Current state: Accepted
Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).
Motivations and Background
In Flink 1.9, we integrated Flink’s FunctionCatalog with CatalogManager from the brand-new catalog API. It enabled Flink to resolve function reference to catalog functions and to support using rich Hive user defined functions (of types UDF, GenericUDF, GenericUDTF, UDAF, GenericUDAFResolver) in Flink SQL. The feature has been a critical integration for Flink to embrace the rich ecosystem of Hive, and attract more users to Flink batch.
Right now there are mainly 3 types of functions in Flink:
- Temporary function: These are functions that are registered via FunctionCatalog#registerScalarFunction() / registerTableFunction() / registeraAggregateFunction(), or created via upcoming function DDL as “CREATE TEMP FUNCTION”. They are held in-memory in FunctionCatalog, last only for the lifespan of a session
- Built-in functions:These are system functions in BuiltInFunctionDefinitions
- Catalog functions in the current catalog and current database of the session
Despite that big advancement, there are still a few problems existing in how Flink table handles functions.
1. Does not clarify the concept of temporary functions
By definition, functions that are created as temporary functions will only be visible to the current session, and will only last till the end of the session.
FunctionCatalog right now holds a map of FunctionDefinitions named ‘userFunctions’ in its memory, which are registered from Table APIs (FunctionCatalog#registerScalarFunction()/registerTableFunction()/registerAggregateFunction()). These are actually temporary functions, though we haven’t officially categorize them as temporary functions explicitly in both the variable name and the API names yet. Not making the nature of these functions clear confuses users, it also doesn’t distinguish them from catalog functions well.
2. Does not clarify the concept of built-in functions
Built-in functions are system information. Similar to temp functions, they don’t have namespaces and should not belong to any catalog or databases.
However, in Flink 1.9, built-in functions are regarded as belongings of Flink’s default catalog and default database. Note that Flink’s default catalog (a.k.a Flink’s “built-in” catalog) is not a system catalog, it’s backed by a GenericInMemoryCatalog impl for the main purpose of backward compatbility and it is no different from any persistent catalog by design except its storage is memory. We should be extra careful on deciding what belongs to the default catalog, instead of throwing everything into it.
3. Flink doesn’t support referencing functions with fully/partialy qualified name
There should be two ways to reference functions in SQL:
- Via ambiguous identifier reference, a.k.a. only with function name
E.g. “select * from myfunc(a) from mytable;” in Flink
- Via unambiguous/precise identifier reference, a.k.a, with a completely/partially qualified name, as ‘<catalog_name>.<db_name>.<function_name>’ or ‘<db_name>.<function_name>’.
E.g. “select * from mycatalog.mydb.myfunc(a) from mytable;” or “select * from mydb.myfunc(a) from mytable;” in Flink
Flink currently only support ambiguous identifier reference. The limitation means that if users want to use a function in a non-current catalog and/or db, they have to set that catalog and db as current catalog and current db.
4. Function resolve order
When there are functions of different types but of the same name, resolving the function name happens in the current order:
- Catalog function - functions from catalog APIs
- userFunctions - functions held in-memory in FunctionCatalog
- Flink built-in functions
This order is problematic in a few ways:
- It’s not deterministic. Depending on the catalog and/or user function names registered, the resolution order for ambugious function reference can be vary. E.g. in a catalog/db where a catalog function registered the same name as a built-in function, the catalog function will be called; in a catalog/db where it doesn’t, the built-in function will be called
- When a catalog or user function has the same name as a built-in function, there’s no way for users to reference the built-in function unless they switch to a catalog and db that doesn’t have such a catalog/user function name
- It doesn’t consider temp functions due to lack of that clear concept
We’d like to redefine that order.
1) Support precise functions reference
Add support for umambiguous function reference, with a completely/partially qualified name, as ‘<catalog_name>.<db_name>.<function_name>’ or ‘<db_name>.<function_name>’.
E.g. “select * from mycatalog.mydb.myfunc(a) from mytable;” or “select * from mydb.myfunc(a) from mytable;” in Flink
With this FLIP, users will be empowered to reference user defined catalog functions precisely and also across catalogs and databases in SQL thru the whole stack.
2) Remove namespaces from Flink built-in functions
We should remove namepaces of Flink built-in functions thru the stack of Flink SQL. Flink built-in functions do not belong to any catalog or database.
3) Redefine function resolution order for ambiguous function reference
We will redefine the ambiguos function resolution order, taking all new types of functions into consideration and making it as user-friendly, flexible, and deterministic as possible for users.
4) Support temporary functions explicitly
This FLIP would add explicit temporary function support by renaming a few variable names, and potential renaming some APIs in favor of new APIs that reflects their nature of dealing with temporary functions.
The following work is out of scope of this FLIP. They will taken care of by other teams/committers.
- Corresponding DDL. E.g. “CREATE [TEMP] FUNCTION xxx AS xxx”
- Some corresponding changes in SQL planner/parser (we’ll do what we can do :)
- Reinitiate Flink functions after reading it class name from catalog
However, they need to be done simultaneously to achieve the functionality in full stack.
We will first clarify the dimensions of functions.
|System (can be used interchangeably with "builtin")||Catalog|
|Non-Temporary||system functions||catalog functions|
|Temporary||temporary system functions||temporary catalog functions|
1. Support Two Types of Temporary Functions
We aim at supporting two types of temporary functions:
- temporary system function that has no namespace and overrides built-in functions
- temporary catalog function that has catalog and database namespaces and overrides catalog functions.
FunctionCatalog’s APIs need to change accordingly to accommodate registering temporary functions with/without namespaces optionally.
a) Temporary System Functions
We will have new APIS in FunctionCatalog as “registerTemporarySystemScalar/Table/AggregateFunctions(String, UserDefinedFunction)” and functions will be stored in “Map<String, FunctionDefinition> tempSystemFunctions” in FunctionCatalog.
Their DDLs are “CREATE/DROP TEMPORARY SYSTEM FUNCTION”.
They will be renamed from “registerScalar/Table/AggregateFunctions()”.
b) Temporary Catalog Functions
We will add a new member variable to FunctionCatalog as “Map<ObjectIdentifier, UserDefinedFunction> tempCatalogFunctions“ to hold those temporary catalog functions in a central place, and new APIs “registerTemporaryScalar/Table/AggregateFunction(ObjectIdentifier, UserDefinedFunction)”.
Lifespan of temp functions are not tied to those of catalogs and databases. Users can create temp catalog functions even though catalogs/dbs in their fully qualified names don't even exist.
Their DDLs are “CREATE/DROP TEMPORARY FUNCTION”.
Some other proposed SQL commands are:
"SHOW FUNCTIONS" - list names of temp and non-temp system/built-in functions, and names of temp and catalog functions in the current catalog and db
"SHOW ALL FUNCTIONS" - list names of temp and non-temp system/built functions, and fully qualified names of temp catalog functions and catalog functions in all catalogs and dbs
"SHOW ALL TEMPORARY FUNCTIONS" - list fully qualified names of temp catalog functions in all catalog and db
"SHOW ALL TEMPORARY SYSTEM FUNCTIONS" - list names of all temp system functions
Lifespan of both types of temporary functions will be within a session, and will destroyed upon session end.
Note: corresponding DDL and SQL commands are not part of this FLIP
2. Support Precise Function Reference
Because system functions don’t have namespaces, a precise function reference in Flink must be either temporary catalog functions or catalog functions.
The resolution order will be
- Temporary catalog functions
- Catalog functions
3. Support Ambiguous Function Reference with a Redefined Resolution Order
For ambiguous function reference, there are 4 types of functions to consider: temporary functions with and without no namespaces, Flink system functions, and catalog functions.
The resolution order for ambiguous function reference only matters when there are functions of different types but the same name, e.g. when there’re three functions all named “myfunc” in temp, catalog, built-in functions. If there’s no function name collision, functions will just be resolved to whichever type it belongs to.
- Temporary system functions
- System functions
- Temporary catalog functions, in the current catalog and current database of the session
- Catalog functions, in the current catalog and current database of the session
The new resolution order will be a breaking change, compared to existing resolution order.
Temp functions should rank above their corresponding persistent/built-in functions due to its temporary nature - users want to overwrite built-in or persistent functions with something temporary that is only visible to themselves and the session, and not impacting other users. In contrary, 1) if users don’t have the intention of overwriting other functions, they can just name the temporary functions to something else, considering the manipulation cost is so low for temporary objects, and 2) if built-in functions precede temporary functions, there’s no way to reference temp functions anymore
System functions should precede catalog functions, because 1) it always give a deterministic resolution order on ambiguous reference by invoking the built-in functions 2) catalog functions can always be precisely referenced with fully/partially qualified names. In contrary, if catalog functions precede built-in functions, built-in functions can never be referenced.
4. Code Changes
Changes to CallExpression and UnresolvedCallExpression
We should replace ObjectIdentifier with FunctionIdentifier in CallExpression and UnresolvedCallExpression. This would be a breaking change for these public evolving classes.
Changes to FunctionLookup interface
In case of a partially qualified name (<db_name>.<function_name>) in SQL, planner/parser should recognize it with context be padding name of the current catalog of the session first to build an ObjectIdentifier.
Changes to FunctionCatalog class
We will not consider the following cases in this FLIP, and will leave them to future efforts:
- By SQL standard, function resolution takes both function name and method parameters into consideration. To simplify our design, we will not consider resolving overloaded functions - those with the same name but different params - in this FLIP.
- it's still valid to have a single function with overloaded eval() methods
Some databases products disallow registering functions with the same name as that of a built-in function
- MySQL (https://dev.mysql.com/doc/refman/8.0/en/function-resolution.html)
- Postgres (https://www.postgresql.org/docs/9.5/sql-createfunction.html), can be the same name but have to be with different args or arg types)