In v0.5.0-0.5.3 releases, streaming was implemented by returning generators from the QueryResult methods
stream_column_blocks, stream_row_blocks, and stream_rows. Safe usage of these methods required executing them
within a Python context created with the QueryResult itself. Otherwise, the HTTP response could be left open in the
case of an exception or other failure to consume the stream.
Using QueryResult as a context is unintuitive, and that usage pattern is deprecated and will be completely disabled in
a future release. Instead, streaming query results should be obtained using the new Client *stream methods described
under New Features in the 0.5.4 release, below.
The send_progress keyword argument to the get_client method has been deprecated. It now does nothing and will be removed
in a future release. Starting with 0.5.9 the driver now requests ClickHouse progress headers in all cases to provide
"keep alive" functionality for both long-running streaming query result processing, and long-running insert/DDL queries.
The secondary effect of the send_progress argument -- to set wait_end_of_query=1 -- is now handled automatically based
on whether the query is streaming or not.
- Fix Pandas dataframe inserts where the Dataframe index does not match the data values (after, for example, creating a new DataFrame from a subset of the original.) ClickHouse#167 Thanks to Georgi Peev for the report and suggested fix, and his continued stress testing of Pandas functionality.
- Compression and other control settings were not properly sent with the request if the corresponding setting was not enabled on the server. Many thanks to Alexander Khmelevskiy for the extended investigation and subsequent fix. ClickHouse#157
- Fix quoting and escaping of array literals in server parameters. See #159. Big thanks to Joachim Jablon for the report and the fix.
- Pandas and numpy Date values were incorrect for values after 2050. This has been fixed. ClickHouse#164
- Fixed server side parameter binding of the NULL value for Nullable types
- Added support for
no_proxy/NO_PROXYenvironment variable. Also added support for lower casehttp_proxyandhttps_proxyvariables. Note that lower case versions have precedence over upper case versions. Fixes ClickHouse#163
- The server timezone will not be applied (and Python datetime types will be timezone naive) if the client and server timezones match
and the
get_clientapply_server_timezone parameter is True (the default). This improves performance where client and server have the same (non-UTC) timezone. To override this behavior and always apply a server timezone to the result, useapply_server_timezone='always'. This should fix ClickHouse#157
- The client
query_dfandquery_df_streammethods now acceptquery_tzandcolumn_tzsparameters like otherquery*methods. - A new boolean parameter
apply_server_timezonehas been added to the mainget_clientmethod. Setting this parameter toTrue(the default) will apply the server timezone (if not UTC) to values returned by the clientquery*methods. The previous behavior would always return timezone naive, UTC based Python and Pandasdatetimeobjects for ClickHouse DateTime and DateTime64 columns without a defined timezone. To revert to the previous behavior, set theapply_server_timezoneparameter toFalse. Closes ClickHouse#152 - The timezone logic applied to query results has been simplified and now uses the following order of precedence:
- Use the column timezone for the column if it is specified using the
column_tzsparameter - Use the column timezone for the column if specified in the ClickHouse column definition (only works for ClickHouse versions 23.2 and later)
- Use the query timezone for the query if it is set using the
query_tzparameter - Use the "response" timezone for the query as read from the
X-ClickHouse-Timezoneheader if different from the server timezone. This closes ClickHouse#138. - Use the ClickHouse server timezone (if the client parameter
apply_server_timezoneisTrue)
- Use the column timezone for the column if it is specified using the
- Note if the detected timezone according to the above precedence is UTC,
clickhouse-connectwill always return a naive datetime object with no timezone information
- ClickHouse external data is now support for all client
querymethods. To send external data, construct adriver.external.ExternalDataobject and send it as theexternal_dataparameter in the appropriate query method. See the ClickHouse documentation for additional details. There are also examples in the test file . Closes ClickHouse#98
- Creating a client would fail if for some reason the user did not have access to the
system.settingstable. Thanks to Filipp Balakin for the fix.
- String columns now accept values of bytes-like objects (bytes/bytearray/etc.) for inserts (as with other inserts, all
values for the inserted column should be the same types, either a bytes-like object or
str). A correspondingbytesread format has been enabled for String columns as well. Thanks to Tim Nooran for opening the issue and providing unit tests. ClickHouse#148 - Cython version upgraded to 3.0.0b1
- Remove unnecessary addition of the client database to the table name for inserts. Fixes ClickHouse#145
- The driver should now work for older versions of ClickHouse back to 19.16. Note that older versions are not
officially tested or supported (like the main ClickHouse database, we officially support the last three monthly ClickHouse
releases and the last two LTS ClickHouse releases). For versions prior to 19.17, you may want change the new
readonlyclickhouse_connect.commonsetting to '1' to allow sending ClickHouse settings with individual queries (if the user has write permissions). Thanks to Aleksey Astafiev for this contribution and for updating the tests to run with these legacy versions!
- Remove direct pandas import that caused an unrecoverable error when pandas was not installed. ClickHouse#139
- By default, reading Pandas Dataframes with query_df and query_df_stream now sets a new QueryContext property
of
use_pandas_natoTrue. Whenuse_pandas_nais True, clickhouse_connect will attempt to use Pandas "missing" values, such as pandas.NaT and pandas.NA, for ClickHouse NULLs (in Nullable columns only), and use the associated extended Pandas dtype. Closes ClickHouse#132 - There are new low level optimizations for reading some Nullable columns, and writing Pandas dataframes
- Timezone information from ClickHouse DateTime columns with a timezone was lost. There was a workaround implemented for this issue in v0.5.8 that allowed assigned timezones to the query or columns on the client side. ClickHouse now support sending this timezone data with the column, but only in server versions 23.2 and later. If such a version is detected, clickhouse-connect will return timezone aware DateTime values without a workaround. Fixes ClickHouse#120
- For certain queries, an incorrect, non-zero "zero value" would be returned for queries where
use_nonewas set toFalse. All NULL values are now properly converted. - Timezone data was lost when a DateTime64 column with a timezone was converted to a Pandas DataFrame. This has been fixed. ClickHouse#136
- send_progress headers were not being correctly requested, which could result in unexpected timeouts for long-running queries. This has been fixed.
- A new keyword parameter
server_host_nameis now recognized by theclickhouse_connect.get_clientmethod. This identifies the "real" ClickHouse server hostname that should be used for HTTPS/TLS certificate validation, in cases where access to the server is through an ssh tunnel or other proxy with a different hostname. For examples of how to use the new parameter, see the updated file https://github.com/ClickHouse/clickhouse-connect/blob/main/examples/ssh_tunnels.py.
- The
databaseelement of a DSN was not recognized when present in thedsnparameter ofclickhouse_connect.get_client. This has been fixed.
- Referencing the QueryResult
named_resultsproperty after other properties such asrow_countwould incorrectly raise a StreamClosedError. Thanks to Stas for the fix.
- A better error message is returned when trying to read a "non-standard" DateTime64 column function for a numpy array
or Pandas DataFrame. "non-standard" means a DateTime64 precision not conforming to seconds, milliseconds, microseconds,
or nanoseconds (0, 3, 6, or 9 respectively). These DateTime64 types are not supported for numpy or Pandas because there is
no corresponding standard numpy datetime64 type and conversion would be unacceptably slow (supported numpy types are
datetime64[s],datetime64[ms],datetime64[us], anddatetime64[ns]). A workaround is to cast the DateTime64 type to a supported type, i.e.SELECT toDateTime64(col_name, 3)for a millisecond column. - The base configuration required for a urllib PoolManager has been broken out into its own help method,
clickhouse_connect.driver.http_util.get_pool_manager_options. This makes it simpler to configure a SOCKSProxyManager as in the new example file https://github.com/ClickHouse/clickhouse-connect/blob/main/examples/ssh_tunnels.py
- Reading Nullable(String) columns has been optimized and should be approximately 2x faster. (This does yet not include LowCardinality(Nullable(String)) columns.)
- Extraction of ClickHouse error messages included in the HTTP Response has been improved
- When reading native Python integer columns, the
use_none=Falsequery parameter would not be respected, and ClickHouse NULLS would be returned as None instead of 0.use_none=Falseshould now work correctly for Nullable(Int) columns - Starting with release 0.5.0, HTTP Connection pools were not always cleanly closed on exit. This has been fixed.
- Large query results using
zstdcompression incorrectly buffered all incoming data at the start of the query, consuming an excessive amount of memory. This has been fixed. ClickHouse#122 Big thanks to Denny Crane for his detailed investigation of the problem. Note that this affected large queries using the defaultcompress=Trueclient setting, as ClickHouse would preferzstdcompression in those cases. - Fixed an issue where a small query_limit would break client initialization due to an incomplete read of the
system.settingstable. ClickHouse#123
- Stream error handling has been improved so exceptions thrown while consuming a stream should be correctly propagated. This includes unexpected stream closures by the ClickHouse server. Errors inserted into the HTTP response by ClickHouse during a query should also be reported as part of a StreamFailureError
- Return empty dataframe instead of empty list when no records returned from
query_dfmethod Fixes ClickHouse#118
- The client
query_limitnow defaults to 0 (unlimited rows returned), since the previous default of 5000 was unintuitive and led to confusion when limited results were returned.
- Allow client side control of datetime.datetime timezones for query results. The client
querymethods for native Python results now accept two new parameters:query_tzis the timezone to be assigned for any DateTime or DateTime64 objects in the results, while timezones can be set per column using thecolumn_tzsdictionary of column names to timezones. See the test file for simple examples. This is a workaround for ClickHouse#120 and the underlying ClickHouse issue ClickHouse/ClickHouse#40397 Note that this issue only affects DateTime columns, not DateTime64, although the query context parameters will override the returned DateTime64 timezone as well.
- Http proxies did not work after removing the requests library. ClickHouse#114.
This should be fixed. Note that socks proxies are still not supported directly, but can be added by creating a correctly
configured urllib3 SOCKSProxyManager and using it as the
pool_mgrargument to tehclickhouse_connect.create_clientmethod.
- Dataframe inserts would incorrectly modify null-like elements of the inserted dataframe. ClickHouse#112. This should be fixed
- Queries of LowCardinality columns using pandas or numpy query methods would result in an exception. ClickHouse#108 This has been fixed.
- Several streaming query methods have been added to the core ClickHouse Connect client. Each of these methods returns a StreamContext object, which must be used as a Python
withContext to stream data (this ensures the underlying streaming response is properly closed/consumed.) For simple examples, see the basic tests.query_column_block_stream-- returns a generator of blocks in column oriented (Native) format. Fastest method for retrieving data in native Python formatquery_row_block_stream-- returns a generator of blocks in row oriented format. Used for processing data in a "batch" of rows at time while limiting memory usagequery_rows_stream-- returns a convenience generator to process rows one at a time (data is still loaded in ClickHouse blocks to preserve memory)query_np_stream-- returns a generator where each ClickHouse data block is transformed into a Numpy arrayquery_df_stream-- returns a generator where each ClickHouse data block is transformed into a Pandas Dataframe
- The
client_nameis now reported in a standardized way to ClickHouse (as thehttp_user_agent). For better tracking of your Python application, use the newproduct_namecommon setting or setclient_nameget_clientparameter to identify your product as<your-product-name>/<product-version>.
- C/Cython optimizations for transforming ClickHouse data to Python types have been improved, and additional datatypes have been optimized in Cython. The performance increase over the previous 0.5.x version is approximately 10% for "normal" read queries.
- Transformation of Numpy arrays and Pandas Dataframes has been completely rewritten to avoid an intermediate conversion to Python types. As a result, querying in Numpy format, and especially Pandas format, has been significantly improved -- from 2x for small datasets to 5x or more for very large Pandas DataFrames (even without streaming). Queries including Numpy datetime64 or Pandas Timestamp objects have particularly benefited from the new implementation.
- The default
maxsizefor concurrent HTTP connections to a single host was accidentally dropped in the 0.5.x release. It has been restored to 8 for better performance when using multiple client objects. - A single low level retry has been restored for HTTP connections on ConnectionReset or RemoteDisconnected exceptions. This should reduce connection errors related to ClickHouse closing expired KeepAlive connections.
- As noted above, streaming, contexts and exception handling have been tightened up to avoid leaving HTTP responses open when querying streams.
- Previous versions used
threading.local()variables to store context information during query processing. The architecture has been changed to pass the relevant Query or Insert Context to transformation methods instead of relying on thread local variables. This is significantly safer in an environment where multiple queries can conceivably be open at the same on the same thread (for example, if using async functions). - Per query formatting logic has moved from
ClickHouseTypeto theQueryContext. ClickHouseTypemethods have been renamed to remove outdated references tonativeformat (everything is native now)- Upgraded Cython Build to 3.0.11alpha release
- Correctly return QueryResult object when created as a context using a
withstatement. This fixes examples and the preferred context syntax for processing query results. Thanks to John McCann Cunniff Jr
- Fix issue where client database is set to None (this normally only happens when deleting the initial database)
- Fix ping check in http client. Closes ClickHouse#96.
The clickhouse_connect get_client method (which proxies the driver.Client constructor) previously accepted arbitrary
keyword arguments that were interpreted as ClickHouse server settings sent with every request. To be consistent with
other client methods, get_client now accepts an optional settings Dict[str, Any] argument that should be used instead
to set ClickHouse server settings.
The driver.HttpClient constructor previously accepted the optional keyword argument http_adapter, which could be used to
pass a custom requests.adapter.HttpAdapter to the client. ClickHouse Connect no longer uses the requests library (see
Dependency Changes below). Instead, the HttpClient constructor now accepts an optional pool_mgr keyword argument which
can be used to set a custom urllib.poolmanager.PoolManager for the client. In most cases the default PoolManager is
all that is needed, but multiple PoolManagers may be required for advanced server/proxy applications with many client instances.
- ClickHouse Connect no longer requires the popular
requestslibrary. Therequestslibrary is built on urllib3, but ClickHouse Connect was utilizing very little of the added functionality. Requests also has very restricted access to theurllib3streaming API, which made adding additional compression methods difficult. Accordingly, the project now interfacdes tourllib3directly. This should not change the public API (except as noted in the warning above), but the HttpClient internals have changed to use the lower level library. - ClickHouse Connect now requires the zstandard and lz4 binding libraries to support zstd and lz4 compression. ClickHouse itself uses these compression algorithms extensively and is optimized to work with them, so ClickHouse Connect now takes advantages of them when compression is desired.
- The core client
querymethod now supports streaming. The returnedQueryResultobject has new streaming methods:stream_column_blocks- returns a generator of smaller result sets matching the ClickHouse blocks returned by the native interface.stream_row_blocks- returns a generator of smaller result sets matching the ClickHouse blocks returned by the native interface, but "pivoted" to return data rows.stream_rows- returns a generator that returns a row of data with each iteration.
These methods should be used within awithcontext to ensure the stream is properly closed when done. In addition, two new propertiesresult_columnsandresult_rowshave been added toQueryResult. Referencing either of these properties will consume the stream and return the full dataset. Note that these properties should be used instead of the ambiguousresult_set, which returns the data oriented based on thecolumn_orientedboolean property. With the addition ofresult_rowsandresult_columnstheresult_setproperty and thecolumn_orientedproperty are unnecessary and may be removed in a future release.
- More compression methods. As noted above, ClickHouse Connect now supports
zstdandlz4compression, as well as brotli (br), if the brotli library is installed. If the clientcompressmethod is set toTrue(the default), ClickHouse Connect will request compression from the ClickHouse server in the orderlz4,zstd,br,gzip,deflate, and will compress inserts to ClickHouse usinglz4. Otherwise, the clientcompressargument can be set to any oflz4,zstd,br, orgzip, and the specific compression method will be used for both queries and inserts. Whilegzipis available, it doesn't perform as well as the other options and should normally not be used.
- More data conversions for query data have been ported to optimized C/Cython code. Rough benchmarks suggest that this improves query performance approximately 20% for standard data types.
- Using the new streaming API to process data in blocks significantly improves performance for large datasets (largely because Python has to allocate significantly less memory and do much less internal data copying otherwise required to build and hold the full dataset). For datasets of a million rows or more, streaming can improve query performance 2x or more.
- As mentioned, ClickHouse
gzipperformance is poor compared tolz4andzstd. Using those compression methods by default avoids the major performance degradation seen in ClickHouse#89. - Passing SqlAlchemy query parameters to the driver.Client constructor was broken by changes in release 0.4.8. ClickHouse#94. This has been fixed.
- Documentation has been expanded to cover recent updates.
- File upload support. The new
driver.toolsmodule adds the functioninsert_fileto simplify directly inserting data files into a table. See the test file for examples. This closes ClickHouse#41. - Added support for server side http query parameters
For queries that contain bindings of the form
{<name>:<datatype>}, the client will automatically convert the query* methodparametersdictionary to the appropriate http query parameters. Closes ClickHouse#49. - The main
clickhouse_connect.get_clientcommand will now accept a standard Pythondsnargument and extract host, port, user, password, and settings (query parameters) from the dsn. Note that values for other keyword parameters will take precedence over values extracted from the dsn. - The QueryResult object now contains convenience properties for the
first_item,first_row, androw_countin the result.
- JSON inserts with the ujson failed, this has been fixed. ClickHouse#84
- The JSON/Object datatype now supports writes using JSON strings as well as Python native types
- Fixed a major settings issue with connecting to a readonly database (introduced in v0.4.4)
- Fix for broken database setup dialog with recent Superset versions using SQLAlchemy 1.4
- Common settings were stored in an immutable named tuple and could not be changed. This is fixed.
- Fixed issue where the query_arrow method would not use the client database
- Ignore all "transport settings" when validating settings. This should fix ClickHouse#80 for older ClickHouse versions
- The get_client method now accepts a http_adapter parameter to allow sharing a requests.HTTPAdapter (and its associated connection pool) across multiple clients.
- The VERSION file is now included in every package installation. Closes ClickHouse#76
- Global/common configuration options are now available in the
clickhouse_connect.commonmodule. The available settings are:autogenerate_session_id[bool] Whether to generate a UUID1 session id used for every client request. Defaults to True. Disabling this can facilitate client sharing and load balancing in some use cases.dict_parameter_format[str] Options are 'json' and 'map'. This controls whether parameterized queries convert a Python dictionary to JSON or ClickHouse Map syntax. Default tojsonfor insert into Object('json') columns.invalid_setting_action[str] Options are 'send' and 'drop'. Client Connect normally validates and drops (with a warning any settings that aren't recognized by the Server or are readonly). Changing this setting to 'send' will include such settings with the request anyway -- which will normally result in an error being returned.
- The
clickhouse_connect.get_clientmethod now accepts asettingsdictionary argument for consistency with other client methods.
- Fixed insert of Pandas Dataframes for Timestamp columns with timezones ClickHouse#77
- Fixed exception when inserting a Pandas Dataframes with NaType values into ClickHouse Float column (see known issue)
When inserting Pandas DataFrame values into a ClickHouse Nullable(Float*) column, a Float NaN value will be converted to a ClickHouse NULL.
This is a side effect of a Pandas issue where df.replace cannot distinguish between NaT and NaN values: pandas-dev/pandas#29024
- Numpy array read and write compatibility has been refined and performance has been improved. This fixes ClickHouse#69
- Pandas Timestamp objects are now correctly handled for all supported ClickHouse Date* types. This fixes ClickHouse#68
- SQLAlchemy datatypes are now correctly mapped to the underlying ClickHouse type regardless of case. This fixes an issue with migrating Superset datasets and queries from clickhouse-sqlalchemy to clickhouse-connect. Thanks to Eugene Torap
- The settings, table information, and insert progress used for client inserts has been centralized in a new reusable InsertContext object. Client insert methods can now accept such objects to simplify code and reduce overhead
- Query results can now be returned in a column oriented format. This is useful to efficiently construct other objects (like Pandas dataframes) that use column storage internally
- The transformation of Pandas data to Python types now bypasses Numpy. As a result compatibility for ClickHouse date, integer, and NULL types has been significantly improved
- An insert using chunked transfer encode could fail in progress during serialization to ClickHouse native format. This would "hang" the request after throwing the exception, leading to ClickHouse reporting "concurrent session" errors. This has been fixed.
- Pandas DataFrame inserts into tables with a "large" integer column would throw an exception. This has been fixed.
- Pandas DataFrame inserts with NaT/NA/nan values would fail, even if inserted into Nullable column types. This has been fixed.
- Numpy inserts into large integer columns are not supported. ClickHouse#69
- Insert of Pandas timestamps with nanosecond precision will lose the nanosecond value. ClickHouse#68
- Fix read compression typo
- Insert performance and memory usage for large inserts has been significantly improved
- Insert blocks now use chunked transfer encoding (by sending a generator instead of a bytearray to the requests POST method)
- If the client is initialized with compress = True, gzip compression is now enabled for inserts
- Pandas DataFrame inserts have been optimized by keep the data in columnar format during the entire insert process
- Fix inserts for date and datetime columns from Pandas dataframes.
- Fix serialization issues for Decimal128 and Decimal256 types
- Update QueryContext.updated_copy method to preserve settings, parameters, etc. ClickHouse#65
- Build Python 3.11 Wheels
- Correctly handle insert into JSON/Object('json') column via SQLAlchemy
- Fix some incompatibilities with SQLAlchemy 1.4
- Fix 'SHOW CREATE' issue. ClickHouse#61
- "Queries" that do not return data results (like DDL and SET queries) are now automatically treated as commands. Closes ClickHouse#59
- A UUID session_id is now generated by default if
session_idis not specified inclickhouse_connect.get_client - Test infrastructure has been simplified and test configuration has moved from pytest options to environment files
- UInt64 types were incorrectly returned as signed Python ints even outside of Superset. This has been fixed
- Superset Engine Spec will now format (U)Int256 and (U)Int128 types as strings to avoid throwing a conversion exception
- The row_binary option for ClickHouse serialization has been removed. The performance is significantly lower than Native format and maintaining the option added complexity with no corresponding benefit
- The Database Connection dialog was broken in the latest Superset development builds. This has been fixed
- IPv6 Addresses fixed for default Superset configuration
- Add single retry for HTTP RemoteDisconnected errors from the ClickHouse Server. This prevents exception spam when requests (in particular inserts) are sent at approximately the same time as the ClickHouse server closes a keep alive connection.
- Fix incorrect validation errors in the Superset connection dialog
- This release updates the build process to include binary wheels for the majority of platforms, include MacOS M1 and Linux Aarch64. This should also fix installation errors on lightweight platforms without build tools.
- Builds are now included for Python 3.11
- Docker images built on MacOS directly from source do not correctly build the C extensions for Linux. However, installing the official wheels from PyPI should work correctly.
- The HTTP client now raises an OperationalError instead of a DatabaseError when the HTTP status code is 429 (too many requests), 503 (service unavailable), or 504 (gateway timeout) to make it easier to determine if it is a retryable exception
- Add
query_retriesclient parameter (default 2) for "retryable" HTTP queries. Does not apply to "commands" like DDL or to inserts
- Fixed an SQLAlchemy dialect issue with SQLAlchemy 1.4 that would cause problems in the most recent Superset version
- Fixed an issue where DBAPI cursors returned an invalid description object for columns. This would cause
'property' object has no attribute 'startswith'errors for some SqlAlchemy and SuperSet queries. - Fixed an issue where datetime parameters would not be correctly rendered as ClickHouse compatible strings
- The "parameters" object passed to client query methods can now be a sequence instead of a dictionary, for compatibility with query strings that contain simple format unnamed format directives, such as
'SELECT * FROM table WHERE value = %s'
- The wait_end_of_query parameter/setting was incorrectly being stripped. This is fixed
- Fix encoding insert of multibyte characters
- Improve identifier handling/quoting for Clickhouse column, table, and database names
- Add client arrow_insert method to directly insert a PyArrow Table insert ClickHouse using Arrow format
- Fix issue when query_limit set to 0
- Fix SQL comment problems in DBAPI cursor
- Support (experimental) JSON/Object datatype. ClickHouse Connect will take advantage of the fast orjson library if available. Note that inserts for JSON columns require ClickHouse server version 22.6.1 or later
- Standardize read format handling and allow specifying a return data format per column or per query.
- Added convenience min_version method to client to see if the server is at least the requested level
- Increase default HTTP timeout to 300 seconds to match ClickHouse server default
- Fixed multiple issues with SQL comments that would cause some queries to fail
- Fixed problem with SQLAlchemy literal binds that would cause an error in Superset filters
- Fixed issue with parameterized queries
- Named Tuples were not supported and would result in throwing an exception. This has been fixed.
- The client query_arrow function would return incomplete results if the query result exceeded the ClickHouse max_block_size. This has been fixed. As part of the fix query_arrow method returns a PyArrow Table object. While this is a breaking change in the API it should be easy to work around.
- Support Nested data types.
- Fix issue with native reads of Nullable(LowCardinality) numeric and date types.
- Empty inserts will now just log a debug message instead of throwing an IndexError.