SQLAlchemy - 在 postgresql 中执行批量更新插入（如果存在，更新，否则插入）

Question

我正在尝试使用 SQLAlchemy 模块（而不是 SQL！）在 python 中编写批量更新插入。

我在 SQLAlchemy 添加上收到以下错误：

sqlalchemy.exc.IntegrityError: (IntegrityError) duplicate key value violates unique constraint "posts_pkey"
DETAIL:  Key (id)=(TEST1234) already exists.

我有一个名为posts的表，在id列上有一个主键。

在这个例子中，我已经在数据库中有一行id=TEST1234 。 当我尝试使用db.session.add()将id设置为TEST1234的新帖子对象时，出现上述错误。 我的印象是，如果主键已经存在，记录就会更新。

如何仅基于主键使用 Flask-SQLAlchemy 进行更新插入？ 有简单的解决方案吗？

如果没有，我总是可以检查并删除具有匹配 id 的任何记录，然后插入新记录，但这对于我的情况来说似乎很昂贵，我不希望有很多更新。

Answer 1

SQLAlchemy 中有一个 upsert-esque 操作：

db.session.merge()

找到这个命令后，我就可以执行 upserts，但值得一提的是，这个操作对于批量“upsert”来说很慢。

另一种方法是获取您要更新插入的主键列表，并查询数据库以查找任何匹配的 id：

# Imagine that post1, post5, and post1000 are posts objects with ids 1, 5 and 1000 respectively
# The goal is to "upsert" these posts.
# we initialize a dict which maps id to the post object

my_new_posts = {1: post1, 5: post5, 1000: post1000} 

for each in posts.query.filter(posts.id.in_(my_new_posts.keys())).all():
    # Only merge those posts which already exist in the database
    db.session.merge(my_new_posts.pop(each.id))

# Only add those posts which did not exist in the database 
db.session.add_all(my_new_posts.values())

# Now we commit our modifications (merges) and inserts (adds) to the database!
db.session.commit()

Answer 2

您可以利用on_conflict_do_update变体。 一个简单的例子如下：

from sqlalchemy.dialects.postgresql import insert

class Post(Base):
    """
    A simple class for demonstration
    """

    id = Column(Integer, primary_key=True)
    title = Column(Unicode)

# Prepare all the values that should be "upserted" to the DB
values = [
    {"id": 1, "title": "mytitle 1"},
    {"id": 2, "title": "mytitle 2"},
    {"id": 3, "title": "mytitle 3"},
    {"id": 4, "title": "mytitle 4"},
]

stmt = insert(Post).values(values)
stmt = stmt.on_conflict_do_update(
    # Let's use the constraint name which was visible in the original posts error msg
    constraint="post_pkey",

    # The columns that should be updated on conflict
    set_={
        "title": stmt.excluded.title
    }
)
session.execute(stmt)

有关更多详细信息，请参阅PG 文档（例如“排除”术语的来源）。

关于重复列名的旁注

上面的代码使用列名作为values列表中的字典键和set_的参数。 如果在类定义中更改了列名，则需要在任何地方更改，否则会中断。 这可以通过访问列定义来避免，使代码更难看，但更健壮：

coldefs = Post.__table__.c

values = [
    {coldefs.id.name: 1, coldefs.title.name: "mytitlte 1"},
    ...
]

stmt = stmt.on_conflict_do_update(
    ...
    set_={
        coldefs.title.name: stmt.excluded.title
        ...
    }
)

Answer 3

使用编译扩展的另一种方法（ https://docs.sqlalchemy.org/en/13/core/compiler.html ）：

from sqlalchemy.ext.compiler import compiles
from sqlalchemy.sql.expression import Insert

@compiles(Insert)
def compile_upsert(insert_stmt, compiler, **kwargs):
    """
    converts every SQL insert to an upsert  i.e;
    INSERT INTO test (foo, bar) VALUES (1, 'a')
    becomes:
    INSERT INTO test (foo, bar) VALUES (1, 'a') ON CONFLICT(foo) DO UPDATE SET (bar = EXCLUDED.bar)
    (assuming foo is a primary key)
    :param insert_stmt: Original insert statement
    :param compiler: SQL Compiler
    :param kwargs: optional arguments
    :return: upsert statement
    """
    pk = insert_stmt.table.primary_key
    insert = compiler.visit_insert(insert_stmt, **kwargs)
    ondup = f'ON CONFLICT ({",".join(c.name for c in pk)}) DO UPDATE SET'
    updates = ', '.join(f"{c.name}=EXCLUDED.{c.name}" for c in insert_stmt.table.columns)
    upsert = ' '.join((insert, ondup, updates))
    return upsert

这应该确保所有插入语句都表现为 upsert。 这个实现是在 Postgres 方言中实现的，但是对于 MySQL 方言，它应该很容易修改。

Answer 4

这不是最安全的方法，但它非常简单且非常快速。 我只是想有选择地覆盖表格的一部分。 我删除了我知道会发生冲突的已知行，然后从 Pandas 数据框中附加了新行。 您的 Pandas 数据框列名需要与您的 sql 表列名相匹配。

eng = create_engine('postgresql://...')
conn = eng.connect()

conn.execute("DELETE FROM my_table WHERE col = %s", val)
df.to_sql('my_table', con=eng, if_exists='append')

Answer 5

我开始研究这个，我想我已经找到了一种非常有效的方法来在 sqlalchemy 中进行 upserts，混合bulk_insert_mappings和bulk_update_mappings而不是merge 。

import time
import sqlite3

from sqlalchemy.ext.declarative import declarative_base
from sqlalchemy import Column, Integer, String, create_engine
from sqlalchemy.orm import scoped_session, sessionmaker
from contextlib import contextmanager


engine = None
Session = sessionmaker()
Base = declarative_base()


def creat_new_database(db_name="sqlite:///bulk_upsert_sqlalchemy.db"):
    global engine
    engine = create_engine(db_name, echo=False)
    local_session = scoped_session(Session)
    local_session.remove()
    local_session.configure(bind=engine, autoflush=False, expire_on_commit=False)
    Base.metadata.drop_all(engine)
    Base.metadata.create_all(engine)


@contextmanager
def db_session():
    local_session = scoped_session(Session)
    session = local_session()

    session.expire_on_commit = False

    try:
        yield session
    except BaseException:
        session.rollback()
        raise
    finally:
        session.close()


class Customer(Base):
    __tablename__ = "customer"
    id = Column(Integer, primary_key=True)
    name = Column(String(255))


def bulk_upsert_mappings(customers):

    entries_to_update = []
    entries_to_put = []
    with db_session() as sess:
        t0 = time.time()

        # Find all customers that needs to be updated and build mappings
        for each in (
            sess.query(Customer.id).filter(Customer.id.in_(customers.keys())).all()
        ):
            customer = customers.pop(each.id)
            entries_to_update.append({"id": customer["id"], "name": customer["name"]})

        # Bulk mappings for everything that needs to be inserted
        for customer in customers.values():
            entries_to_put.append({"id": customer["id"], "name": customer["name"]})

        sess.bulk_insert_mappings(Customer, entries_to_put)
        sess.bulk_update_mappings(Customer, entries_to_update)
        sess.commit()

    print(
        "Total time for upsert with MAPPING update "
        + str(len(customers))
        + " records "
        + str(time.time() - t0)
        + " sec"
        + " inserted : "
        + str(len(entries_to_put))
        + " - updated : "
        + str(len(entries_to_update))
    )


def bulk_upsert_merge(customers):

    entries_to_update = 0
    entries_to_put = []
    with db_session() as sess:
        t0 = time.time()

        # Find all customers that needs to be updated and merge
        for each in (
            sess.query(Customer.id).filter(Customer.id.in_(customers.keys())).all()
        ):
            values = customers.pop(each.id)
            sess.merge(Customer(id=values["id"], name=values["name"]))
            entries_to_update += 1

        # Bulk mappings for everything that needs to be inserted
        for customer in customers.values():
            entries_to_put.append({"id": customer["id"], "name": customer["name"]})

        sess.bulk_insert_mappings(Customer, entries_to_put)
        sess.commit()

    print(
        "Total time for upsert with MERGE update "
        + str(len(customers))
        + " records "
        + str(time.time() - t0)
        + " sec"
        + " inserted : "
        + str(len(entries_to_put))
        + " - updated : "
        + str(entries_to_update)
    )


if __name__ == "__main__":

    batch_size = 10000

    # Only inserts
    customers_insert = {
        i: {"id": i, "name": "customer_" + str(i)} for i in range(batch_size)
    }

    # 50/50 inserts update
    customers_upsert = {
        i: {"id": i, "name": "customer_2_" + str(i)}
        for i in range(int(batch_size / 2), batch_size + int(batch_size / 2))
    }

    creat_new_database()
    bulk_upsert_mappings(customers_insert.copy())
    bulk_upsert_mappings(customers_upsert.copy())
    bulk_upsert_mappings(customers_insert.copy())

    creat_new_database()
    bulk_upsert_merge(customers_insert.copy())
    bulk_upsert_merge(customers_upsert.copy())
    bulk_upsert_merge(customers_insert.copy())

基准测试结果：

Total time for upsert with MAPPING: 0.17138004302978516 sec inserted : 10000 - updated : 0
Total time for upsert with MAPPING: 0.22074174880981445 sec inserted : 5000 - updated : 5000
Total time for upsert with MAPPING: 0.22307634353637695 sec inserted : 0 - updated : 10000
Total time for upsert with MERGE: 0.1724097728729248 sec inserted : 10000 - updated : 0
Total time for upsert with MERGE: 7.852903842926025 sec inserted : 5000 - updated : 5000
Total time for upsert with MERGE: 15.11970829963684 sec inserted : 0 - updated : 10000

Answer 6

我知道这有点晚了，但我已经建立在@mgoldwasser 给出的答案的基础上，并将其变成了一个可以在任何模型（表）上使用的函数，

def upsert_data(entries, model, key):
    for each in session.query(model).filter(getattr(model, key).in_(entries.keys())).all():
        # Only merge those posts which already exist in the database
        session.merge(entries.pop(getattr(each, key)))

    # Only add those posts which did not exist in the database 
    session.add_all(entries.values())

    # Now we commit our modifications (merges) and inserts (adds) to the database!
    session.commit()

entries的格式应与已接受的答案中解释的格式相同。 它应该是一个字典，其中每个对象的键应该是数据库的主键。

model是您要更新插入的 ORM 模型。

key是表的主键。