Basic Clickhouse support for use as a backend with the data-linkage and deduplication package Splink.
Supports clickhouse server connected via clickhouse connect.
Also supports in-process chDB version if installed with the chdb extras.
Install from PyPI using pip:
# just installs the Clickhouse server dependencies
pip install splinkclickhouse
# or to install with support for chdb:
pip install splinkclickhouse[chdb]or you can install the package directly from github:
# Replace with any version you want, or specify a branch after '@'
pip install git+https://github.com/ADBond/[email protected]If instead you are using conda, splinkclickhouse is available on conda-forge:
conda install conda-forge::splinkclickhouseNote that the conda version will only be able to use the Clickhouse server functionality as chdb is not currently available within conda.
While the package is in early development there will may be breaking changes in new versions without warning, although these should only occur in new minor versions. Nevertheless if you depend on this package it is recommended to pin a version to avoid any disruption that this may cause.
Import ClickhouseAPI, which accepts a clickhouse_connect client, configured with attributes relevant for your connection:
import clickhouse_connect
import splink.comparison_library as cl
from splink import Linker, SettingsCreator, block_on, splink_datasets
from splinkclickhouse import ClickhouseAPI
df = splink_datasets.fake_1000
conn_atts = {
"host": "localhost",
"port": 8123,
"username": "splinkognito",
"password": "splink123!",
}
db_name = "__temp_splink_db"
default_client = clickhouse_connect.get_client(**conn_atts)
default_client.command(f"CREATE DATABASE IF NOT EXISTS {db_name}")
client = clickhouse_connect.get_client(
**conn_atts,
database=db_name,
)
db_api = ClickhouseAPI(client)
# can have at most one tf-adjusted comparison, see caveats below
settings = SettingsCreator(
link_type="dedupe_only",
comparisons=[
cl.JaroWinklerAtThresholds("first_name"),
cl.JaroAtThresholds("surname"),
cl.DateOfBirthComparison(
"dob",
input_is_string=True,
),
cl.DamerauLevenshteinAtThresholds("city").configure(
term_frequency_adjustments=True
),
cl.JaccardAtThresholds("email"),
],
blocking_rules_to_generate_predictions=[
block_on("first_name", "dob"),
block_on("surname"),
],
)
linker = Linker(df, settings, db_api=db_api)See Splink documentation for use of the Linker.
To use chdb as a Splink backend you must install the chdb package.
This is automatically installed if you install with the chdb extras (pip install splinkclickhouse[chdb]).
Import ChDBAPI, which accepts a connection from chdb.api:
import splink.comparison_library as cl
from chdb import dbapi
from splink import Linker, SettingsCreator, block_on, splink_datasets
from splinkclickhouse import ChDBAPI
con = dbapi.connect()
db_api = ChDBAPI(con)
df = splink_datasets.fake_1000
settings = SettingsCreator(
link_type="dedupe_only",
comparisons=[
cl.NameComparison("first_name"),
cl.JaroAtThresholds("surname"),
cl.DateOfBirthComparison(
"dob",
input_is_string=True,
),
cl.DamerauLevenshteinAtThresholds("city").configure(
term_frequency_adjustments=True
),
cl.EmailComparison("email"),
],
blocking_rules_to_generate_predictions=[
block_on("first_name", "dob"),
block_on("surname"),
],
)
linker = Linker(df, settings, db_api=db_api)See Splink documentation for use of the Linker.
splinkclickhouse is compatible with all of the in-built splinks comparisons and comparison levels in splink.comparison_library and splink.comparison_level_library.
However, splinkclickhouse provides a few pre-made extras to leverage Clickhouse-specific functionality.
These can be used in exactly the same way as the native Splink libraries, for example:
import splink.comparison_library as cl
from splink import SettingsCreator
import splinkclickhouse.comparison_library as cl_ch
...
settings = SettingsCreator(
link_type="dedupe_only",
comparisons=[
cl.ExactMatch("name"),
cl_ch.DistanceInKMAtThresholds(
"latitude",
"longitude",
[10, 50, 100, 200, 500],
),
],
)or with individual comparison-levels:
import splink.comparison_level_library as cll
import splink.comparison_library as cl
from splink import SettingsCreator
import splinkclickhouse.comparison_level_library as cll_ch
...
settings = SettingsCreator(
link_type="dedupe_only",
comparisons=[
cl.ExactMatch("name"),
cl.CustomComparison(
comparison_levels = [
cll.And(
cll.NullLevel("city"),
cll.NullLevel("postcode"),
cll.Or(cll.NullLevel("latitude"), cll.NullLevel("longitude"))
),
cll.ExactMatch("postcode"),
cll_ch.DistanceInKMLevel("latitude", "longitude", 5),
cll_ch.DistanceInKMLevel("latitude", "longitude", 10),
cll.ExactMatch("city"),
cll_ch.DistanceInKMLevel("latitude", "longitude", 50),
cll.ElseLevel(),
],
output_column_name="location",
),
],
)If you have difficulties with the package you can open an issue. You may also suggest changes by opening a PR, although it may be best to discuss in an issue beforehand.
This package is 'unofficial', in that it is not directly supported by the Splink team. Maintenance / improvements will be done on a 'best effort' basis where resources allow.
Clickhouse offers several different date formats.
The basic Date format cannot handle dates before the Unix epoch (1970-01-01), which makes it unsuitable for many use-cases for holding date-of-births.
The parsing function parseDateTime (and variants) which support providing custom formats return a DateTime, which also has the above limited range.
In splinkclickhouse we use the function parseDateTime64BestEffortOrNull so that we can use the extended-range DateTime64 data type, which supports dates back to 1900-01-01, but does not allow custom date formats. Currently no DateTime64 equivalent of parseDateTime exists.
If you require different behaviour (for instance if you have an unusual date format and know that you do not need dates outside of the DateTime range) you will either need to derive a new column in your source data, or construct the relevant SQL expression manually.
There is not currently a way in Clickhouse to deal directly with date values before 1900. However, splinkclickhouse offers some tools to help with this.
It creates a SQL UDF (which can be opted-out of) days_since_epoch, to convert a date string (in YYYY-MM-DD format) into an integer, representing the number of days since 1970-01-01 to handle dates well outside the range of DateTime64, based on the proleptic Gregorian calendar.
This can be used with column expression extension splinkclickhouse.column_expression.ColumnExpression via the transform .parse_date_to_int(), or using custom versions of Splink library functions cll.AbsoluteDateDifferenceLevel, cl.AbsoluteDateDifferenceAtThresholds, and cl.DateOfBirthComparison.
These functions can be used with string columns (which will be wrapped in the above parsing function), or integer columns if the conversion via days_since_epoch is already done in the data-preparation stage.
When passing data into chdb from pandas or pyarrow tables, NULL values in String columns are converted into empty strings, instead of remaining NULL.
For now this is not handled within the package. You can workaround the issue by wrapping column names in NULLIF:
import splink.comparison_level as cl
first_name_comparison = cl.DamerauLevenshteinAtThresholds("NULLIF(first_name, '')")Currently at most one term frequency adjustment can be used with ClickhouseAPI.
This also applies to ChDBAPI but only in debug_mode. With debug_mode off there is no limit on term frequency adjustments.
ClickhouseAPI will allow registration of pandas dataframes, by inferring the types of columns. It currently only does this for string, integer, and float columns, and will always make them Nullable.
If you require other data types, or more fine-grained control, it is recommended to import the data into Clickhouse yourself, and then pass the table name (as a string) to the Linker instead.