Strongly suggest using pangres, which is much more efficient and elegant than this package.
pandas_upsert_to_mysql
Enhanced to_sql
method in pandas DataFrame, for MySQL database only. It provides a relatively convenient upsert (insert or update) feature inplementation through temporary table.
- Whether a record needs to be update or not is determined by primary key or unique constraint
- The MySQL database table structure requires to be well designed (need to use SQLAlchemy)
- The primary key must be MySQL auto-increment. Not allow composite primary key mixing auto-increment and other columns
Installation
pip install pandas_upsert_to_mysql
Usage and Example
Let's use an order table as instance. Here the row_id
is the auto-incremented primary key. order_id
and product_id
make up of the unique contraint (a single order can have more than one kind of product).
-- Here we use native SQL to create the table for illustration convenience. In the actual pratice we need to use SQLAlchemy.
CREATE TABLE `order_info` (
`row_id` int(11) NOT NULL AUTO_INCREMENT COMMENT 'auto_incremented_ID',
`order_id` varchar(5) NOT NULL DEFAULT '-9999' COMMENT 'order_id',
`product_id` varchar(5) NOT NULL DEFAULT '-9999' COMMENT 'product_id',
`qty` int(11) DEFAULT NULL COMMENT 'purchase_quantity',
`refund_qty` int(11) DEFAULT NULL COMMENT 'refund_quantity',
`update_time` timestamp NOT NULL DEFAULT CURRENT_TIMESTAMP ON UPDATE CURRENT_TIMESTAMP COMMENT 'last_update_time',
`create_time` timestamp NOT NULL DEFAULT CURRENT_TIMESTAMP COMMENT 'first_create_time',
PRIMARY KEY (`row_id`),
UNIQUE KEY `main` (`order_id`,`product_id`)
) ENGINE=InnoDB DEFAULT CHARSET=utf8 COMMENT='Order Info'
As time moving, this table needs two operations as below:
- New orders: insert new records
- Existed orders apply refund: update the
refund_qty
('qty' for quantity) field of the specific orders
We insert the base data first. This step can be easily done with pandas official native method to_sql
row_id | order_id | product_id | qty | refund_qty | update_time | create_time |
---|
1 | A0001 | PD100 | 10 | 0 | 2020-06-26 11:11:55 | 2020-06-26 11:11:55 |
2 | A0002 | PD200 | 20 | 0 | 2020-06-26 11:11:55 | 2020-06-26 11:11:55 |
3 | A0002 | PD201 | 22 | 0 | 2020-06-26 11:11:55 | 2020-06-26 11:11:55 |
Assume the database already has the data above. Now we need to upsert the new data below. Pay attention that A0002-PD201 now has two refund_qty, and we have the new record A0003-PD300.
order_id | product_id | qty | refund_qty |
---|
A0001 | PD100 | 10 | 0 |
A0002 | PD200 | 20 | 0 |
A0002 | PD201 | 22 | 2 |
A0003 | PD300 | 30 | 0 |
import pandas_upsert_to_mysql.table as table
from pandas_upsert_to_mysql import Upsert
from connection import Connector
engine = Connector(schema='dev').get_engine()
Upsert(engine=engine).to_mysql(df=table.ExampleOrderTable.new_df,
target_table=table.Order,
temp_table=table.OrderTemp,
if_record_exists='update')
Then we can get the result. Be aware that the update_time
only changed in the record whose row_id=3, completely as expected.
row_id | order_id | product_id | qty | refund_qty | update_time | create_time |
---|
1 | A0001 | PD100 | 10 | 0 | 2020-06-26 11:11:55 | 2020-06-26 11:11:55 |
2 | A0002 | PD200 | 20 | 0 | 2020-06-26 11:11:55 | 2020-06-26 11:11:55 |
3 | A0002 | PD201 | 22 | 2 | 2020-06-26 11:13:19 | 2020-06-26 11:11:55 |
4 | A0003 | PD300 | 30 | 0 | 2020-06-26 11:13:19 | 2020-06-26 11:13:19 |
Caveats
- Any conlumn in the unique constraint must not have null value, otherwise the update will violate the constraint and cause duplicates. This is considered as a bug #8173 of MySQL with long history, which means before we write the DataFrame to the database, we should assign a default value for each column in unique constraint to replace the possible null values.
- Implement using
session scope
of SQLAlchemy. But since the author is not a professional engineer, I still can not gurantee the reliability in concurrent condition - Only tested in pandas >= 1.0.3 and MySQL 5.7 innodb environment. The transaction isolation level is the default REPEATABLE-READ
Origin
Pandas official (up to 1.0.5 version) to_sql
method does not implement upsert feature. Its parameter if_exist
has avaliable values as below:
if_exists: {‘fail’, ‘replace’, ‘append’}, default ‘fail’
How to behave if the table already exists.
-
fail: Raise a ValueError.
-
replace: Drop the table before inserting new values.
-
append: Insert new values to the existing table.
Notice that the replace here takes effect on the whole table rather than each specific row, which means the native method can only truncate the whole table and re-insert the entire DataFrame.
Demand for the upsert feature has been discussed in the official repo long before, see issue #14553. But the discussion has lasted for almost 4 years (up to 1.0.5 version), and the officials still can not deliver this feature in a stable release. For now it is only called "may be fixed by #29636"
In those discussion, the major consideration of the officials are:
- Consistent support for multiple databases, but different databases can have huge difference on their native upsert pratice
- Also due to the disparity of databases, they would only support upsert by primary key in the first plan, not by unique constraints
- The conciseness of API
Therefore it can be expected that, the officials may be conservative on their first implementation for the balance of simplicity ("to protect users that have a poorly designed database") and function abundance: only support upsert by primary key for example. However, since the auto-increment has to be (part of) the primary key in MySQL, you won't be able to distinguish those duplicate records which need to be updated once you define an auto-increment. One possible solution is that the primary key contains only an auto-increment, other columns to decide a unique record serve as the unique constraint. For this goal, this repo gives an unofficial solution only for MySQL, and it is highly probably not a best pratice.
Up to the last commit of this repo, pandas 1.0.5 version's native to_sql
method still does not support upsert.
Reference
- pangres: support multiple databases; upsert by primary key, require to set the primary key as the DataFrame index (In fact it is completely compatible with unique constraint); implement without temporary table, much faster.
- pandabase: support multiple databases; upsert by primary key, require to set the primary key as the DataFrame index; seems not supporting MySQL
- pandas-to-mysql
- pandas-sql
- Pandas-to_sql-upsert