I have a record that I want to exist in the database if it is not there, and if it is there already (primary key exists) I want the fields to be updated to the current state. This is often called an upsert.

The following incomplete code snippet demonstrates what will work, but it seems excessively clunky (especially if there were a lot more columns). What is the better/best way?

Base = declarative_base()
class Template(Base):
    __tablename__ = 'templates'
    id = Column(Integer, primary_key = True)
    name = Column(String(80), unique = True, index = True)
    template = Column(String(80), unique = True)
    description = Column(String(200))
    def __init__(self, Name, Template, Desc):
        self.name = Name
        self.template = Template
        self.description = Desc

def UpsertDefaultTemplate():
    sess = Session()
    desired_default = Template("default", "AABBCC", "This is the default template")
        q = sess.query(Template).filter_by(name = desiredDefault.name)
        existing_default = q.one()
    except sqlalchemy.orm.exc.NoResultFound:
        #default does not exist yet, so add it...
        #default already exists.  Make sure the values are what we want...
        assert isinstance(existing_default, Template)
        existing_default.name = desired_default.name
        existing_default.template = desired_default.template
        existing_default.description = desired_default.description

Is there a better or less verbose way of doing this? Something like this would be great:

sess.upsert_this(desired_default, unique_key = "name")

although the unique_key kwarg is obviously unnecessary (the ORM should be able to easily figure this out) I added it just because SQLAlchemy tends to only work with the primary key. eg: I’ve been looking at whether Session.merge would be applicable, but this works only on primary key, which in this case is an autoincrementing id which is not terribly useful for this purpose.

A sample use case for this is simply when starting up a server application that may have upgraded its default expected data. ie: no concurrency concerns for this upsert.

SQLAlchemy does have a “save-or-update” behavior, which in recent versions has been built into session.add, but previously was the separate session.saveorupdate call. This is not an “upsert” but it may be good enough for your needs.

It is good that you are asking about a class with multiple unique keys; I believe this is precisely the reason there is no single correct way to do this. The primary key is also a unique key. If there were no unique constraints, only the primary key, it would be a simple enough problem: if nothing with the given ID exists, or if ID is None, create a new record; else update all other fields in the existing record with that primary key.

However, when there are additional unique constraints, there are logical issues with that simple approach. If you want to “upsert” an object, and the primary key of your object matches an existing record, but another unique column matches a different record, then what do you do? Similarly, if the primary key matches no existing record, but another unique column does match an existing record, then what? There may be a correct answer for your particular situation, but in general I would argue there is no single correct answer.

That would be the reason there is no built in “upsert” operation. The application must define what this means in each particular case.

SQLAlchemy supports ON CONFLICT with two methods on_conflict_do_update() and on_conflict_do_nothing().

Copying from the documentation:

from sqlalchemy.dialects.postgresql import insert

stmt = insert(my_table).values(user_email="[email protected]", data="inserted data")
stmt = stmt.on_conflict_do_update(

Nowadays, SQLAlchemy provides two helpful functions on_conflict_do_nothing and on_conflict_do_update. Those functions are useful but require you to swich from the ORM interface to the lower-level one – SQLAlchemy Core.

Although those two functions make upserting using SQLAlchemy’s syntax not that difficult, these functions are far from providing a complete out-of-the-box solution to upserting.

My common use case is to upsert a big chunk of rows in a single SQL query/session execution. I usually encounter two problems with upserting:

For example, higher level ORM functionalities we’ve gotten used to are missing. You cannot use ORM objects but instead have to provide ForeignKeys at the time of insertion.

I’m using this following function I wrote to handle both of those issues:

def upsert(session, model, rows):
    table = model.__table__
    stmt = postgresql.insert(table)
    primary_keys = [key.name for key in inspect(table).primary_key]
    update_dict = {c.name: c for c in stmt.excluded if not c.primary_key}

    if not update_dict:
        raise ValueError("insert_or_update resulted in an empty update_dict")

    stmt = stmt.on_conflict_do_update(index_elements=primary_keys,

    seen = set()
    foreign_keys = {col.name: list(col.foreign_keys)[0].column for col in table.columns if col.foreign_keys}
    unique_constraints = [c for c in table.constraints if isinstance(c, UniqueConstraint)]
    def handle_foreignkeys_constraints(row):
        for c_name, c_value in foreign_keys.items():
            foreign_obj = row.pop(c_value.table.name, None)
            row[c_name] = getattr(foreign_obj, c_value.name) if foreign_obj else None

        for const in unique_constraints:
            unique = tuple([const,] + [row[col.name] for col in const.columns])
            if unique in seen:
                return None

        return row

    rows = list(filter(None, (handle_foreignkeys_constraints(row) for row in rows)))
    session.execute(stmt, rows)

I use a “look before you leap” approach:

# first get the object from the database if it exists
# we're guaranteed to only get one or zero results
# because we're filtering by primary key
switch_command = session.query(Switch_Command).\
    filter(Switch_Command.switch_id == switch.id).\
    filter(Switch_Command.command_id == command.id).first()

# If we didn't get anything, make one
if not switch_command:
    switch_command = Switch_Command(switch_id=switch.id, command_id=command.id)

# update the stuff we care about
switch_command.lastseen = datetime.datetime.utcnow()

# This will generate either an INSERT or UPDATE
# depending on whether we have a new object or not

The advantage is that this is db-neutral and I think it’s clear to read. The disadvantage is that there’s a potential race condition in a scenario like the following:

  • we query the db for a switch_command and don’t find one
  • we create a switch_command
  • another process or thread creates a switch_command with the same primary key as ours
  • we try to commit our switch_command

The below works fine for me with redshift database and will also work for combined primary key constraint.

SOURCE : this

Just few modifications required for creating SQLAlchemy engine in the function
def start_engine()

from sqlalchemy import Column, Integer, Date ,Metadata
from sqlalchemy.ext.declarative import declarative_base
from sqlalchemy.dialects.postgresql import insert
from sqlalchemy import create_engine
from sqlalchemy.orm import sessionmaker
from sqlalchemy.dialects import postgresql

Base = declarative_base()

def start_engine():
    engine = create_engine(os.getenv('SQLALCHEMY_URI', 
     connect = engine.connect()
    meta = MetaData(bind=engine)
    return engine

class DigitalSpend(Base):
    __tablename__ = 'digital_spend'
    report_date = Column(Date, nullable=False)
    day = Column(Date, nullable=False, primary_key=True)
    impressions = Column(Integer)
    conversions = Column(Integer)

    def __repr__(self):
        return str([getattr(self, c.name, None) for c in self.__table__.c])

def compile_query(query):
    compiler = query.compile if not hasattr(query, 'statement') else 
    return compiler(dialect=postgresql.dialect())

def upsert(session, model, rows, as_of_date_col="report_date", no_update_cols=[]):
    table = model.__table__

    stmt = insert(table).values(rows)

    update_cols = [c.name for c in table.c
                   if c not in list(table.primary_key.columns)
                   and c.name not in no_update_cols]

    on_conflict_stmt = stmt.on_conflict_do_update(
        set_={k: getattr(stmt.excluded, k) for k in update_cols},
        index_where=(getattr(model, as_of_date_col) < getattr(stmt.excluded, as_of_date_col))


session = start_engine()
upsert(session, DigitalSpend, initial_rows, no_update_cols=['conversions'])

This allows access to the underlying models based on string names

def get_class_by_tablename(tablename):
  """Return class reference mapped to table.
  :param tablename: String with name of table.
  :return: Class reference or None.
  for c in Base._decl_class_registry.values():
    if hasattr(c, '__tablename__') and c.__tablename__ == tablename:
      return c

sqla_tbl = get_class_by_tablename(table_name)

def handle_upsert(record_dict, table):
    handles updates when there are primary key conflicts

        # Here we'll assume the error is caused by an integrity error
        # We do this because the error classes are passed from the
        # underlying package (pyodbc / sqllite) SQLAlchemy doesn't mask
        # them with it's own code - this should be updated to have
        # explicit error handling for each new db engine

        # <update>add explicit error handling for each db engine</update> 
        # Query for conflic class, use update method to change values based on dict
        c_tbl_primary_keys = [i.name for i in table.__table__.primary_key] # List of primary key col names
        c_tbl_cols = dict(sqla_tbl.__table__.columns) # String:Col Object crosswalk

        c_query_dict = {k:record_dict[k] for k in c_tbl_primary_keys if k in record_dict} # sub-dict from data of primary key:values
        c_oo_query_dict = {c_tbl_cols[k]:v for (k,v) in c_query_dict.items()} # col-object:query value for primary key cols

        c_target_record = session.query(sqla_tbl).filter(*[k==v for (k,v) in oo_query_dict.items()]).first()

        # apply new data values to the existing record
        for k, v in record_dict.items()
            setattr(c_target_record, k, v)

This works for me with sqlite3 and postgres. Albeit it might fail with combined primary key constraints and will most likely fail with additional unique constraints.

        t = self._meta.tables[data['table']]
    except KeyError:
        self._log.error('table "%s" unknown', data['table'])

        q = insert(t, values=data['values'])
    except IntegrityError:
        self._log.warning('integrity error')
        where_clause = [c.__eq__(data['values'][c.name]) for c in t.c if c.primary_key]
        update_dict = {c.name: data['values'][c.name] for c in t.c if not c.primary_key}
        q = update(t, values=update_dict).where(*where_clause)
    except Exception as e:
        self._log.error('%s: %s', t.name, e)

There are multiple answers and here comes yet another answer (YAA). Other answers are not that readable due to the metaprogramming involved. Here is an example that

  • Uses SQLAlchemy ORM

  • Shows how to create a row if there are zero rows using on_conflict_do_nothing

  • Shows how to update the existing row (if any) without creating a new row using on_conflict_do_update

  • Uses the table primary key as the constraint

A longer example in the original question what this code is related to.

import sqlalchemy as sa
import sqlalchemy.orm as orm
from sqlalchemy import text
from sqlalchemy.dialects.postgresql import insert
from sqlalchemy.orm import Session

class PairState(Base):

    __tablename__ = "pair_state"

    # This table has 1-to-1 relationship with Pair
    pair_id = sa.Column(sa.ForeignKey("pair.id"), nullable=False, primary_key=True, unique=True)
    pair = orm.relationship(Pair,
                                        cascade="all, delete-orphan",
                                        single_parent=True, ), )

    # First raw event in data stream
    first_event_at = sa.Column(sa.TIMESTAMP(timezone=True), nullable=False, server_default=text("TO_TIMESTAMP(0)"))

    # Last raw event in data stream
    last_event_at = sa.Column(sa.TIMESTAMP(timezone=True), nullable=False, server_default=text("TO_TIMESTAMP(0)"))

    # The last hypertable entry added
    last_interval_at = sa.Column(sa.TIMESTAMP(timezone=True), nullable=False, server_default=text("TO_TIMESTAMP(0)"))

    def create_first_event_if_not_exist(dbsession: Session, pair_id: int, ts: datetime.datetime):
        """Sets the first event value if not exist yet."""
            values(pair_id=pair_id, first_event_at=ts).

    def update_last_event(dbsession: Session, pair_id: int, ts: datetime.datetime):
        """Replaces the the column last_event_at for a named pair."""
        # Based on the original example of https://stackoverflow.com/a/49917004/315168
            values(pair_id=pair_id, last_event_at=ts).
            on_conflict_do_update(constraint=PairState.__table__.primary_key, set_={"last_event_at": ts})

    def update_last_interval(dbsession: Session, pair_id: int, ts: datetime.datetime):
        """Replaces the the column last_interval_at for a named pair."""
            values(pair_id=pair_id, last_interval_at=ts).
            on_conflict_do_update(constraint=PairState.__table__.primary_key, set_={"last_interval_at": ts})

In case of sqlite, the sqlite_on_conflict="REPLACE" option can be used when defining a UniqueConstraint, and sqlite_on_conflict_unique for unique constraint on a single column. Then session.add will work in a way just like upsert. See the official documentation.