This document describes the current specification of the new _data connectors_s feature of `graphql-engine`, which is under active development.
The data connectors feature allows `graphql-engine` to delegate the execution of operations to external web services called _agents_. Such agents provide access to a data set, allowing `graphql-engine` to query that data set over a web API.
This document specifies (1) the web API that must be presented by agents, and (2) the precise behaviour of agents for specific reference data sets.
For further reference, the directory in which this document resides contains some implementations of different agents:
- [Reference Implementation](./reference/) - A reference implementation in TypeScript that serves some static in-memory data
## Stability
This specification is complete with regards to the current implementation, but should be considered _unstable_ until the Data Connectors feature is officially released and explicitly marked as a non-experimental feature.
## Setting up Data Connector agents with `graphql-engine`
In order to run one of the example agents, follow the steps in its respective README document.
Once an agent is running, import the following metadata into `graphql-engine`:
The `backend_configs.dataconnector` section lets you set the URIs for as many agents as you'd like. In this case, we've defined one called "reference". When you create a source, the `kind` of the source should be set to the name you gave the agent in the `backend_configs.dataconnector` section (in this case, "reference").
The `configuration` property under the source can contain an 'arbitrary' JSON object, and this JSON will be sent to the agent on every request via the `X-Hasura-DataConnector-Config` header. The example here is configuration that the reference agent uses. The JSON object must conform to the schema specified by the agent from its `/capabilities` endpoint.
The `name` property under the source will be sent to the agent on every request via the `X-Hasura-DataConnector-SourceName` header. This name uniquely identifies a source within an instance of HGE.
The `albums` and `artists` tables should now be available in the GraphiQL console. You should be able to issue queries via the web service. For example:
```graphql
query {
artists {
name
albums {
title
}
}
}
```
## Implementing Data Connector agents
This section is a guide to implementing Data Connector agents for `graphql-engine`. You may find it useful to consult the code examples for reference.
The entry point to the reference agent application is a Fastify HTTP server. Raw data is loaded from JSON files on disk, and the server provides the following endpoints:
-`GET /capabilities`, which returns the capabilities of the agent and a schema that describes the type of the configuration expected to be sent on the `X-Hasura-DataConnector-Config` header
-`GET /schema`, which returns information about the provided _data schema_, its tables and their columns
-`POST /query`, which receives a query structure to be executed, encoded as the JSON request body, and returns JSON containing the requested fields. The query will be over the data schema described by the `/schema` endpoint.
The `/schema`, `/query` and `/mutation` endpoints require the request to have the `X-Hasura-DataConnector-Config` header set. That header contains configuration information that agent can use to configure itself. For example, the header could contain a connection string to the database, if the agent requires a connection string to know how to connect to a specific database. The header must be a JSON object, but the specific properties that are required are up to the agent to define.
The `/schema`, `/query` and `/mutation` endpoints also require the request to have the `X-Hasura-DataConnector-SourceName` header set. This header contains the name of the data source configured in HGE that will be querying the agent. This can be used by the agent to maintain things like connection pools and configuration maps on a per-source basis.
We'll look at the implementation of each of the endpoints in turn.
### Capabilities and configuration schema
The `GET /capabilities` endpoint is used by `graphql-engine` to discover the capabilities supported by the agent, and so that it can know the correct shape of configuration data that needs to be collected from the user and sent to the agent in the `X-Hasura-DataConnector-Config` header. It should return a JSON object similar to the following:
The `config_schema` property contains an [OpenAPI 3 Schema](https://swagger.io/specification/#schema-object) object that represents the schema of the configuration object. It can use references (`$ref`) to refer to other schemas defined in the `other_schemas` object by name.
`graphql-engine` will use the `config_schema` OpenAPI 3 Schema to validate the user's configuration JSON before putting it into the `X-Hasura-DataConnector-Config` header.
The agent can declare whether or not it supports primary keys or foreign keys by setting the `supports_primary_keys` and `supports_foreign_keys` properties under the `data_schema` object on capabilities. If it does not declare support, it is expected that it will not return any such primary/foreign keys in the schema it exposes on the `/schema` endpoint.
If the agent only supports table columns that are always nullable, then it should set `column_nullability` to `"only_nullable"`. However, if it supports both nullable and non-nullable columns, then it should set `"nullable_and_non_nullable"`.
Agents should declare the scalar types they support, along with the comparison operators and aggregate functions on those types.
The agent may optionally specify how to parse values of scalar type by associating it with one of the built-in GraphQL types (`Int`, `Float`, `String`, `Boolean` or `ID`)
Update column operators are operators that can defined to allow custom mutation operations to be performed on columns of the particular scalar type.
These can be defined by adding an `update_column_operators` property, which maps the operator name to an object that defines the operator's `argument_type`. The name specified in capabilities will be prefixed with `_` when it is used in the GraphQL mutation schema.
The `graphql_type` property can be used to tell Hasura GraphQL Engine to parse values of the scalar type as though they were one of the built-in GraphQL scalar types `Int`, `Float`, `String`, `Boolean`, or `ID`.
The agent can declare whether it supports mutations (ie. changing data) against its data source. If it supports mutations, it needs to declare a `mutations` capability with agent-specific values for the following properties:
The agent is able to specify whether or not it supports inserts, updates and deletes separately. For inserts, it can specify whether it supports nested inserts, where the user can insert related rows nested inside the one row insert.
It also should specify its supported level of transactional atomicity when performing mutations. It can choose between the following levels:
-`row`: If multiple rows are affected in a single operation but one fails, only the failed row's changes will be reverted. For example, if one mutation operation inserts four rows, but one row fails, the other three rows will still be inserted, and the failed one will not.
-`single_operation`: If multiple rows are affected in a single operation but one fails, all affected rows in the operation will be reverted. For example, if one mutation operation inserts four rows, but one row fails, none of the rows will be inserted.
-`homogeneous_operations`: If multiple operations of only the same type exist in the one mutation request, a failure in one will result in all changes being reverted. For example, if one mutation request contains two insert operations, one to Table A and one to Table B, and Table B's insert fails, no rows will have been inserted into either Table A nor B.
-`heterogeneous_operations`: If multiple operations of any type exist in the one mutation request, a failure in one will result in all changes being reverted. For example, if one mutation request contains three operations, one to insert some rows, one to update some rows, and one to delete some rows, and the deletion fails, all changes (inserts, updates and deletes) will be reverted.
The preference would be to support the highest level of atomicity possible (ie `heteregeneous_operations` is preferred over `row`). It is also possible to omit the property, which would imply no atomicity at all (failures cannot be rolled back whatsoever).
The agent can also specify whether or not it supports `returning` data from mutations. This refers to the ability to return the data that was mutated by mutation operations (for example, the updated rows in an update, or the deleted rows in a delete).
The agent can declare whether it supports datasets (ie. api for creating/cloning schemas). If it supports datasets, it needs to declare a `datasets` capability:
```json
{
"capabilities": {
"datasets": { }
}
}
```
See [Datasets](#datasets) for information on how this capability is used.
Table names are described as an array of strings. This allows agents to fully qualify their table names with whatever namespacing requirements they have. For example, if the agent connects to a database that puts tables inside schemas, the agent could use table names such as `["my_schema", "my_table"]`.
The `type` of a table can be either a "table" or a "view".
Tables have mutability properties, namely, whether they are "insertable", "updatable" and "deletable", which refers to whether rows can be inserted/updated/deleted from the table. Typically, in an RDBMS, tables are insertable, updatable and deletable, but views may not be. However, an agent may declare the mutability properties in any combination that suits its data source.
Columns also have "insertable" and "updatable" mutability properties. Typically, in an RDBMS, computed columns are neither insertable not updatable, primary keys are insertable but not updatable and normal columns are insertable and updatable. Agents may declare whatever combination suits their data source.
If the agent declares a lack of mutability support in its capabilities, it should not declare tables/columns as mutable in its schema here.
The `SchemaResponse` TypeScript type from [the reference implementation](./reference/src/types/index.ts) describes the valid response body for the `GET /schema` endpoint.
The implementation of the service is responsible for intepreting this data structure and producing a JSON response body which is compatible with both the query and the schema.
- The `table` field tells us which table to fetch the data from, namely the `Artist` table. The table name (ie. the array of strings) must be one that was returned previously by the `/schema` endpoint.
- The `table_relationships` field that lists any relationships used to join between tables in the query. This query does not use any relationships, so this is just an empty list here.
- The `query` field contains further information about how to query the specified table:
- The `where` field tells us that there is currently no (interesting) predicate being applied to the rows of the data set (just an empty conjunction, which ought to return every row).
- The `order_by` field tells us that there is no particular ordering to use, and that we can return data in its natural order.
- The `limit` and `offset` fields tell us that there is no pagination required.
- The `fields` field tells us that we ought to return two fields per row (`ArtistId` and `Name`), and that these fields should be fetched from the columns with the same names. The scalar types of the columns are also denoted with `column_type`.
The rows returned by the query must be put into the `rows` property array in the query response object. Each object within this array represents a row, and the row object properties are the fields requested in the query. The value of the row object properties can be one of two types:
-`column`: The field was a column field, then value of that column for this row is used
-`relationship`: If the field was a relationship field, then a new query response object that contains the results of navigating that relationship for the current row must be used. (The query response structure is recursive via relationship-typed field values). Examples of this can be seen in the Relationships section below.
If the GraphQL query contains pagination information, then the `limit` and `offset` fields may be set to integer values, indicating the number of rows to return, and the index of the first row to return, respectively.
#### Filters
The `where` field contains a recursive expression data structure which should be interpreted as a predicate in the context of each record.
Each node of this recursive expression structure is tagged with a `type` property, which indicates the type of that node, and the node will contain one or more additional fields depending on that type. The valid expression types are enumerated below, along with these additional fields:
| `and` | `expressions` | A conjunction of several subexpressions |
| `or` | `expressions` | A disjunction of several subexpressions |
| `not` | `expression` | The negation of a single subexpression |
| `exists` | `in_table`, `where` | Test if a row exists that matches the `where` subexpression in the specified table (`in_table`) |
| `binary_op` | `operator`, `column`, `value` | Test the specified `column` against a single `value` using a particular binary comparison `operator` |
| `binary_arr_op` | `operator`, `column`, `values`, `value_type` | Test the specified `column` against an array of `values` using a particular binary comparison `operator` where the type of each value is a `value_type` |
| `unary_op` | `operator`, `column` | Test the specified `column` against a particular unary comparison `operator` |
The value of the `in_table` property of the `exists` expression is an object that describes which table to look for rows in. The object is tagged with a `type` property:
| type | Additional fields | Description |
|-------------|---------------------------------|
| `related` | `relationship` | The table is related to the current table via the relationship name specified in `relationship` (this means it should be joined to the current table via the relationship) |
| `unrelated` | `table` | The table specified by `table` is unrelated to the current table and therefore is not explicitly joined to the current table |
The "current table" during expression evaluation is the table specified by the closest ancestor `exists` expression, or if there is no `exists` ancestor, it is the table involved in the Query that the whole `where` Expression is from.
| `in` | The SQL `IN` operator (ie. the column must be any of the array of specified values) |
The available unary comparison operators that can be used against a column:
| Unary comparison operator | Description |
|---------------------------|-------------|
| `is_null` | Tests if a column is null |
Values (as used in `value` in `binary_op` and the `values` array in `binary_arr_op`) are specified as either a literal value, or a reference to another column, which could potentially be in another related table in the same query. The value object is tagged with a `type` property and has different fields based on the type.
| type | Additional fields | Description |
|----------|-------------------|-------------|
| `scalar` | `value` | A scalar `value` to compare against |
| `column` | `column` | A `column` in the current table being queried to compare against |
Columns (as used in `column` fields in `binary_op`, `binary_arr_op`, `unary_op` and in `column`-typed Values) are specified as a column `name`, a `column_type` to denote the scalar type of the column, as well as optionally a `path` to the table that contains the column. If the `path` property is missing/null or an empty array, then the column is on the current table. However, if the path is `["$"]`, then the column is on the table involved in the Query that the whole `where` expression is from. At this point in time, these are the only valid values of `path`.
Exists expressions can be nested, but the `["$"]` path always refers to the query table. So in this example, a person table is filtered by whether or not that person has any children that have any friends that have the same first name as the parent:
In this example, a table is filtered by whether or not an unrelated administrators table contains an admin called "superuser". Note that this means if the administrators table contains the "superuser" admin, then all rows of the table are returned, but if not, no rows are returned.
If the call to `GET /capabilities` returns a `capabilities` record with a `relationships` field then the query structure may include fields corresponding to relationships.
Relationship fields are indicated by a `type` field containing the string `relationship`. Such fields will also include the name of the relationship in a field called `relationship`. This name refers to a relationship that is specified on the top-level query request object in the `table_relationships` field.
This `table_relationships` is a list of tables, and for each table, a map of relationship name to relationship information. The information is an object that has a field `target_table` that specifies the name of the related table. It has a field called `relationship_type` that specified either an `object` (many to one) or an `array` (one to many) relationship. There is also a `column_mapping` field that indicates the mapping from columns in the source table to columns in the related table.
It is intended that the backend should execute the `query` contained in the relationship field and return the resulting query response as the value of this field, with the additional record-level predicate that any mapped columns should be equal in the context of the current record of the current table.
The top-level `table_relationships` can be looked up by starting from the source table (in this case `Artist`), locating the `ArtistAlbums` relationship under that table, then extracting the relationship information. This information includes the `target_table` field which indicates the table to be queried when following this relationship is the `Album` table. The `relationship_type` field indicates that this relationship is an `array` relationship (ie. that it will return zero to many Album rows per Artist row). The `column_mapping` field indicates the column mapping for this relationship, namely that the Artist's `ArtistId` must equal the Album's `ArtistId`.
Back on the relationship field inside the query, there is another `query` field. This indicates the query that should be executed against the `Album` table, but we must remember to enforce the additional constraint between Artist's `ArtistId` and Album's `ArtistId`. That is, in the context of any single outer `Artist` record, we should populate the `Albums` field with the query response containing the array of Album records for which the `ArtistId` field is equal to the outer record's `ArtistId` field.
It is possible to form queries that filter their results by comparing columns across tables via relationships. One way this can happen in Hasura GraphQL Engine is when configuring permissions on a table. It is possible to configure a filter on a table such that it joins to another table in order to compare some data in the filter expression.
The following metadata when used with HGE configures a `Customer` and `Employee` table, and sets up a select permission rule on `Customer` such that only customers that live in the same country as their SupportRep Employee would be visible to users in the `user` role:
The key point of interest here is in the `where` field where we are comparing between columns. Our first expression is an `exists` expression that specifies a row must exist in the table related to the `Customer` table by the `SupportRep` relationship (ie. the `Employee` table). These rows must match a subexpression that compares the related `Employee`'s `Country` column with `equal` to `Customer`'s `Country` column (as indicated by the `["$"]` path). So, in order to evaluate this condition, we'd need to join the `Employee` table using the `column_mapping` specified in the `SupportRep` relationship. Then if any of the related rows (in this case, only one because it is an `object` relation) contain a `Country` that is equal to Customer row's `Country` the `binary_op` would evaluate to True. This would mean a row exists, so the `exists` evaluates to true, and we don't filter out the Customer row.
#### Filtering by Unrelated Tables
It is possible to filter a table by a predicate evaluated against a completely unrelated table. This can happen in Hasura GraphQL Engine when configuring permissions on a table.
In the following example, we are configuring HGE's metadata such that when the Customer table is queried by the employee role, the employee currently doing the query (as specified by the `X-Hasura-EmployeeId` session variable) must be an employee from the city of Calgary, otherwise no rows are returned.
Given this GraphQL query (where the `X-Hasura-Role` session variable is set to `employee`, and the `X-Hasura-EmployeeId` session variable is set to `2`):
The key part in this query is the `where` expression. The root expression in the where is an `exists` expression which specifies that at least one row must exist in the unrelated `["Employee"]` table that satisfies a subexpression. This subexpression asserts that the rows from the Employee table have both `EmployeeId` as `2` and `City` as `Calgary`. The columns referenced inside this subexpression don't have `path` properties, which means they refer the columns on the Employee table because that is the closest ancestor `exists` table.
HGE supports forming GraphQL queries that allow clients to aggregate over the data in their data sources. This type of query can be passed through to Data Connector agents as a part of the Query structure sent to `/query`.
For example, consider the following GraphQL query:
```graphql
query {
Artist_aggregate {
aggregate {
max {
ArtistId
}
}
}
}
```
This would cause the following query request to be performed:
Notice the `Query` has an `aggregates` property; this property contains an object where the property name is the field name of the aggregate, and the value is a description of the aggregate. In the example above, we're using the `max` function on the `ArtistId` column. The `max` function is a function that operates on a single column, so the type of the aggregate is `single_column`.
These are the supported `single_column` functions:
-`avg`
-`max`
-`min`
-`stddev_pop`
-`stddev_samp`
-`sum`
-`var_pop`
-`var_samp`
The aggregate function is to be run over all rows that match the `Query`. In this case, the query has no filters on it (ie. no `where`, `limit` or `offset` properties), so the query would be selecting all rows in the Artist table.
There are two other types of aggregates, `column_count` and `star_count`, as demonstrated in this GraphQL query, and its resultant `QueryRequest`:
A `column_count` aggregate counts the number of rows that have non-null values in the specified `columns`. If `distinct` is set to `true`, then the count should only count unique values of those columns. This is like a `COUNT(x,y,z)` or a `COUNT(DISTINCT x,y,z)` in SQL.
A `star_count` aggregate simply counts the number of rows matched by the query (similar to a `COUNT(*)` in SQL).
The results of the aggregate functions must be returned in an `aggregates` property on the query response. For example:
```json
{
"aggregates": {
"aggregate_distinct_count": 347,
"aggregate_count": 347
}
}
```
HGE's aggregate GraphQL queries can also return the rows involved in the aggregates, as well as apply all the standard filtering operations, for example:
```graphql
query {
Artist_aggregate(where: {Name: {_gt: "Z"}}) {
aggregate {
count
}
nodes {
ArtistId
Name
}
}
}
```
The `nodes` part of the query ends up as standard `fields` in the `Query`, and therefore are treated exactly the same as discussed in previous sections:
This would be expected to return the following response, with the rows from the Artist table, and the aggregates from the related Albums nested under the relationship field values for each Album row:
The `elements` field specifies an array of one-or-more ordering elements. Each element represents a "target" to order, and a direction to order by. The direction can either be `asc` (ascending) or `desc` (descending). If there are multiple elements specified, then rows should be ordered with earlier elements in the array taking precedence. In the above example, rows are principally ordered by `last_name`, delegating to `first_name` in the case where two last names are equal.
The order by element `target` is specified as an object, whose `type` property specifies a different sort of ordering target:
| type | Additional fields | Description |
|------|-------------------|-------------|
| `column` | `column` | Sort by the `column` specified |
| `star_count_aggregate` | - | Sort by the count of all rows on the related target table (a non-empty `target_path` will always be specified) |
| `single_column_aggregate` | `function`, `column` | Sort by the value of applying the specified aggregate function to the column values of the rows in the related target table (a non-empty `target_path` will always be specified) |
The `target_path` property is a list of relationships to navigate before finding the `target` to sort on. This is how sorting on columns or aggregates on related tables is expressed. Note that aggregate-typed targets will never be found on the current table (ie. a `target_path` of `[]`) and are always applied to a related table.
Here's an example of applying an ordering by a related table; the Album table is being queried and sorted by the Album's Artist's Name.
Note that the `target_path` specifies the relationship path of `["Artist"]`, and that this relationship is defined in the top-level `table_relationships`. The ordering element target column `Name` would therefore be found on the `Artist` table after joining to it from each `Album`. (See the [Relationships](#Relationships) section for more information about relationships.)
The `relations` property of `order_by` will contain all the relations used in the order by, for the purpose of specifying filters that must be applied to the joined tables before using them for sorting. The `relations` property captures all `target_path`s used in the `order_by` in a recursive fashion, so for example, if the following `target_path`s were used in the `order_by`'s `elements`:
*`["Artist", "Albums"]`
*`["Artist"]`
*`["Tracks"]`
Then the value of the `relations` property would look like this:
```json
{
"Artist": {
"where": null,
"subrelations": {
"Albums": {
"where": null,
"subrelations": {}
}
}
},
"Tracks": {
"where": null,
"subrelations": {}
}
}
```
The `where` properties may contain filtering expressions that must be applied to the joined table before using it for sorting. The filtering expressions are defined in the same manner as specified in the [Filters](#Filters) section of this document, where they are used on the `where` property of Queries.
For example, here's a query that retrieves artists ordered descending by the count of all their albums where the album title is greater than 'T'.
The `QueryRequest` TypeScript type in the [reference implementation](./reference/src/types/index.ts) describes the valid request body payloads which may be passed to the `POST /query` endpoint. The response body structure is captured by the `QueryResponse` type.
Agents must expose a `/health` endpoint which must return a 204 No Content HTTP response code if the agent is up and running. This does not mean that the agent is able to connect to any data source it performs queries against, only that the agent is running and can accept requests, even if some of those requests might fail because a dependant service is unavailable.
However, this endpoint can also be used to check whether the ability of the agent to talk to a particular data source is healthy. If the endpoint is sent the `X-Hasura-DataConnector-Config` and `X-Hasura-DataConnector-SourceName` headers, then the agent is expected to check that it can successfully talk to whatever data source is being specified by those headers. If it can do so, then it must return a 204 No Content response code.
Any non-200 response code from an Agent (except for the `/health` endpoint) will be interpreted as an error. These should be handled gracefully by `graphql-engine` but provide limited details to users. If you wish to return structured error information to users you can return a status of `500`, or `400` from the `/capabilities`, `/schema`, and `/query` endpoints with the following JSON format:
"type": String, // A specific error type, see below
"message": String, // A plain-text message for display purposes
"details": Value // An arbitrary JSON Value containing error details
}
```
The available error types are:
*`mutation-constraint-violation`: For when a mutation request fails because the mutation causes a violation of data constraints (for example, primary key constraint) in the data source
*`mutation-permission-check-failure`: For when a permissions check fails during a mutation and the mutation is rejected
*`uncaught-error`: For all other errors
### Mutations
The `POST /mutation` endpoint is invoked when the user issues a mutation GraphQL request to `graphql-engine`, assuming the agent has declared itself capable of mutations in its capabilities. The basic structure of a mutation request is as follows:
```jsonc
{
"table_relationships": [], // Any relationships between tables are described in here in the same manner as in queries
"operations": [ // A mutation request can contain multiple mutation operations
{
"type": "insert", // Also: "update" and "delete"
"returning_fields": { // The fields to return for every affected row
"ArtistId": {
"type": "column",
"column": "ArtistId",
"column_type": "number"
}
},
... // Other operation type-specific properties, detailed below
}
]
}
```
There are three types of mutation operations: `insert`, `update` and `delete`. A request can involve multiple mutation operations, potentially of differing types. A mutation operation can specify `returning_fields` which are the fields that are expected to be returned in the response for each row affected by the mutation operation.
The response to a mutation request takes this basic structure:
```jsonc
{
"operation_results": [ // There will be a result object per operation, returned here in the same order as in the request
{
"affected_rows": 1, // The number of rows affected by the mutation operation
"returning": [ // The rows that were affected; each row object contains the fields requested in `returning_fields`
{
"FieldName": "FieldValue"
}
]
}
]
}
```
If any mutation operation causes an error, for example, if a mutation violates a constraint such as a primary key or a foreign key constraint in an RDBMS, then an error should be returned as a response in the same manner as described in the [Reporting Errors](#reporting-errors) section. Changes should be rolled back to the extent described by the `atomicity_level` declared by in the agent's [mutation capabilities](#mutations-capabilities). The error type should be `mutation_constraint_violation` and the HTTP response code should be 400. For example:
```json
{
"type": "mutation-constraint-violation",
"message": "Violation of PRIMARY KEY constraint PK_Artist. Cannot insert duplicate key in table Artist. The duplicate key value is (1).", // Can be any helpfully descriptive error message
"details": { // Any helpful structured error information, the below is just an example
"constraint_name": "PK_Artist",
"table": ["Artist"],
"key_value": 1
}
}
```
If a mutation fails because it fails a permissions check (eg a `post-insert-check`), then the error code that should be used is `mutation-permission-check-failure`.
#### Insert Operations
Here's an example GraphQL mutation that inserts two artists:
```graphql
mutation InsertArtists {
insert_Artist(objects: [
{ArtistId: 300, Name: "Taylor Swift"},
{ArtistId: 301, Name: "Phil Collins"}
]) {
affected_rows
returning {
ArtistId
Name
}
}
}
```
This would result in a mutation request like this:
```json
{
"table_relationships": [],
"insert_schema": [
{
"table": ["Artist"],
"fields": {
"ArtistId": {
"type": "column",
"column": "ArtistId",
"column_type": "number"
},
"Name": {
"type": "column",
"column": "Name",
"column_type": "string"
}
}
}
],
"operations": [
{
"type": "insert",
"table": ["Artist"],
"rows": [
[
{
"ArtistId": 300,
"Name": "Taylor Swift"
},
{
"ArtistId": 301,
"Name": "Phil Collins"
}
]
],
"post_insert_check": {
"type": "and",
"expressions": []
},
"returning_fields": {
"ArtistId": {
"type": "column",
"column": "ArtistId",
"column_type": "number"
},
"Name": {
"type": "column",
"column": "Name",
"column_type": "string"
}
}
}
]
}
```
The first thing to notice is the `insert_schema` property at the mutation request level. This contains the definition of the fields that will be used inside any insert operation in this request on a per-table basis. The schema for the row data to insert is placed here, separate to the row data itself, in order to reduce the amount of duplicate data that would exist were it inlined into the row data structures themselves.
So, because in this request we are inserting into the Artist table, we have the definition of what "ArtistId" and "Name" properties mean when they are found in the rows to insert for the Artist table. In this case, both fields are columns (`"type": "column"`) with their names (`column`) and types (`column_type`) specified.
Next, let's break the `insert`-typed operation's properties down:
*`table`: specifies the table we're inserting rows into
*`rows`: An array of rows to insert. Each row is an object with properties, where what the properties correspond to (eg. column values) is defined by the `insert_schema`.
*`post_insert_check`: The post-insert check is an expression (in the same format as `Query`'s `where` property) that all inserted rows must match otherwise their insertion must be reverted. This expression comes from `graphql-engine`'s permissions system. The reason that it is a "post-insert" check is because it can involve joins via relationships to other tables and potentially data that is only available post-insert such as computed columns. If the agent knows it can compute the result of such a check without actually performing an insert, it is free to do so, but it must produce a result that is indistinguishable from that which was done post-insert. If the post-insert check fails, the mutation request should fail with an error using the error code `mutation-permission-check-failure`.
*`returning_fields`: This specifies a list of fields to return in the response. The property takes the same format as the `fields` property on Queries. It is expected that the specified fields will be returned for all rows affected by the insert (ie. all inserted rows).
The result of this request would be the following response:
```json
{
"operation_results": [
{
"affected_rows": 2,
"returning": [
{
"ArtistId": 300,
"Name": "Taylor Swift"
},
{
"ArtistId": 301,
"Name": "Phil Collins"
}
]
}
]
}
```
Notice that the two affected rows in `returning` are the two that we inserted.
#### Nested Insert Operations
If a user wishes to insert multiple related rows in one go, they can issue a nested insert GraphQL query:
```graphql
mutation InsertAlbum {
insert_Album(objects: [
{
AlbumId: 400,
Title: "Fearless",
Artist: {
data: {
ArtistId: 300,
Name: "Taylor Swift"
}
},
Tracks: {
data: [
{ TrackId: 4000, Name: "Fearless" },
{ TrackId: 4001, Name: "Fifteen" }
]
}
}
]) {
affected_rows
returning {
AlbumId
Title
Artist {
ArtistId
Name
}
Tracks {
TrackId
Name
}
}
}
}
```
This would result in the following request:
```json
{
"table_relationships": [
{
"source_table": ["Album"],
"relationships": {
"Artist": {
"target_table": ["Artist"],
"relationship_type": "object",
"column_mapping": {
"ArtistId": "ArtistId"
}
},
"Tracks": {
"target_table": ["Track"],
"relationship_type": "array",
"column_mapping": {
"AlbumId": "AlbumId"
}
}
}
}
],
"insert_schema": [
{
"table": ["Album"],
"fields": {
"AlbumId": {
"type": "column",
"column": "AlbumId",
"column_type": "number"
},
"Title": {
"type": "column",
"column": "Title",
"column_type": "string"
},
"Artist": {
"type": "object_relation",
"relationship": "Artist",
"insert_order": "before_parent"
},
"Tracks": {
"type": "array_relation",
"relationship": "Tracks"
},
}
},
{
"table": ["Artist"],
"fields": {
"ArtistId": {
"type": "column",
"column": "ArtistId",
"column_type": "number"
},
"Name": {
"type": "column",
"column": "Name",
"column_type": "string"
}
}
},
{
"table": ["Track"],
"fields": {
"TrackId": {
"type": "column",
"column": "TrackId",
"column_type": "number"
},
"Name": {
"type": "column",
"column": "Name",
"column_type": "string"
}
}
}
],
"operations": [
{
"type": "insert",
"table": ["Album"],
"rows": [
[
{
"AlbumId": 400,
"Title": "Fearless",
"Artist": {
"ArtistId": 300,
"Name": "Taylor Swift"
},
"Tracks": [
{
"TrackId": 4000,
"Name": "Fearless"
},
{
"TrackId": 4001,
"Name": "Fifteen"
}
]
}
]
],
"post_insert_check": {
"type": "and",
"expressions": []
},
"returning_fields": {
"AlbumId": {
"type": "column",
"column": "AlbumId",
"column_type": "number"
},
"Title": {
"type": "column",
"column": "Title",
"column_type": "string"
},
"Artist": {
"type": "relationship",
"relationship": "Artist",
"query": {
"fields": {
"ArtistId": {
"type": "column",
"column": "ArtistId",
"column_type": "number"
},
"Name": {
"type": "column",
"column": "Name",
"column_type": "string"
}
}
}
},
"Tracks": {
"type": "relationship",
"relationship": "Tracks",
"query": {
"fields": {
"TrackId": {
"type": "column",
"column": "TrackId",
"column_type": "number"
},
"Name": {
"type": "column",
"column": "Name",
"column_type": "string"
}
}
}
}
}
}
]
}
```
Note that there are two new types of fields in the `insert_schema` in this query to capture the nested inserts:
*`object_relation`: This captures a nested insert across an object relationship. In this case, we're inserting the related Artist row.
*`relationship`: The name of the relationship across which to insert the related row. The information about this relationship can be looked up in `table_relationships`.
*`insert_order`: This can be either `before_parent` or `after_parent` and indicates whether or not the related row needs to be inserted before the parent row or after it.
*`array_relation`: This captures a nested insert across an array relationship. In this case, we're inserting the related Tracks rows.
*`relationship`: The name of the relationship across which to insert the related rows. The information about this relationship can be looked up in `table_relationships`.
The agent is expected to set the necessary values of foreign key columns itself when inserting all the rows. In this example, the agent would:
* First insert the Artist.
* Then insert the Album, using the Artist.ArtistId primary key column for the Album.ArtistId foreign key column.
* Then insert the two Track rows, using the Album.AlbumId primary key column for the Track.AlbumId foreign key column.
This is particularly important where the value of primary keys are not known until they are generated in the database itself and cannot be provided by the user.
Note that in `returning_fields` we have used fields of type `relationship` to navigate relationships in the returned affected rows. This works in the same way as in Queries.
The response to this mutation request would be:
```json
{
"operation_results": [
{
"affected_rows": 4,
"returning": [
{
"AlbumId": 400,
"Title": "Fearless",
"Artist": {
"rows": [
{
"ArtistId": 300,
"Name": "Taylor Swift"
}
]
},
"Tracks": {
"rows": [
{
"TrackId": 4000,
"Name": "Fearless"
},
{
"ArtistId": 4001,
"Name": "Fifteen"
}
]
}
}
]
}
]
}
```
Note that relationship fields are returned in the response in the same fashion as they are in a Query response; ie. inside a nested object with `rows` and `aggregates` (if specified) properties.
#### Update Operations
Here's an example of a mutation that updates a Track row:
Breaking down the properties in the `update`-typed mutation operation:
*`table`: specifies the table we're updating rows in
*`where`: An expression (same as the expression in a Query's `where` property) that is used to select the matching rows to update
*`updates`: An array of `RowUpdate` objects that describe the individual updates to be applied to each row that matches the expression in `where`. There are two types of `RowUpdate`s:
*`set` - This sets the specified column to the specified value
*`custom_operator` - This defines a mutation to a column using a custom update column operator defined by the agent in its capabilities. The `operator_name` property specifies which operator was used. In this case, `inc` is the custom operator.
*`post_update_check`: The post-update check is an expression (in the same format as `Query`'s `where` property) that all updated rows must match otherwise the changes made must be reverted. This expression comes from `graphql-engine`'s permissions system. The reason that it is a "post-update" check is because it operates on the post-update data (such as the results of increment updates), can involve joins via relationships to other tables, and can potentially involve data that is only available post-insert such as computed columns. If the agent knows it can compute the result of such a check without actually performing an update, it is free to do so, but it must produce a result that is indistinguishable from that which was done post-update. If the post-update check fails, the mutation request should fail with an error using the error code `mutation-permission-check-failure`.
*`returning_fields`: This specifies a list of fields to return in the response. The property takes the same format as the `fields` property on Queries. It is expected that the specified fields will be returned for all rows affected by the update (ie. all updated rows).
Update operations return responses that are the same as insert operations, except the affected rows in `returning` are naturally the updated rows instead.
#### Delete Operations
Here's an example of a mutation that deletes a Track row:
```graphql
mutation UpdateTrack {
delete_Track(
where: { TrackId: { _eq: 1 } },
) {
affected_rows
returning {
TrackId
Milliseconds
}
}
}
```
This would cause a mutation request to be send that looks like this:
Breaking down the properties in the `delete`-typed mutation operation:
*`table`: specifies the table we're deleting rows from
*`where`: An expression (same as the expression in a Query's `where` property) that is used to select the matching rows to delete
*`returning_fields`: This specifies a list of fields to return in the response. The property takes the same format as the `fields` property on Queries. It is expected that the specified fields will be returned for all rows affected by the deletion (ie. all deleted rows).
Delete operations return responses that are the same as insert and update operations, except the affected rows in `returning` are the deleted rows instead.
The `/datasets` resource is available to use in order to create new databases/schemas from templates.
Datasets are represented by abstract names referencing database-schema templates that can be cloned from and clones that can be used via config and deleted. This feature is required for testing the mutations feature, but may also have non-test uses - for example - spinning up interactive demo projects.
The `/datasets/:name` resource has the following methods:
The `POST` method is the most significant way to interact with the API. It allows for cloning a dataset template to a new name. The new name can be used to delete the dataset, and the config returned from the POST API call can be used as the config header for non-dataset interactions such as queries.
The following diagram shows the interactions between the various datatypes and resource methods: