so there is no reason to change them at runtime.
## distributed_connections_pool_size
The maximum number of simultaneous connections with remote servers for distributed processing of all queries to a single Distributed table. We recommend setting a value no less than the number of servers in the cluster.
The timeout in milliseconds for connecting to a remote server for a Distributed table engine, if the 'shard' and 'replica' sections are used in the cluster definition.
If unsuccessful, several attempts are made to connect to various replicas.
Whether to use a cache of uncompressed blocks. Accepts 0 or 1. By default, 0 (disabled).
Using the uncompressed cache (only for tables in the MergeTree family) can significantly reduce latency and increase throughput when working with a large number of short queries. Enable this setting for users who send frequent short requests. Also pay attention to the [uncompressed_cache_size](../server_settings/settings.md#server-settings-uncompressed_cache_size) configuration parameter (only set in the config file) – the size of uncompressed cache blocks. By default, it is 8 GiB. The uncompressed cache is filled in as needed and the least-used data is automatically deleted.
For queries that read at least a somewhat large volume of data (one million rows or more), the uncompressed cache is disabled automatically in order to save space for truly small queries. This means that you can keep the 'use_uncompressed_cache' setting always set to 1.
When using the HTTP interface, the 'query_id' parameter can be passed. This is any string that serves as the query identifier.
If a query from the same user with the same 'query_id' already exists at this time, the behavior depends on the 'replace_running_query' parameter.
`0` (default) – Throw an exception (don't allow the query to run if a query with the same 'query_id' is already running).
`1`– Cancel the old query and start running the new one.
Yandex.Metrica uses this parameter set to 1 for implementing suggestions for segmentation conditions. After entering the next character, if the old query hasn't finished yet, it should be canceled.
This parameter is useful when you are using formats that require a schema definition, such as [Cap'n Proto](https://capnproto.org/). The value depends on the format.
The number of errors is counted for each replica. The query is sent to the replica with the fewest errors, and if there are several of these, to any one of them.
Disadvantages: Server proximity is not accounted for; if the replicas have different data, you will also get different data.
The number of errors is counted for each replica. Every 5 minutes, the number of errors is integrally divided by 2. Thus, the number of errors is calculated for a recent time with exponential smoothing. If there is one replica with a minimal number of errors (i.e. errors occurred recently on the other replicas), the query is sent to it. If there are multiple replicas with the same minimal number of errors, the query is sent to the replica with a host name that is most similar to the server's host name in the config file (for the number of different characters in identical positions, up to the minimum length of both host names).
For instance, example01-01-1 and example01-01-2.yandex.ru are different in one position, while example01-01-1 and example01-02-2 differ in two places.
This method might seem primitive, but it doesn't require external data about network topology, and it doesn't compare IP addresses, which would be complicated for our IPv6 addresses.
Thus, if there are equivalent replicas, the closest one by name is preferred.
We can also assume that when sending a query to the same server, in the absence of failures, a distributed query will also go to the same servers. So even if different data is placed on the replicas, the query will return mostly the same results.
This algorithm chooses the first replica in the set or a random replica if the first is unavailable. It's effective in cross-replication topology setups, but useless in other configurations.
The `first_or_random` algorithm solves the problem of the `in_order` algorithm. With `in_order`, if one replica goes down, the next one gets a double load while the remaining replicas handle the usual amount of traffic. When using the `first_or_random` algorithm, load is evenly distributed among replicas that are still available.
If this portion of the pipeline was compiled, the query may run faster due to deployment of short cycles and inlining aggregate function calls. The maximum performance improvement (up to four times faster in rare cases) is seen for queries with multiple simple aggregate functions. Typically, the performance gain is insignificant. In very rare cases, it may slow down query execution.
For testing, the value can be set to 0: compilation runs synchronously and the query waits for the end of the compilation process before continuing execution. For all other cases, use values starting with 1. Compilation normally takes about 5-10 seconds.
If the value is 1 or more, compilation occurs asynchronously in a separate thread. The result will be used as soon as it is ready, including queries that are currently running.
Compiled code is required for each different combination of aggregate functions used in the query and the type of keys in the GROUP BY clause.
The results of compilation are saved in the build directory in the form of .so files. There is no restriction on the number of compilation results, since they don't use very much space. Old results will be used after server restarts, except in the case of a server upgrade – in this case, the old results are deleted.
If the value is true, integers appear in quotes when using JSON\* Int64 and UInt64 formats (for compatibility with most JavaScript implementations); otherwise, integers are output without the quotes.
`INSERT` succeeds only when ClickHouse manages to correctly write data to the `insert_quorum` of replicas during the `insert_quorum_timeout`. If for any reason the number of replicas with successful writes does not reach the `insert_quorum`, the write is considered failed and ClickHouse will delete the inserted block from all the replicas where data has already been written.
All the replicas in the quorum are consistent, i.e., they contain data from all previous `INSERT` queries. The `INSERT` sequence is linearized.
When reading the data written from the `insert_quorum`, you can use the [select_sequential_consistency](#settings-select_sequential_consistency) option.
- If the number of available replicas at the time of the query is less than the `insert_quorum`.
- At an attempt to write data when the previous block has not yet been inserted in the `insert_quorum` of replicas. This situation may occur if the user tries to perform an `INSERT` before the previous one with the `insert_quorum` is completed.
Quorum write timeout in seconds. If the timeout has passed and no write has taken place yet, ClickHouse will generate an exception and the client must repeat the query to write the same block to the same or any other replica.
When sequential consistency is enabled, ClickHouse allows the client to execute the `SELECT` query only for those replicas that contain data from all previous `INSERT` queries executed with `insert_quorum`. If the client refers to a partial replica, ClickHouse will generate an exception. The SELECT query will not include data that has not yet been written to the quorum of replicas.
Limits the data volume (in bytes) that is received or transmitted over the network when executing a query. This setting applies for every individual query.
Limits speed of data exchange over the network in bytes per second. This setting applies for all concurrently running queries performed by a single user.
1. Rewriting queries for join from the syntax with commas to the `JOIN ON/USING` syntax. If the setting value is 0, ClickHouse doesn't process queries with syntax that uses commas, and throws an exception.
2. Converting `CROSS JOIN` to `INNER JOIN` if `WHERE` conditions allow it.
Specifies which of the `uniq*` functions should be used for performing the [COUNT(DISTINCT ...)](../../query_language/agg_functions/reference.md#agg_function-count) construction.
