flyrain commented on code in PR #14117: URL: https://github.com/apache/iceberg/pull/14117#discussion_r2761879852
########## format/udf-spec.md: ########## @@ -0,0 +1,349 @@ +--- +title: "SQL UDF Spec" +--- +<!-- + - Licensed to the Apache Software Foundation (ASF) under one or more + - contributor license agreements. See the NOTICE file distributed with + - this work for additional information regarding copyright ownership. + - The ASF licenses this file to You under the Apache License, Version 2.0 + - (the "License"); you may not use this file except in compliance with + - the License. You may obtain a copy of the License at + - + - http://www.apache.org/licenses/LICENSE-2.0 + - + - Unless required by applicable law or agreed to in writing, software + - distributed under the License is distributed on an "AS IS" BASIS, + - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + - See the License for the specific language governing permissions and + - limitations under the License. + --> + +# Iceberg UDF Spec + +## Background and Motivation + +A SQL user-defined function (UDF or UDTF) is a callable routine that accepts input parameters and executes a function body. +Depending on the function type, the result can be: + +- **Scalar function (UDF)** – returns a single value, which may be a primitive type (e.g., `int`, `string`) or a non-primitive type (e.g., `struct`, `list`). +- **Table function (UDTF)** – returns a table with zero or more rows of columns with a uniform schema. + +This specification introduces a standardized metadata format for UDFs in Iceberg. + +## Goals + +* Define a portable metadata format for both scalar and table SQL UDFs. The metadata is self-contained and can be moved across catalogs. +* Support function evolution through versioning and rollback. +* Provide consistent semantics for representing UDFs across engines. + +## Overview + +UDF metadata follows the same design principles as Iceberg table and view metadata: each function is represented by a +**self-contained metadata file**. Metadata captures definitions, parameters, return types, documentation, security, +properties, and engine-specific representations. + +* UDF metadata files are immutable. Any modification (new definition, updated representation, changed properties, etc.) + creates a new metadata file. Catalogs use an atomic swap to update the metadata file linked to a catalog identifier. +* Each metadata file includes recent definition versions, enabling rollbacks without external state. + +## Specification + +### UDF Metadata +The UDF metadata file has the following fields: + +| Requirement | Field name | Type | Description | +|-------------|------------------|------------------------|-----------------------------------------------------------------------| +| *required* | `function-uuid` | `string` | A UUID that identifies this UDF, generated once at creation. | +| *required* | `format-version` | `int` | UDF specification format version (must be `1`). | +| *required* | `definitions` | `list<definition>` | List of function [definition](#definition) entities. | +| *required* | `definition-log` | `list<definition-log>` | History of [definition version history](#definition-log). | +| *optional* | `location` | `string` | The function's base location; used to create metadata file locations. | +| *optional* | `properties` | `map<string,string>` | A string-to-string map of properties. | +| *optional* | `secure` | `boolean` | Whether it is a secure function. Default: `false`. | +| *optional* | `doc` | `string` | Documentation string. | + +Notes: +1. When `secure` is set to `true`, engines should prevent leakage of sensitive information to end users. Each engine + may have its own security definition and mechanisms. It is the administrator's responsibility to ensure that secure + UDFs are properly configured and protected in their environment. +2. Entries in `properties` are treated as hints, not strict rules. +3. UDF names are not stored in metadata. It's the catalog's responsibility to map UDF names to metadata file locations. + +### Definition + +Each `definition` represents one function signature (e.g., `add_one(int)` vs `add_one(float)`). A definition is uniquely +identified by its signature (the ordered list of parameter types). There can be only one definition for a given signature. +All versions within a definition must accept the same signature as specified in the definition's `parameters` field and +must produce values of the declared `return-type`. + +| Requirement | Field name | Type | Description | +|-------------|----------------------|--------------------------------|---------------------------------------------------------------------------------------------------| +| *required* | `definition-id` | `string` | An identifier derived from canonical parameter-type tuple (see [Definition ID](#definition-id)). | +| *required* | `parameters` | `list<parameter>` | Ordered list of [function parameters](#parameter). Invocation order **must** match this list. | +| *required* | `return-type` | `string` or `object` | Declared return type using [Types](#types). | +| *optional* | `return-nullable` | `boolean` | A hint to indicate whether the return value is nullable or not. Default: `true`. | +| *required* | `versions` | `list<definition-version>` | [Versioned implementations](#definition-version) of this definition. | +| *required* | `current-version-id` | `int` | Identifier of the current version for this definition. | +| *required* | `function-type` | `string` (`"udf"` or `"udtf"`) | If `"udtf"`, `return-type` must be a `struct` (see [Types](#types)) describing the output schema. | +| *optional* | `doc` | `string` | Documentation string. | + +### Parameter +| Requirement | Field | Type | Description | +|-------------|--------|----------|--------------------------------------------| +| *required* | `type` | `string` | Parameter data type (see [Types](#types)). | +| *required* | `name` | `string` | Parameter name. | +| *optional* | `doc` | `string` | Parameter documentation. | + +Notes: +1. Variadic (vararg) parameters are not supported. Each definition must declare a fixed number of parameters. + +#### Types +[iceberg-type-json]: https://iceberg.apache.org/spec/#appendix-c-json-serialization +Types are based on the [Iceberg Type](https://iceberg.apache.org/spec/#schemas-and-data-types). + +Primitive and semi-structured type strings are encoded based on [Iceberg Type JSON Representation][iceberg-type-json] +(e.g., `int`, `string`, `timestamp`, `decimal(9,2)`, `variant`). Type strings must contain no spaces or quote characters. + +Nested types (`struct`, `list`, `map`) use the [Iceberg Type JSON Representation][iceberg-type-json] with the +following fields required. Any other fields must be ignored. +* `list` requires `type` and `element`, e.g., `{ "type": "list", "element": "string" }` +* `map` requires `type`, `key`, and `value`, e.g., `{ "type": "map", "key": "string", "value": "int" }` +* `struct` requires `type` and `fields`, where each field requires `name` and `type`, + e.g., `{ "type": "struct", "fields": [ { "name": "id", "type": "int" }, { "name": "name", "type": "string" } ] }` + +#### Definition ID +The `definition-id` is a canonical string derived from the parameter types, formatted as a comma-separated list with no +spaces. Each type uses the string representation described in [Types](#types): + +* Primitives: the type name (e.g., `int`) +* List: `list<element-type>` (e.g., `list<int>`) +* Map: `map<key-type,value-type>` (e.g., `map<string,int>`) +* Struct: `struct<name1:type1,name2:type2,...>` with field names and types (e.g., `struct<id:int,name:string>`) + +Examples of complete definition-id signatures: +* `int` – single int parameter +* `int,string` – two parameters: int and string +* `int,list<int>,struct<id:int,name:string>` – three parameters: an int, a list and a struct + +### Definition Version + +Each definition can evolve over time by introducing new versions. +A `definition version` represents a specific implementation of that definition at a given point in time. A definition version can have multiple SQL representations of different dialects, but only one SQL representation per dialect. + +| Requirement | Field name | Type | Description | +|-------------|-------------------|----------------------------------------------------------|----------------------------------------------------------------| +| *required* | `version-id` | `int` | Monotonically increasing identifier of the definition version. | +| *required* | `representations` | `list<representation>` | [Dialect-specific implementations](#representation). | +| *optional* | `deterministic` | `boolean` (default `false`) | Whether the function is deterministic. | +| *optional* | `on-null-input` | `string` (`"return-null"` or `"call"`, default `"call"`) | Defines how the UDF behaves when any input parameter is NULL. | +| *required* | `timestamp-ms` | `long` (unix epoch millis) | Creation timestamp of this version. | Review Comment: Discussed with @stevenzwu a bit, some points to share: 1. Do we support time-travel for UDF versions? We may never. 2. Do we support rollback to a historical version? We might. Do we support time-based rollback, like rollback to a UDF version 3 days ago? Maybe not. It feels really weird to me, that a user just want to rollback to a version 3 days ago without checking the exact version he/she want to rollback. Version-id based rollback should just work. 3. Most engines(Spark, Trino, Snowflake, Bigquery, Postgres) don't even support UDF versioning. With that, I think it's fine to not enforce monotonically increasing. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
