Error Messages

This section lists descriptions and background for common error messages and warnings raised or emitted by SQLAlchemy.

SQLAlchemy normally raises errors within the context of a SQLAlchemy-specific exception class. For details on these classes, see Core Exceptions and ORM Exceptions.

SQLAlchemy errors can roughly be separated into two categories, the programming-time error and the runtime error. Programming-time errors are raised as a result of functions or methods being called with incorrect arguments, or from other configuration-oriented methods such as mapper configurations that can’t be resolved. The programming-time error is typically immediate and deterministic. The runtime error on the other hand represents a failure that occurs as a program runs in response to some condition that occurs arbitrarily, such as database connections being exhausted or some data-related issue occurring. Runtime errors are more likely to be seen in the logs of a running application as the program encounters these states in response to load and data being encountered.

Since runtime errors are not as easy to reproduce and often occur in response to some arbitrary condition as the program runs, they are more difficult to debug and also affect programs that have already been put into production.

Within this section, the goal is to try to provide background on some of the most common runtime errors as well as programming time errors.

Connections and Transactions

QueuePool limit of size <x> overflow <y> reached, connection timed out, timeout <z>

This is possibly the most common runtime error experienced, as it directly involves the work load of the application surpassing a configured limit, one which typically applies to nearly all SQLAlchemy applications.

The following points summarize what this error means, beginning with the most fundamental points that most SQLAlchemy users should already be familiar with.

  • The SQLAlchemy Engine object uses a pool of connections by default - What this means is that when one makes use of a SQL database connection resource of an Engine object, and then releases that resource, the database connection itself remains connected to the database and is returned to an internal queue where it can be used again. Even though the code may appear to be ending its conversation with the database, in many cases the application will still maintain a fixed number of database connections that persist until the application ends or the pool is explicitly disposed.

  • Because of the pool, when an application makes use of a SQL database connection, most typically from either making use of Engine.connect() or when making queries using an ORM Session, this activity does not necessarily establish a new connection to the database at the moment the connection object is acquired; it instead consults the connection pool for a connection, which will often retrieve an existing connection from the pool to be re-used. If no connections are available, the pool will create a new database connection, but only if the pool has not surpassed a configured capacity.

  • The default pool used in most cases is called QueuePool. When you ask this pool to give you a connection and none are available, it will create a new connection if the total number of connections in play are less than a configured value. This value is equal to the pool size plus the max overflow. That means if you have configured your engine as:

    1. engine = create_engine("mysql+mysqldb://u:p@host/db", pool_size=10, max_overflow=20)

    The above Engine will allow at most 30 connections to be in play at any time, not including connections that were detached from the engine or invalidated. If a request for a new connection arrives and 30 connections are already in use by other parts of the application, the connection pool will block for a fixed period of time, before timing out and raising this error message.

    In order to allow for a higher number of connections be in use at once, the pool can be adjusted using the create_engine.pool_size and create_engine.max_overflow parameters as passed to the create_engine() function. The timeout to wait for a connection to be available is configured using the create_engine.pool_timeout parameter.

  • The pool can be configured to have unlimited overflow by setting create_engine.max_overflow to the value “-1”. With this setting, the pool will still maintain a fixed pool of connections, however it will never block upon a new connection being requested; it will instead unconditionally make a new connection if none are available.

    However, when running in this way, if the application has an issue where it is using up all available connectivity resources, it will eventually hit the configured limit of available connections on the database itself, which will again return an error. More seriously, when the application exhausts the database of connections, it usually will have caused a great amount of resources to be used up before failing, and can also interfere with other applications and database status mechanisms that rely upon being able to connect to the database.

    Given the above, the connection pool can be looked at as a safety valve for connection use, providing a critical layer of protection against a rogue application causing the entire database to become unavailable to all other applications. When receiving this error message, it is vastly preferable to repair the issue using up too many connections and/or configure the limits appropriately, rather than allowing for unlimited overflow which does not actually solve the underlying issue.

What causes an application to use up all the connections that it has available?

  • The application is fielding too many concurrent requests to do work based on the configured value for the pool - This is the most straightforward cause. If you have an application that runs in a thread pool that allows for 30 concurrent threads, with one connection in use per thread, if your pool is not configured to allow at least 30 connections checked out at once, you will get this error once your application receives enough concurrent requests. Solution is to raise the limits on the pool or lower the number of concurrent threads.

  • The application is not returning connections to the pool - This is the next most common reason, which is that the application is making use of the connection pool, but the program is failing to release these connections and is instead leaving them open. The connection pool as well as the ORM Session do have logic such that when the session and/or connection object is garbage collected, it results in the underlying connection resources being released, however this behavior cannot be relied upon to release resources in a timely manner.

    A common reason this can occur is that the application uses ORM sessions and does not call Session.close() upon them one the work involving that session is complete. Solution is to make sure ORM sessions if using the ORM, or engine-bound Connection objects if using Core, are explicitly closed at the end of the work being done, either via the appropriate .close() method, or by using one of the available context managers (e.g. “with:” statement) to properly release the resource.

  • The application is attempting to run long-running transactions - A database transaction is a very expensive resource, and should never be left idle waiting for some event to occur. If an application is waiting for a user to push a button, or a result to come off of a long running job queue, or is holding a persistent connection open to a browser, don’t keep a database transaction open for the whole time. As the application needs to work with the database and interact with an event, open a short-lived transaction at that point and then close it.

  • The application is deadlocking - Also a common cause of this error and more difficult to grasp, if an application is not able to complete its use of a connection either due to an application-side or database-side deadlock, the application can use up all the available connections which then leads to additional requests receiving this error. Reasons for deadlocks include:

    • Using an implicit async system such as gevent or eventlet without properly monkeypatching all socket libraries and drivers, or which has bugs in not fully covering for all monkeypatched driver methods, or less commonly when the async system is being used against CPU-bound workloads and greenlets making use of database resources are simply waiting too long to attend to them. Neither implicit nor explicit async programming frameworks are typically necessary or appropriate for the vast majority of relational database operations; if an application must use an async system for some area of functionality, it’s best that database-oriented business methods run within traditional threads that pass messages to the async part of the application.

    • A database side deadlock, e.g. rows are mutually deadlocked

    • Threading errors, such as mutexes in a mutual deadlock, or calling upon an already locked mutex in the same thread

Keep in mind an alternative to using pooling is to turn off pooling entirely. See the section Switching Pool Implementations for background on this. However, note that when this error message is occurring, it is always due to a bigger problem in the application itself; the pool just helps to reveal the problem sooner.

See also

Connection Pooling

Working with Engines and Connections

Can’t reconnect until invalid transaction is rolled back. Please rollback() fully before proceeding

This error condition refers to the case where a Connection was invalidated, either due to a database disconnect detection or due to an explicit call to Connection.invalidate(), but there is still a transaction present that was initiated either explicitly by the Connection.begin() method, or due to the connection automatically beginning a transaction as occurs in the 2.x series of SQLAlchemy when any SQL statements are emitted. When a connection is invalidated, any Transaction that was in progress is now in an invalid state, and must be explicitly rolled back in order to remove it from the Connection.

DBAPI Errors

The Python database API, or DBAPI, is a specification for database drivers which can be located at Pep-249. This API specifies a set of exception classes that accommodate the full range of failure modes of the database.

SQLAlchemy does not generate these exceptions directly. Instead, they are intercepted from the database driver and wrapped by the SQLAlchemy-provided exception DBAPIError, however the messaging within the exception is generated by the driver, not SQLAlchemy.

InterfaceError

Exception raised for errors that are related to the database interface rather than the database itself.

This error is a DBAPI Error and originates from the database driver (DBAPI), not SQLAlchemy itself.

The InterfaceError is sometimes raised by drivers in the context of the database connection being dropped, or not being able to connect to the database. For tips on how to deal with this, see the section Dealing with Disconnects.

DatabaseError

Exception raised for errors that are related to the database itself, and not the interface or data being passed.

This error is a DBAPI Error and originates from the database driver (DBAPI), not SQLAlchemy itself.

DataError

Exception raised for errors that are due to problems with the processed data like division by zero, numeric value out of range, etc.

This error is a DBAPI Error and originates from the database driver (DBAPI), not SQLAlchemy itself.

OperationalError

Exception raised for errors that are related to the database’s operation and not necessarily under the control of the programmer, e.g. an unexpected disconnect occurs, the data source name is not found, a transaction could not be processed, a memory allocation error occurred during processing, etc.

This error is a DBAPI Error and originates from the database driver (DBAPI), not SQLAlchemy itself.

The OperationalError is the most common (but not the only) error class used by drivers in the context of the database connection being dropped, or not being able to connect to the database. For tips on how to deal with this, see the section Dealing with Disconnects.

IntegrityError

Exception raised when the relational integrity of the database is affected, e.g. a foreign key check fails.

This error is a DBAPI Error and originates from the database driver (DBAPI), not SQLAlchemy itself.

InternalError

Exception raised when the database encounters an internal error, e.g. the cursor is not valid anymore, the transaction is out of sync, etc.

This error is a DBAPI Error and originates from the database driver (DBAPI), not SQLAlchemy itself.

The InternalError is sometimes raised by drivers in the context of the database connection being dropped, or not being able to connect to the database. For tips on how to deal with this, see the section Dealing with Disconnects.

ProgrammingError

Exception raised for programming errors, e.g. table not found or already exists, syntax error in the SQL statement, wrong number of parameters specified, etc.

This error is a DBAPI Error and originates from the database driver (DBAPI), not SQLAlchemy itself.

The ProgrammingError is sometimes raised by drivers in the context of the database connection being dropped, or not being able to connect to the database. For tips on how to deal with this, see the section Dealing with Disconnects.

NotSupportedError

Exception raised in case a method or database API was used which is not supported by the database, e.g. requesting a .rollback() on a connection that does not support transaction or has transactions turned off.

This error is a DBAPI Error and originates from the database driver (DBAPI), not SQLAlchemy itself.

SQL Expression Language

Object will not produce a cache key, Performance Implications

SQLAlchemy as of version 1.4 includes a SQL compilation caching facility which will allow Core and ORM SQL constructs to cache their stringified form, along with other structural information used to fetch results from the statement, allowing the relatively expensive string compilation process to be skipped when another structurally equivalent construct is next used. This system relies upon functionality that is implemented for all SQL constructs, including objects such as Column, select(), and TypeEngine objects, to produce a cache key which fully represents their state to the degree that it affects the SQL compilation process.

If the warnings in question refer to widely used objects such as Column objects, and are shown to be affecting the majority of SQL constructs being emitted (using the estimation techniques described at Estimating Cache Performance Using Logging) such that caching is generally not enabled for an application, this will negatively impact performance and can in some cases effectively produce a performance degradation compared to prior SQLAlchemy versions. The FAQ at Why is my application slow after upgrading to 1.4 and/or 2.x? covers this in additional detail.

Caching disables itself if there’s any doubt

Caching relies on being able to generate a cache key that accurately represents the complete structure of a statement in a consistent fashion. If a particular SQL construct (or type) does not have the appropriate directives in place which allow it to generate a proper cache key, then caching cannot be safely enabled:

  • The cache key must represent the complete structure: If the usage of two separate instances of that construct may result in different SQL being rendered, caching the SQL against the first instance of the element using a cache key that does not capture the distinct differences between the first and second elements will result in incorrect SQL being cached and rendered for the second instance.

  • The cache key must be consistent: If a construct represents state that changes every time, such as a literal value, producing unique SQL for every instance of it, this construct is also not safe to cache, as repeated use of the construct will quickly fill up the statement cache with unique SQL strings that will likely not be used again, defeating the purpose of the cache.

For the above two reasons, SQLAlchemy’s caching system is extremely conservative about deciding to cache the SQL corresponding to an object.

Assertion attributes for caching

The warning is emitted based on the criteria below. For further detail on each, see the section Why is my application slow after upgrading to 1.4 and/or 2.x?.

See also

Estimating Cache Performance Using Logging - background on observing cache behavior and efficiency

Why is my application slow after upgrading to 1.4 and/or 2.x? - in the Frequently Asked Questions section

Compiler StrSQLCompiler can’t render element of type <element type>

This error usually occurs when attempting to stringify a SQL expression construct that includes elements which are not part of the default compilation; in this case, the error will be against the StrSQLCompiler class. In less common cases, it can also occur when the wrong kind of SQL expression is used with a particular type of database backend; in those cases, other kinds of SQL compiler classes will be named, such as SQLCompiler or sqlalchemy.dialects.postgresql.PGCompiler. The guidance below is more specific to the “stringification” use case but describes the general background as well.

Normally, a Core SQL construct or ORM Query object can be stringified directly, such as when we use print():

  1. >>> from sqlalchemy import column
  2. >>> print(column("x") == 5)
  3. x = :x_1

When the above SQL expression is stringified, the StrSQLCompiler compiler class is used, which is a special statement compiler that is invoked when a construct is stringified without any dialect-specific information.

However, there are many constructs that are specific to some particular kind of database dialect, for which the StrSQLCompiler doesn’t know how to turn into a string, such as the PostgreSQL “insert on conflict” construct:

  1. >>> from sqlalchemy.dialects.postgresql import insert
  2. >>> from sqlalchemy import table, column
  3. >>> my_table = table("my_table", column("x"), column("y"))
  4. >>> insert_stmt = insert(my_table).values(x="foo")
  5. >>> insert_stmt = insert_stmt.on_conflict_do_nothing(index_elements=["y"])
  6. >>> print(insert_stmt)
  7. Traceback (most recent call last):
  8. ...
  9. sqlalchemy.exc.UnsupportedCompilationError:
  10. Compiler <sqlalchemy.sql.compiler.StrSQLCompiler object at 0x7f04fc17e320>
  11. can't render element of type
  12. <class 'sqlalchemy.dialects.postgresql.dml.OnConflictDoNothing'>

In order to stringify constructs that are specific to particular backend, the ClauseElement.compile() method must be used, passing either an Engine or a Dialect object which will invoke the correct compiler. Below we use a PostgreSQL dialect:

  1. >>> from sqlalchemy.dialects import postgresql
  2. >>> print(insert_stmt.compile(dialect=postgresql.dialect()))
  3. INSERT INTO my_table (x) VALUES (%(x)s) ON CONFLICT (y) DO NOTHING

For an ORM Query object, the statement can be accessed using the Query.statement accessor:

  1. statement = query.statement
  2. print(statement.compile(dialect=postgresql.dialect()))

See the FAQ link below for additional detail on direct stringification / compilation of SQL elements.

See also

How do I render SQL expressions as strings, possibly with bound parameters inlined?

TypeError: <operator> not supported between instances of ‘ColumnProperty’ and <something>

This often occurs when attempting to use a column_property() or deferred() object in the context of a SQL expression, usually within declarative such as:

  1. class Bar(Base):
  2. __tablename__ = "bar"
  3. id = Column(Integer, primary_key=True)
  4. cprop = deferred(Column(Integer))
  5. __table_args__ = (CheckConstraint(cprop > 5),)

Above, the cprop attribute is used inline before it has been mapped, however this cprop attribute is not a Column, it’s a ColumnProperty, which is an interim object and therefore does not have the full functionality of either the Column object or the InstrumentedAttribute object that will be mapped onto the Bar class once the declarative process is complete.

While the ColumnProperty does have a __clause_element__() method, which allows it to work in some column-oriented contexts, it can’t work in an open-ended comparison context as illustrated above, since it has no Python __eq__() method that would allow it to interpret the comparison to the number “5” as a SQL expression and not a regular Python comparison.

The solution is to access the Column directly using the ColumnProperty.expression attribute:

  1. class Bar(Base):
  2. __tablename__ = "bar"
  3. id = Column(Integer, primary_key=True)
  4. cprop = deferred(Column(Integer))
  5. __table_args__ = (CheckConstraint(cprop.expression > 5),)

A value is required for bind parameter <x> (in parameter group <y>)

This error occurs when a statement makes use of bindparam() either implicitly or explicitly and does not provide a value when the statement is executed:

  1. stmt = select(table.c.column).where(table.c.id == bindparam("my_param"))
  2. result = conn.execute(stmt)

Above, no value has been provided for the parameter “my_param”. The correct approach is to provide a value:

  1. result = conn.execute(stmt, my_param=12)

When the message takes the form “a value is required for bind parameter <x> in parameter group <y>”, the message is referring to the “executemany” style of execution. In this case, the statement is typically an INSERT, UPDATE, or DELETE and a list of parameters is being passed. In this format, the statement may be generated dynamically to include parameter positions for every parameter given in the argument list, where it will use the first set of parameters to determine what these should be.

For example, the statement below is calculated based on the first parameter set to require the parameters, “a”, “b”, and “c” - these names determine the final string format of the statement which will be used for each set of parameters in the list. As the second entry does not contain “b”, this error is generated:

  1. m = MetaData()
  2. t = Table("t", m, Column("a", Integer), Column("b", Integer), Column("c", Integer))
  3. e.execute(
  4. t.insert(),
  5. [
  6. {"a": 1, "b": 2, "c": 3},
  7. {"a": 2, "c": 4},
  8. {"a": 3, "b": 4, "c": 5},
  9. ],
  10. )
  1. sqlalchemy.exc.StatementError: (sqlalchemy.exc.InvalidRequestError)
  2. A value is required for bind parameter 'b', in parameter group 1
  3. [SQL: u'INSERT INTO t (a, b, c) VALUES (?, ?, ?)']
  4. [parameters: [{'a': 1, 'c': 3, 'b': 2}, {'a': 2, 'c': 4}, {'a': 3, 'c': 5, 'b': 4}]]

Since “b” is required, pass it as None so that the INSERT may proceed:

  1. e.execute(
  2. t.insert(),
  3. [
  4. {"a": 1, "b": 2, "c": 3},
  5. {"a": 2, "b": None, "c": 4},
  6. {"a": 3, "b": 4, "c": 5},
  7. ],
  8. )

See also

Sending Parameters

Expected FROM clause, got Select. To create a FROM clause, use the .subquery() method

This refers to a change made as of SQLAlchemy 1.4 where a SELECT statement as generated by a function such as select(), but also including things like unions and textual SELECT expressions are no longer considered to be FromClause objects and can’t be placed directly in the FROM clause of another SELECT statement without them being wrapped in a Subquery first. This is a major conceptual change in the Core and the full rationale is discussed at A SELECT statement is no longer implicitly considered to be a FROM clause.

Given an example as:

  1. m = MetaData()
  2. t = Table("t", m, Column("a", Integer), Column("b", Integer), Column("c", Integer))
  3. stmt = select(t)

Above, stmt represents a SELECT statement. The error is produced when we want to use stmt directly as a FROM clause in another SELECT, such as if we attempted to select from it:

  1. new_stmt_1 = select(stmt)

Or if we wanted to use it in a FROM clause such as in a JOIN:

  1. new_stmt_2 = select(some_table).select_from(some_table.join(stmt))

In previous versions of SQLAlchemy, using a SELECT inside of another SELECT would produce a parenthesized, unnamed subquery. In most cases, this form of SQL is not very useful as databases like MySQL and PostgreSQL require that subqueries in FROM clauses have named aliases, which means using the SelectBase.alias() method or as of 1.4 using the SelectBase.subquery() method to produce this. On other databases, it is still much clearer for the subquery to have a name to resolve any ambiguity on future references to column names inside the subquery.

Beyond the above practical reasons, there are a lot of other SQLAlchemy-oriented reasons the change is being made. The correct form of the above two statements therefore requires that SelectBase.subquery() is used:

  1. subq = stmt.subquery()
  2. new_stmt_1 = select(subq)
  3. new_stmt_2 = select(some_table).select_from(some_table.join(subq))

See also

A SELECT statement is no longer implicitly considered to be a FROM clause

An alias is being generated automatically for raw clauseelement

New in version 1.4.26.

This deprecation warning refers to a very old and likely not well known pattern that applies to the legacy Query.join() method as well as the 2.0 style Select.join() method, where a join can be stated in terms of a relationship() but the target is the Table or other Core selectable to which the class is mapped, rather than an ORM entity such as a mapped class or aliased() construct:

  1. a1 = Address.__table__
  2. q = (
  3. s.query(User)
  4. .join(a1, User.addresses)
  5. .filter(Address.email_address == "ed@foo.com")
  6. .all()
  7. )

The above pattern also allows an arbitrary selectable, such as a Core Join or Alias object, however there is no automatic adaptation of this element, meaning the Core element would need to be referred towards directly:

  1. a1 = Address.__table__.alias()
  2. q = (
  3. s.query(User)
  4. .join(a1, User.addresses)
  5. .filter(a1.c.email_address == "ed@foo.com")
  6. .all()
  7. )

The correct way to specify a join target is always by using the mapped class itself or an aliased object, in the latter case using the PropComparator.of_type() modifier to set up an alias:

  1. # normal join to relationship entity
  2. q = s.query(User).join(User.addresses).filter(Address.email_address == "ed@foo.com")
  3. # name Address target explicitly, not necessary but legal
  4. q = (
  5. s.query(User)
  6. .join(Address, User.addresses)
  7. .filter(Address.email_address == "ed@foo.com")
  8. )

Join to an alias:

  1. from sqlalchemy.orm import aliased
  2. a1 = aliased(Address)
  3. # of_type() form; recommended
  4. q = (
  5. s.query(User)
  6. .join(User.addresses.of_type(a1))
  7. .filter(a1.email_address == "ed@foo.com")
  8. )
  9. # target, onclause form
  10. q = s.query(User).join(a1, User.addresses).filter(a1.email_address == "ed@foo.com")

An alias is being generated automatically due to overlapping tables

New in version 1.4.26.

This warning is typically generated when querying using the Select.join() method or the legacy Query.join() method with mappings that involve joined table inheritance. The issue is that when joining between two joined inheritance models that share a common base table, a proper SQL JOIN between the two entities cannot be formed without applying an alias to one side or the other; SQLAlchemy applies an alias to the right side of the join. For example given a joined inheritance mapping as:

  1. class Employee(Base):
  2. __tablename__ = "employee"
  3. id = Column(Integer, primary_key=True)
  4. manager_id = Column(ForeignKey("manager.id"))
  5. name = Column(String(50))
  6. type = Column(String(50))
  7. reports_to = relationship("Manager", foreign_keys=manager_id)
  8. __mapper_args__ = {
  9. "polymorphic_identity": "employee",
  10. "polymorphic_on": type,
  11. }
  12. class Manager(Employee):
  13. __tablename__ = "manager"
  14. id = Column(Integer, ForeignKey("employee.id"), primary_key=True)
  15. __mapper_args__ = {
  16. "polymorphic_identity": "manager",
  17. "inherit_condition": id == Employee.id,
  18. }

The above mapping includes a relationship between the Employee and Manager classes. Since both classes make use of the “employee” database table, from a SQL perspective this is a self referential relationship. If we wanted to query from both the Employee and Manager models using a join, at the SQL level the “employee” table needs to be included twice in the query, which means it must be aliased. When we create such a join using the SQLAlchemy ORM, we get SQL that looks like the following:

  1. >>> stmt = select(Employee, Manager).join(Employee.reports_to)
  2. >>> print(stmt)
  3. SELECT employee.id, employee.manager_id, employee.name,
  4. employee.type, manager_1.id AS id_1, employee_1.id AS id_2,
  5. employee_1.manager_id AS manager_id_1, employee_1.name AS name_1,
  6. employee_1.type AS type_1
  7. FROM employee JOIN
  8. (employee AS employee_1 JOIN manager AS manager_1 ON manager_1.id = employee_1.id)
  9. ON manager_1.id = employee.manager_id

Above, the SQL selects FROM the employee table, representing the Employee entity in the query. It then joins to a right-nested join of employee AS employee_1 JOIN manager AS manager_1, where the employee table is stated again, except as an anonymous alias employee_1. This is the “automatic generation of an alias” that the warning message refers towards.

When SQLAlchemy loads ORM rows that each contain an Employee and a Manager object, the ORM must adapt rows from what above is the employee_1 and manager_1 table aliases into those of the un-aliased Manager class. This process is internally complex and does not accommodate for all API features, notably when trying to use eager loading features such as contains_eager() with more deeply nested queries than are shown here. As the pattern is unreliable for more complex scenarios and involves implicit decisionmaking that is difficult to anticipate and follow, the warning is emitted and this pattern may be considered a legacy feature. The better way to write this query is to use the same patterns that apply to any other self-referential relationship, which is to use the aliased() construct explicitly. For joined-inheritance and other join-oriented mappings, it is usually desirable to add the use of the aliased.flat parameter, which will allow a JOIN of two or more tables to be aliased by applying an alias to the individual tables within the join, rather than embedding the join into a new subquery:

  1. >>> from sqlalchemy.orm import aliased
  2. >>> manager_alias = aliased(Manager, flat=True)
  3. >>> stmt = select(Employee, manager_alias).join(Employee.reports_to.of_type(manager_alias))
  4. >>> print(stmt)
  5. SELECT employee.id, employee.manager_id, employee.name,
  6. employee.type, manager_1.id AS id_1, employee_1.id AS id_2,
  7. employee_1.manager_id AS manager_id_1, employee_1.name AS name_1,
  8. employee_1.type AS type_1
  9. FROM employee JOIN
  10. (employee AS employee_1 JOIN manager AS manager_1 ON manager_1.id = employee_1.id)
  11. ON manager_1.id = employee.manager_id

If we then wanted to use contains_eager() to populate the reports_to attribute, we refer to the alias:

  1. >>> stmt = (
  2. ... select(Employee)
  3. ... .join(Employee.reports_to.of_type(manager_alias))
  4. ... .options(contains_eager(Employee.reports_to.of_type(manager_alias)))
  5. ... )

Without using the explicit aliased() object, in some more nested cases the contains_eager() option does not have enough context to know where to get its data from, in the case that the ORM is “auto-aliasing” in a very nested context. Therefore it’s best not to rely on this feature and instead keep the SQL construction as explicit as possible.

Object Relational Mapping

Parent instance <x> is not bound to a Session; (lazy load/deferred load/refresh/etc.) operation cannot proceed

This is likely the most common error message when dealing with the ORM, and it occurs as a result of the nature of a technique the ORM makes wide use of known as lazy loading. Lazy loading is a common object-relational pattern whereby an object that’s persisted by the ORM maintains a proxy to the database itself, such that when various attributes upon the object are accessed, their value may be retrieved from the database lazily. The advantage to this approach is that objects can be retrieved from the database without having to load all of their attributes or related data at once, and instead only that data which is requested can be delivered at that time. The major disadvantage is basically a mirror image of the advantage, which is that if lots of objects are being loaded which are known to require a certain set of data in all cases, it is wasteful to load that additional data piecemeal.

Another caveat of lazy loading beyond the usual efficiency concerns is that in order for lazy loading to proceed, the object has to remain associated with a Session in order to be able to retrieve its state. This error message means that an object has become de-associated with its Session and is being asked to lazy load data from the database.

The most common reason that objects become detached from their Session is that the session itself was closed, typically via the Session.close() method. The objects will then live on to be accessed further, very often within web applications where they are delivered to a server-side templating engine and are asked for further attributes which they cannot load.

Mitigation of this error is via these techniques:

  • Try not to have detached objects; don’t close the session prematurely - Often, applications will close out a transaction before passing off related objects to some other system which then fails due to this error. Sometimes the transaction doesn’t need to be closed so soon; an example is the web application closes out the transaction before the view is rendered. This is often done in the name of “correctness”, but may be seen as a mis-application of “encapsulation”, as this term refers to code organization, not actual actions. The template that uses an ORM object is making use of the proxy pattern which keeps database logic encapsulated from the caller. If the Session can be held open until the lifespan of the objects are done, this is the best approach.

  • Otherwise, load everything that’s needed up front - It is very often impossible to keep the transaction open, especially in more complex applications that need to pass objects off to other systems that can’t run in the same context even though they’re in the same process. In this case, the application should prepare to deal with detached objects, and should try to make appropriate use of eager loading to ensure that objects have what they need up front.

  • And importantly, set expire_on_commit to False - When using detached objects, the most common reason objects need to re-load data is because they were expired from the last call to Session.commit(). This expiration should not be used when dealing with detached objects; so the Session.expire_on_commit parameter be set to False. By preventing the objects from becoming expired outside of the transaction, the data which was loaded will remain present and will not incur additional lazy loads when that data is accessed.

    Note also that Session.rollback() method unconditionally expires all contents in the Session and should also be avoided in non-error scenarios.

    See also

    Relationship Loading Techniques - detailed documentation on eager loading and other relationship-oriented loading techniques

    Committing - background on session commit

    Refreshing / Expiring - background on attribute expiry

This Session’s transaction has been rolled back due to a previous exception during flush

The flush process of the Session, described at Flushing, will roll back the database transaction if an error is encountered, in order to maintain internal consistency. However, once this occurs, the session’s transaction is now “inactive” and must be explicitly rolled back by the calling application, in the same way that it would otherwise need to be explicitly committed if a failure had not occurred.

This is a common error when using the ORM and typically applies to an application that doesn’t yet have correct “framing” around its Session operations. Further detail is described in the FAQ at “This Session’s transaction has been rolled back due to a previous exception during flush.” (or similar).

For relationship <relationship>, delete-orphan cascade is normally configured only on the “one” side of a one-to-many relationship, and not on the “many” side of a many-to-one or many-to-many relationship.

This error arises when the “delete-orphan” cascade is set on a many-to-one or many-to-many relationship, such as:

  1. class A(Base):
  2. __tablename__ = "a"
  3. id = Column(Integer, primary_key=True)
  4. bs = relationship("B", back_populates="a")
  5. class B(Base):
  6. __tablename__ = "b"
  7. id = Column(Integer, primary_key=True)
  8. a_id = Column(ForeignKey("a.id"))
  9. # this will emit the error message when the mapper
  10. # configuration step occurs
  11. a = relationship("A", back_populates="bs", cascade="all, delete-orphan")
  12. configure_mappers()

Above, the “delete-orphan” setting on B.a indicates the intent that when every B object that refers to a particular A is deleted, that the A should then be deleted as well. That is, it expresses that the “orphan” which is being deleted would be an A object, and it becomes an “orphan” when every B that refers to it is deleted.

The “delete-orphan” cascade model does not support this functionality. The “orphan” consideration is only made in terms of the deletion of a single object which would then refer to zero or more objects that are now “orphaned” by this single deletion, which would result in those objects being deleted as well. In other words, it is designed only to track the creation of “orphans” based on the removal of one and only one “parent” object per orphan, which is the natural case in a one-to-many relationship where a deletion of the object on the “one” side results in the subsequent deletion of the related items on the “many” side.

The above mapping in support of this functionality would instead place the cascade setting on the one-to-many side, which looks like:

  1. class A(Base):
  2. __tablename__ = "a"
  3. id = Column(Integer, primary_key=True)
  4. bs = relationship("B", back_populates="a", cascade="all, delete-orphan")
  5. class B(Base):
  6. __tablename__ = "b"
  7. id = Column(Integer, primary_key=True)
  8. a_id = Column(ForeignKey("a.id"))
  9. a = relationship("A", back_populates="bs")

Where the intent is expressed that when an A is deleted, all of the B objects to which it refers are also deleted.

The error message then goes on to suggest the usage of the relationship.single_parent flag. This flag may be used to enforce that a relationship which is capable of having many objects refer to a particular object will in fact have only one object referring to it at a time. It is used for legacy or other less ideal database schemas where the foreign key relationships suggest a “many” collection, however in practice only one object would actually refer to a given target object at at time. This uncommon scenario can be demonstrated in terms of the above example as follows:

  1. class A(Base):
  2. __tablename__ = "a"
  3. id = Column(Integer, primary_key=True)
  4. bs = relationship("B", back_populates="a")
  5. class B(Base):
  6. __tablename__ = "b"
  7. id = Column(Integer, primary_key=True)
  8. a_id = Column(ForeignKey("a.id"))
  9. a = relationship(
  10. "A",
  11. back_populates="bs",
  12. single_parent=True,
  13. cascade="all, delete-orphan",
  14. )

The above configuration will then install a validator which will enforce that only one B may be associated with an A at at time, within the scope of the B.a relationship:

  1. >>> b1 = B()
  2. >>> b2 = B()
  3. >>> a1 = A()
  4. >>> b1.a = a1
  5. >>> b2.a = a1
  6. sqlalchemy.exc.InvalidRequestError: Instance <A at 0x7eff44359350> is
  7. already associated with an instance of <class '__main__.B'> via its
  8. B.a attribute, and is only allowed a single parent.

Note that this validator is of limited scope and will not prevent multiple “parents” from being created via the other direction. For example, it will not detect the same setting in terms of A.bs:

  1. >>> a1.bs = [b1, b2]
  2. >>> session.add_all([a1, b1, b2])
  3. >>> session.commit()
  4. INSERT INTO a DEFAULT VALUES
  5. ()
  6. INSERT INTO b (a_id) VALUES (?)
  7. (1,)
  8. INSERT INTO b (a_id) VALUES (?)
  9. (1,)

However, things will not go as expected later on, as the “delete-orphan” cascade will continue to work in terms of a single lead object, meaning if we delete either of the B objects, the A is deleted. The other B stays around, where the ORM will usually be smart enough to set the foreign key attribute to NULL, but this is usually not what’s desired:

  1. >>> session.delete(b1)
  2. >>> session.commit()
  3. UPDATE b SET a_id=? WHERE b.id = ?
  4. (None, 2)
  5. DELETE FROM b WHERE b.id = ?
  6. (1,)
  7. DELETE FROM a WHERE a.id = ?
  8. (1,)
  9. COMMIT

For all the above examples, similar logic applies to the calculus of a many-to-many relationship; if a many-to-many relationship sets single_parent=True on one side, that side can use the “delete-orphan” cascade, however this is very unlikely to be what someone actually wants as the point of a many-to-many relationship is so that there can be many objects referring to an object in either direction.

Overall, “delete-orphan” cascade is usually applied on the “one” side of a one-to-many relationship so that it deletes objects in the “many” side, and not the other way around.

Changed in version 1.3.18: The text of the “delete-orphan” error message when used on a many-to-one or many-to-many relationship has been updated to be more descriptive.

See also

Cascades

delete-orphan

Instance <instance> is already associated with an instance of <instance> via its <attribute> attribute, and is only allowed a single parent.

Instance <instance> is already associated with an instance of <instance> via its <attribute> attribute, and is only allowed a single parent.

This error is emitted when the relationship.single_parent flag is used, and more than one object is assigned as the “parent” of an object at once.

Given the following mapping:

  1. class A(Base):
  2. __tablename__ = "a"
  3. id = Column(Integer, primary_key=True)
  4. class B(Base):
  5. __tablename__ = "b"
  6. id = Column(Integer, primary_key=True)
  7. a_id = Column(ForeignKey("a.id"))
  8. a = relationship(
  9. "A",
  10. single_parent=True,
  11. cascade="all, delete-orphan",
  12. )

The intent indicates that no more than a single B object may refer to a particular A object at once:

  1. >>> b1 = B()
  2. >>> b2 = B()
  3. >>> a1 = A()
  4. >>> b1.a = a1
  5. >>> b2.a = a1
  6. sqlalchemy.exc.InvalidRequestError: Instance <A at 0x7eff44359350> is
  7. already associated with an instance of <class '__main__.B'> via its
  8. B.a attribute, and is only allowed a single parent.

When this error occurs unexpectedly, it is usually because the relationship.single_parent flag was applied in response to the error message described at For relationship <relationship>, delete-orphan cascade is normally configured only on the “one” side of a one-to-many relationship, and not on the “many” side of a many-to-one or many-to-many relationship., and the issue is in fact a misunderstanding of the “delete-orphan” cascade setting. See that message for details.

See also

For relationship <relationship>, delete-orphan cascade is normally configured only on the “one” side of a one-to-many relationship, and not on the “many” side of a many-to-one or many-to-many relationship.

relationship X will copy column Q to column P, which conflicts with relationship(s): ‘Y’

This warning refers to the case when two or more relationships will write data to the same columns on flush, but the ORM does not have any means of coordinating these relationships together. Depending on specifics, the solution may be that two relationships need to be referred towards one another using relationship.back_populates, or that one or more of the relationships should be configured with relationship.viewonly to prevent conflicting writes, or sometimes that the configuration is fully intentional and should configure relationship.overlaps to silence each warning.

For the typical example that’s missing relationship.back_populates, given the following mapping:

  1. class Parent(Base):
  2. __tablename__ = "parent"
  3. id = Column(Integer, primary_key=True)
  4. children = relationship("Child")
  5. class Child(Base):
  6. __tablename__ = "child"
  7. id = Column(Integer, primary_key=True)
  8. parent_id = Column(ForeignKey("parent.id"))
  9. parent = relationship("Parent")

The above mapping will generate warnings:

  1. SAWarning: relationship 'Child.parent' will copy column parent.id to column child.parent_id,
  2. which conflicts with relationship(s): 'Parent.children' (copies parent.id to child.parent_id).

The relationships Child.parent and Parent.children appear to be in conflict. The solution is to apply relationship.back_populates:

  1. class Parent(Base):
  2. __tablename__ = "parent"
  3. id = Column(Integer, primary_key=True)
  4. children = relationship("Child", back_populates="parent")
  5. class Child(Base):
  6. __tablename__ = "child"
  7. id = Column(Integer, primary_key=True)
  8. parent_id = Column(ForeignKey("parent.id"))
  9. parent = relationship("Parent", back_populates="children")

For more customized relationships where an “overlap” situation may be intentional and cannot be resolved, the relationship.overlaps parameter may specify the names of relationships for which the warning should not take effect. This typically occurs for two or more relationships to the same underlying table that include custom relationship.primaryjoin conditions that limit the related items in each case:

  1. class Parent(Base):
  2. __tablename__ = "parent"
  3. id = Column(Integer, primary_key=True)
  4. c1 = relationship(
  5. "Child",
  6. primaryjoin="and_(Parent.id == Child.parent_id, Child.flag == 0)",
  7. backref="parent",
  8. overlaps="c2, parent",
  9. )
  10. c2 = relationship(
  11. "Child",
  12. primaryjoin="and_(Parent.id == Child.parent_id, Child.flag == 1)",
  13. overlaps="c1, parent",
  14. )
  15. class Child(Base):
  16. __tablename__ = "child"
  17. id = Column(Integer, primary_key=True)
  18. parent_id = Column(ForeignKey("parent.id"))
  19. flag = Column(Integer)

Above, the ORM will know that the overlap between Parent.c1, Parent.c2 and Child.parent is intentional.

Object cannot be converted to ‘persistent’ state, as this identity map is no longer valid.

New in version 1.4.26.

This message was added to accommodate for the case where a Result object that would yield ORM objects is iterated after the originating Session has been closed, or otherwise had its Session.expunge_all() method called. When a Session expunges all objects at once, the internal identity map used by that Session is replaced with a new one, and the original one discarded. An unconsumed and unbuffered Result object will internally maintain a reference to that now-discarded identity map. Therefore, when the Result is consumed, the objects that would be yielded cannot be associated with that Session. This arrangement is by design as it is generally not recommended to iterate an unbuffered Result object outside of the transactional context in which it was created:

  1. # context manager creates new Session
  2. with Session(engine) as session_obj:
  3. result = sess.execute(select(User).where(User.id == 7))
  4. # context manager is closed, so session_obj above is closed, identity
  5. # map is replaced
  6. # iterating the result object can't associate the object with the
  7. # Session, raises this error.
  8. user = result.first()

The above situation typically will not occur when using the asyncio ORM extension, as when AsyncSession returns a sync-style Result, the results have been pre-buffered when the statement was executed. This is to allow secondary eager loaders to invoke without needing an additional await call.

To pre-buffer results in the above situation using the regular Session in the same way that the asyncio extension does it, the prebuffer_rows execution option may be used as follows:

  1. # context manager creates new Session
  2. with Session(engine) as session_obj:
  3. # result internally pre-fetches all objects
  4. result = sess.execute(
  5. select(User).where(User.id == 7), execution_options={"prebuffer_rows": True}
  6. )
  7. # context manager is closed, so session_obj above is closed, identity
  8. # map is replaced
  9. # pre-buffered objects are returned
  10. user = result.first()
  11. # however they are detached from the session, which has been closed
  12. assert inspect(user).detached
  13. assert inspect(user).session is None

Above, the selected ORM objects are fully generated within the session_obj block, associated with session_obj and buffered within the Result object for iteration. Outside the block, session_obj is closed and expunges these ORM objects. Iterating the Result object will yield those ORM objects, however as their originating Session has expunged them, they will be delivered in the detached state.

Note

The above reference to a “pre-buffered” vs. “un-buffered” Result object refers to the process by which the ORM converts incoming raw database rows from the DBAPI into ORM objects. It does not imply whether or not the underlying cursor object itself, which represents pending results from the DBAPI, is itself buffered or unbuffered, as this is essentially a lower layer of buffering. For background on buffering of the cursor results itself, see the section Using Server Side Cursors (a.k.a. stream results).

Type annotation can’t be interpreted for Annotated Declarative Table form

SQLAlchemy 2.0 introduces a new Annotated Declarative Table declarative system which derives ORM mapped attribute information from PEP 484 annotations within class definitions at runtime. A requirement of this form is that all ORM annotations must make use of a generic container called Mapped to be properly annotated. Legacy SQLAlchemy mappings which include explicit PEP 484 typing annotations, such as those which use the legacy Mypy extension for typing support, may include directives such as those for relationship() that don’t include this generic.

To resolve, the classes may be marked with the __allow_unmapped__ boolean attribute until they can be fully migrated to the 2.0 syntax. See the migration notes at Migration to 2.0 Step Six - Add __allow_unmapped__ to explicitly typed ORM models for an example.

See also

Migration to 2.0 Step Six - Add __allow_unmapped__ to explicitly typed ORM models - in the SQLAlchemy 2.0 - Major Migration Guide document

AsyncIO Exceptions

AwaitRequired

The SQLAlchemy async mode requires an async driver to be used to connect to the db. This error is usually raised when trying to use the async version of SQLAlchemy with a non compatible DBAPI.

See also

Asynchronous I/O (asyncio)

MissingGreenlet

A call to the async DBAPI was initiated outside the greenlet spawn context usually setup by the SQLAlchemy AsyncIO proxy classes. Usually this error happens when an IO was attempted in an unexpected place, without using the provided async api. When using the ORM this may be due to a lazy loading attempt, which is unsupported when using SQLAlchemy with AsyncIO dialects.

See also

Preventing Implicit IO when Using AsyncSession - covers most ORM scenarios where this problem can occur and how to mitigate.

No Inspection Available

Using the inspect() function directly on an AsyncConnection or AsyncEngine object is not currently supported, as there is not yet an awaitable form of the Inspector object available. Instead, the object is used by acquiring it using the inspect() function in such a way that it refers to the underlying AsyncConnection.sync_connection attribute of the AsyncConnection object; the Inspector is then used in a “synchronous” calling style by using the AsyncConnection.run_sync() method along with a custom function that performs the desired operations:

  1. async def async_main():
  2. async with engine.connect() as conn:
  3. tables = await conn.run_sync(
  4. lambda sync_conn: inspect(sync_conn).get_table_names()
  5. )

See also

Using the Inspector to inspect schema objects - additional examples of using inspect() with the asyncio extension.

Core Exception Classes

See Core Exceptions for Core exception classes.

ORM Exception Classes

See ORM Exceptions for ORM exception classes.

Legacy Exceptions

Exceptions in this section are not generated by current SQLAlchemy versions, however are provided here to suit exception message hyperlinks.

The <some function> in SQLAlchemy 2.0 will no longer <something>

SQLAlchemy 2.0 represents a major shift for a wide variety of key SQLAlchemy usage patterns in both the Core and ORM components. The goal of the 2.0 release is to make a slight readjustment in some of the most fundamental assumptions of SQLAlchemy since its early beginnings, and to deliver a newly streamlined usage model that is hoped to be significantly more minimalist and consistent between the Core and ORM components, as well as more capable.

Introduced at SQLAlchemy 2.0 - Major Migration Guide, the SQLAlchemy 2.0 project includes a comprehensive future compatibility system that’s integrated into the 1.4 series of SQLAlchemy, such that applications will have a clear, unambiguous, and incremental upgrade path in order to migrate applications to being fully 2.0 compatible. The RemovedIn20Warning deprecation warning is at the base of this system to provide guidance on what behaviors in an existing codebase will need to be modified. An overview of how to enable this warning is at SQLAlchemy 2.0 Deprecations Mode.

See also

SQLAlchemy 2.0 - Major Migration Guide - An overview of the upgrade process from the 1.x series, as well as the current goals and progress of SQLAlchemy 2.0.

SQLAlchemy 2.0 Deprecations Mode - specific guidelines on how to use “2.0 deprecations mode” in SQLAlchemy 1.4.

Object is being merged into a Session along the backref cascade

This message refers to the “backref cascade” behavior of SQLAlchemy, removed in version 2.0. This refers to the action of an object being added into a Session as a result of another object that’s already present in that session being associated with it. As this behavior has been shown to be more confusing than helpful, the relationship.cascade_backrefs and backref.cascade_backrefs parameters were added, which can be set to False to disable it, and in SQLAlchemy 2.0 the “cascade backrefs” behavior has been removed entirely.

For older SQLAlchemy versions, to set relationship.cascade_backrefs to False on a backref that is currently configured using the relationship.backref string parameter, the backref must be declared using the backref() function first so that the backref.cascade_backrefs parameter may be passed.

Alternatively, the entire “cascade backrefs” behavior can be turned off across the board by using the Session in “future” mode, by passing True for the Session.future parameter.

See also

cascade_backrefs behavior deprecated for removal in 2.0 - background on the change for SQLAlchemy 2.0.

select() construct created in “legacy” mode; keyword arguments, etc.

The select() construct has been updated as of SQLAlchemy 1.4 to support the newer calling style that is standard in SQLAlchemy 2.0. For backwards compatibility within the 1.4 series, the construct accepts arguments in both the “legacy” style as well as the “new” style.

The “new” style features that column and table expressions are passed positionally to the select() construct only; any other modifiers to the object must be passed using subsequent method chaining:

  1. # this is the way to do it going forward
  2. stmt = select(table1.c.myid).where(table1.c.myid == table2.c.otherid)

For comparison, a select() in legacy forms of SQLAlchemy, before methods like Select.where() were even added, would like:

  1. # this is how it was documented in original SQLAlchemy versions
  2. # many years ago
  3. stmt = select([table1.c.myid], whereclause=table1.c.myid == table2.c.otherid)

Or even that the “whereclause” would be passed positionally:

  1. # this is also how it was documented in original SQLAlchemy versions
  2. # many years ago
  3. stmt = select([table1.c.myid], table1.c.myid == table2.c.otherid)

For some years now, the additional “whereclause” and other arguments that are accepted have been removed from most narrative documentation, leading to a calling style that is most familiar as the list of column arguments passed as a list, but no further arguments:

  1. # this is how it's been documented since around version 1.0 or so
  2. stmt = select([table1.c.myid]).where(table1.c.myid == table2.c.otherid)

The document at select() no longer accepts varied constructor arguments, columns are passed positionally describes this change in terms of 2.0 Migration.

See also

select() no longer accepts varied constructor arguments, columns are passed positionally

SQLAlchemy 2.0 - Major Migration Guide

A bind was located via legacy bound metadata, but since future=True is set on this Session, this bind is ignored.

The concept of “bound metadata” is present up until SQLAlchemy 1.4; as of SQLAlchemy 2.0 it’s been removed.

This error refers to the MetaData.bind parameter on the MetaData object that in turn allows objects like the ORM Session to associate a particular mapped class with an Engine. In SQLAlchemy 2.0, the Session must be linked to each Engine directly. That is, instead of instantiating the Session or sessionmaker without any arguments, and associating the Engine with the MetaData:

  1. engine = create_engine("sqlite://")
  2. Session = sessionmaker()
  3. metadata_obj = MetaData(bind=engine)
  4. Base = declarative_base(metadata=metadata_obj)
  5. class MyClass(Base):
  6. ...
  7. session = Session()
  8. session.add(MyClass())
  9. session.commit()

The Engine must instead be associated directly with the sessionmaker or Session. The MetaData object should no longer be associated with any engine:

  1. engine = create_engine("sqlite://")
  2. Session = sessionmaker(engine)
  3. Base = declarative_base()
  4. class MyClass(Base):
  5. ...
  6. session = Session()
  7. session.add(MyClass())
  8. session.commit()

In SQLAlchemy 1.4, this 2.0 style behavior is enabled when the Session.future flag is set on sessionmaker or Session.

This Compiled object is not bound to any Engine or Connection

This error refers to the concept of “bound metadata”, which is a legacy SQLAlchemy pattern present only in 1.x versions. The issue occurs when one invokes the Executable.execute() method directly off of a Core expression object that is not associated with any Engine:

  1. metadata_obj = MetaData()
  2. table = Table("t", metadata_obj, Column("q", Integer))
  3. stmt = select(table)
  4. result = stmt.execute() # <--- raises

What the logic is expecting is that the MetaData object has been bound to a Engine:

  1. engine = create_engine("mysql+pymysql://user:pass@host/db")
  2. metadata_obj = MetaData(bind=engine)

Where above, any statement that derives from a Table which in turn derives from that MetaData will implicitly make use of the given Engine in order to invoke the statement.

Note that the concept of bound metadata is not present in SQLAlchemy 2.0. The correct way to invoke statements is via the Connection.execute() method of a Connection:

  1. with engine.connect() as conn:
  2. result = conn.execute(stmt)

When using the ORM, a similar facility is available via the Session:

  1. result = session.execute(stmt)

See also

Basics of Statement Execution

This connection is on an inactive transaction. Please rollback() fully before proceeding

This error condition was added to SQLAlchemy as of version 1.4, and does not apply to SQLAlchemy 2.0. The error refers to the state where a Connection is placed into a transaction using a method like Connection.begin(), and then a further “marker” transaction is created within that scope; the “marker” transaction is then rolled back using Transaction.rollback() or closed using Transaction.close(), however the outer transaction is still present in an “inactive” state and must be rolled back.

The pattern looks like:

  1. engine = create_engine(...)
  2. connection = engine.connect()
  3. transaction1 = connection.begin()
  4. # this is a "sub" or "marker" transaction, a logical nesting
  5. # structure based on "real" transaction transaction1
  6. transaction2 = connection.begin()
  7. transaction2.rollback()
  8. # transaction1 is still present and needs explicit rollback,
  9. # so this will raise
  10. connection.execute(text("select 1"))

Above, transaction2 is a “marker” transaction, which indicates a logical nesting of transactions within an outer one; while the inner transaction can roll back the whole transaction via its rollback() method, its commit() method has no effect except to close the scope of the “marker” transaction itself. The call to transaction2.rollback() has the effect of deactivating transaction1 which means it is essentially rolled back at the database level, however is still present in order to accommodate a consistent nesting pattern of transactions.

The correct resolution is to ensure the outer transaction is also rolled back:

  1. transaction1.rollback()

This pattern is not commonly used in Core. Within the ORM, a similar issue can occur which is the product of the ORM’s “logical” transaction structure; this is described in the FAQ entry at “This Session’s transaction has been rolled back due to a previous exception during flush.” (or similar).

The “subtransaction” pattern is removed in SQLAlchemy 2.0 so that this particular programming pattern is no longer be available, preventing this error message.