Adapters

Schema adapters

A schema adapter allows Calcite to read particular kind of data, presenting the data as tables within a schema.

Other language interfaces

Engines

Many projects and products use Apache Calcite for SQL parsing, query optimization, data virtualization/federation, and materialized view rewrite. Some of them are listed on the “powered by Calcite” page.

Drivers

A driver allows you to connect to Calcite from your application.

The JDBC driver is powered by Avatica. Connections can be local or remote (JSON over HTTP or Protobuf over HTTP).

The basic form of the JDBC connect string is

jdbc:calcite:property=value;property2=value2

where property, property2 are properties as described below. (Connect strings are compliant with OLE DB Connect String syntax, as implemented by Avatica’s ConnectStringParser.)

JDBC connect string parameters

Property Description
approximateDecimal Whether approximate results from aggregate functions on DECIMAL types are acceptable.
approximateDistinctCount Whether approximate results from COUNT(DISTINCT ...) aggregate functions are acceptable.
approximateTopN Whether approximate results from “Top N” queries (ORDER BY aggFun() DESC LIMIT n) are acceptable.
caseSensitive Whether identifiers are matched case-sensitively. If not specified, value from lex is used.
conformance SQL conformance level. Values: DEFAULT (the default, similar to PRAGMATIC_2003), LENIENT, MYSQL_5, ORACLE_10, ORACLE_12, PRAGMATIC_99, PRAGMATIC_2003, STRICT_92, STRICT_99, STRICT_2003, SQL_SERVER_2008.
createMaterializations Whether Calcite should create materializations. Default false.
defaultNullCollation How NULL values should be sorted if neither NULLS FIRST nor NULLS LAST are specified in a query. The default, HIGH, sorts NULL values the same as Oracle.
druidFetch How many rows the Druid adapter should fetch at a time when executing SELECT queries.
forceDecorrelate Whether the planner should try de-correlating as much as possible. Default true.
fun Collection of built-in functions and operators. Valid values are “standard” (the default), “oracle”, “spatial”, and may be combined using commas, for example “oracle,spatial”.
lex Lexical policy. Values are ORACLE (default), MYSQL, MYSQL_ANSI, SQL_SERVER, JAVA.
materializationsEnabled Whether Calcite should use materializations. Default false.
model URI of the JSON model file.
parserFactory Parser factory. The name of a class that implements interface SqlParserImplFactory and has a public default constructor or an INSTANCE constant.
quoting How identifiers are quoted. Values are DOUBLE_QUOTE, BACK_QUOTE, BRACKET. If not specified, value from lex is used.
quotedCasing How identifiers are stored if they are quoted. Values are UNCHANGED, TO_UPPER, TO_LOWER. If not specified, value from lex is used.
schema Name of initial schema.
schemaFactory Schema factory. The name of a class that implements interface SchemaFactory and has a public default constructor or an INSTANCE constant. Ignored if model is specified.
schemaType Schema type. Value must be “MAP” (the default), “JDBC”, or “CUSTOM” (implicit if schemaFactory is specified). Ignored if model is specified.
spark Specifies whether Spark should be used as the engine for processing that cannot be pushed to the source system. If false (the default), Calcite generates code that implements the Enumerable interface.
timeZone Time zone, for example “gmt-3”. Default is the JVM’s time zone.
typeSystem Type system. The name of a class that implements interface RelDataTypeSystem and has a public default constructor or an INSTANCE constant.
unquotedCasing How identifiers are stored if they are not quoted. Values are UNCHANGED, TO_UPPER, TO_LOWER. If not specified, value from lex is used.

To make a connection to a single schema based on a built-in schema type, you don’t need to specify a model. For example,

jdbc:calcite:schemaType=JDBC; schema.jdbcUser=SCOTT; schema.jdbcPassword=TIGER; schema.jdbcUrl=jdbc:hsqldb:res:foodmart

creates a connection with a schema mapped via the JDBC schema adapter to the foodmart database.

Similarly, you can connect to a single schema based on a user-defined schema adapter. For example,

jdbc:calcite:schemaFactory=org.apache.calcite.adapter.cassandra.CassandraSchemaFactory; schema.host=localhost; schema.keyspace=twissandra

makes a connection to the Cassandra adapter, equivalent to writing the following model file:

{
  "version": "1.0",
  "defaultSchema": "foodmart",
  "schemas": [
    {
      type: 'custom',
      name: 'twissandra',
      factory: 'org.apache.calcite.adapter.cassandra.CassandraSchemaFactory',
      operand: {
        host: 'localhost',
        keyspace: 'twissandra'
      }
    }
  ]
}

Note how each key in the operand section appears with a schema. prefix in the connect string.

Server

Calcite’s core module (calcite-core) supports SQL queries (SELECT) and DML operations (INSERT, UPDATE, DELETE, MERGE) but does not support DDL operations such as CREATE SCHEMA or CREATE TABLE. As we shall see, DDL complicates the state model of the repository and makes the parser more difficult to extend, so we left DDL out of core.

The server module (calcite-server) adds DDL support to Calcite. It extends the SQL parser, using the same mechanism used by sub-projects, adding some DDL commands:

  • CREATE and DROP SCHEMA
  • CREATE and DROP FOREIGN SCHEMA
  • CREATE and DROP TABLE (including CREATE TABLE ... AS SELECT)
  • CREATE and DROP MATERIALIZED VIEW
  • CREATE and DROP VIEW

Commands are described in the SQL reference.

To enable, include calcite-server.jar in your class path, and add parserFactory=org.apache.calcite.sql.parser.ddl.SqlDdlParserImpl#FACTORY to the JDBC connect string (see connect string property parserFactory). Here is an example using the sqlline shell.

$ ./sqlline
sqlline version 1.3.0
> !connect jdbc:calcite:parserFactory=org.apache.calcite.sql.parser.ddl.SqlDdlParserImpl#FACTORY sa ""
> CREATE TABLE t (i INTEGER, j VARCHAR(10));
No rows affected (0.293 seconds)
> INSERT INTO t VALUES (1, 'a'), (2, 'bc');
2 rows affected (0.873 seconds)
> CREATE VIEW v AS SELECT * FROM t WHERE i > 1;
No rows affected (0.072 seconds)
> SELECT count(*) FROM v;
+---------------------+
|       EXPR$0        |
+---------------------+
| 1                   |
+---------------------+
1 row selected (0.148 seconds)
> !quit

The calcite-server module is optional. One of its goals is to showcase Calcite’s capabilities (for example materialized views, foreign tables and generated columns) using concise examples that you can try from the SQL command line. All of the capabilities used by calcite-server are available via APIs in calcite-core.

If you are the author of a sub-project, it is unlikely that your syntax extensions match those in calcite-server, so we recommend that you add your SQL syntax extensions by extending the core parser; if you want DDL commands, you may be able to copy-paste from calcite-server into your project.

At present, the repository is not persisted. As you execute DDL commands, you are modifying an in-memory repository by adding and removing objects reachable from a root Schema. All commands within the same SQL session will see those objects. You can create the same objects in a future session by executing the same script of SQL commands.

Calcite could also act as a data virtualization or federation server: Calcite manages data in multiple foreign schemas, but to a client the data all seems to be in the same place. Calcite chooses where processing should occur, and whether to create copies of data for efficiency. The calcite-server module is a step towards that goal; an industry-strength solution would require further on packaging (to make Calcite runnable as a service), repository persistence, authorization and security.

Extensibility

There are many other APIs that allow you to extend Calcite’s capabilities.

In this section, we briefly describe those APIs, to give you an idea what is possible. To fully use these APIs you will need to read other documentation such as the javadoc for the interfaces, and possibly seek out the tests that we have written for them.

Functions and operators

There are several ways to add operators or functions to Calcite. We’ll describe the simplest (and least powerful) first.

User-defined functions are the simplest (but least powerful). They are straightforward to write (you just write a Java class and register it in your schema) but do not offer much flexibility in the number and type of arguments, resolving overloaded functions, or deriving the return type.

It you want that flexibility, you probably need to write you a user-defined operator (see interface SqlOperator).

If your operator does not adhere to standard SQL function syntax, “f(arg1, arg2, ...)”, then you need to extend the parser.

There are many good examples in the tests: class UdfTest tests user-defined functions and user-defined aggregate functions.

Aggregate functions

User-defined aggregate functions are similar to user-defined functions, but each function has several corresponding Java methods, one for each stage in the life-cycle of an aggregate:

  • init creates an accumulator;
  • add adds one row’s value to an accumulator;
  • merge combines two accumulators into one;
  • result finalizes an accumulator and converts it to a result.

For example, the methods (in pseudo-code) for SUM(int) are as follows:

struct Accumulator {
  final int sum;
}
Accumulator init() {
  return new Accumulator(0);
}
Accumulator add(Accumulator a, int x) {
  return new Accumulator(a.sum + x);
}
Accumulator merge(Accumulator a, Accumulator a2) {
  return new Accumulator(a.sum + a2.sum);
}
int result(Accumulator a) {
  return new Accumulator(a.sum + x);
}

Here is the sequence of calls to compute the sum of two rows with column values 4 and 7:

a = init()    # a = {0}
a = add(a, 4) # a = {4}
a = add(a, 7) # a = {11}
return result(a) # returns 11

Window functions

A window function is similar to an aggregate function but it is applied to a set of rows gathered by an OVER clause rather than by a GROUP BY clause. Every aggregate function can be used as a window function, but there are some key differences. The rows seen by a window function may be ordered, and window functions that rely upon order (RANK, for example) cannot be used as aggregate functions.

Another difference is that windows are non-disjoint: a particular row can appear in more than one window. For example, 10:37 appears in both the 9:00-10:00 hour and also the 9:15-9:45 hour.

Window functions are computed incrementally: when the clock ticks from 10:14 to 10:15, two rows might enter the window and three rows leave. For this, window functions have have an extra life-cycle operation:

  • remove removes a value from an accumulator.

It pseudo-code for SUM(int) would be:

Accumulator remove(Accumulator a, int x) {
  return new Accumulator(a.sum - x);
}

Here is the sequence of calls to compute the moving sum, over the previous 2 rows, of 4 rows with values 4, 7, 2 and 3:

a = init()       # a = {0}
a = add(a, 4)    # a = {4}
emit result(a)   # emits 4
a = add(a, 7)    # a = {11}
emit result(a)   # emits 11
a = remove(a, 4) # a = {7}
a = add(a, 2)    # a = {9}
emit result(a)   # emits 9
a = remove(a, 7) # a = {2}
a = add(a, 3)    # a = {5}
emit result(a)   # emits 5

Grouped window functions

Grouped window functions are functions that operate the GROUP BY clause to gather together records into sets. The built-in grouped window functions are HOP, TUMBLE and SESSION. You can define additional functions by implementing interface SqlGroupedWindowFunction.

Table functions and table macros

User-defined table functions are defined in a similar way to regular “scalar” user-defined functions, but are used in the FROM clause of a query. The following query uses a table function called Ramp:

SELECT * FROM TABLE(Ramp(3, 4))

User-defined table macros use the same SQL syntax as table functions, but are defined differently. Rather than generating data, they generate an relational expression. Table macros are invoked during query preparation and the relational expression they produce can then be optimized. (Calcite’s implementation of views uses table macros.)

class TableFunctionTest tests table functions and contains several useful examples.

Extending the parser

Suppose you need to extend Calcite’s SQL grammar in a way that will be compatible with future changes to the grammar. Making a copy of the grammar file Parser.jj in your project would be foolish, because the grammar is edited quite frequently.

Fortunately, Parser.jj is actually an Apache FreeMarker template that contains variables that can be substituted. The parser in calcite-core instantiates the template with default values of the variables, typically empty, but you can override. If your project would like a different parser, you can provide your own config.fmpp and parserImpls.ftl files and therefore generate an extended parser.

The calcite-server module, which was created in [CALCITE-707] and adds DDL statements such as CREATE TABLE, is an example that you could follow. Also see class ExtensionSqlParserTest.

Customizing SQL dialect accepted and generated

To customize what SQL extensions the parser should accept, implement interface SqlConformance or use one of the built-in values in enum SqlConformanceEnum.

To control how SQL is generated for an external database (usually via the JDBC adapter), use class SqlDialect. The dialect also describes the engine’s capabilities, such as whether it supports OFFSET and FETCH clauses.

Defining a custom schema

To define a custom schema, you need to implement interface SchemaFactory.

During query preparation, Calcite will call this interface to find out what tables and sub-schemas your schema contains. When a table in your schema is referenced in a query, Calcite will ask your schema to create an instance of interface Table.

That table will be wrapped in a TableScan and will undergo the query optimization process.

Reflective schema

A reflective schema (class ReflectiveSchema) is a way of wrapping a Java object so that it appears as a schema. Its collection-valued fields will appear as tables.

It is not a schema factory but an actual schema; you have to create the object and wrap it in the schema by calling APIs.

See class ReflectiveSchemaTest.

Defining a custom table

To define a custom table, you need to implement interface TableFactory. Whereas a schema factory a set of named tables, a table factory produces a single table when bound to a schema with a particular name (and optionally a set of extra operands).

Modifying data

If your table is to support DML operations (INSERT, UPDATE, DELETE, MERGE), your implementation of interface Table must implement interface ModifiableTable.

Streaming

If your table is to support streaming queries, your implementation of interface Table must implement interface StreamableTable.

See class StreamTest for examples.

Pushing operations down to your table

If you wish to push processing down to your custom table’s source system, consider implementing either interface FilterableTable or interface ProjectableFilterableTable.

If you want more control, you should write a planner rule. This will allow you to push down expressions, to make a cost-based decision about whether to push down processing, and push down more complex operations such as join, aggregation, and sort.

Type system

You can customize some aspects of the type system by implementing interface RelDataTypeSystem.

Relational operators

All relational operators implement interface RelNode and most extend class AbstractRelNode. The core operators (used by SqlToRelConverter and covering conventional relational algebra) are TableScan, TableModify, Values, Project, Filter, Aggregate, Join, Sort, Union, Intersect, Minus, Window and Match.

Each of these has a “pure” logical sub-class, LogicalProject and so forth. Any given adapter will have counterparts for the operations that its engine can implement efficiently; for example, the Cassandra adapter has CassandraProject but there is no CassandraJoin.

You can define your own sub-class of RelNode to add a new operator, or an implementation of an existing operator in a particular engine.

To make an operator useful and powerful, you will need planner rules to combine it with existing operators. (And also provide metadata, see below). This being algebra, the effects are combinatorial: you write a few rules, but they combine to handle an exponential number of query patterns.

If possible, make your operator a sub-class of an existing operator; then you may be able to re-use or adapt its rules. Even better, if your operator is a logical operation that you can rewrite (again, via a planner rule) in terms of existing operators, you should do that. You will be able to re-use the rules, metadata and implementations of those operators with no extra work.

Planner rule

A planner rule (class RelOptRule) transforms a relational expression into an equivalent relational expression.

A planner engine has many planner rules registered and fires them to transform the input query into something more efficient. Planner rules are therefore central to the optimization process, but surprisingly each planner rule does not concern itself with cost. The planner engine is responsible for firing rules in a sequence that produces an optimal plan, but each individual rules only concerns itself with correctness.

Calcite has two built-in planner engines: class VolcanoPlanner uses dynamic programming and is good for exhaustive search, whereas class HepPlanner fires a sequence of rules in a more fixed order.

Calling conventions

A calling convention is a protocol used by a particular data engine. For example, the Cassandra engine has a collection of relational operators, CassandraProject, CassandraFilter and so forth, and these operators can be connected to each other without the data having to be converted from one format to another.

If data needs to be converted from one calling convention to another, Calcite uses a special sub-class of relational expression called a converter (see class Converter). But of course converting data has a runtime cost.

When planning a query that uses multiple engines, Calcite “colors” regions of the relational expression tree according to their calling convention. The planner pushes operations into data sources by firing rules. If the engine does not support a particular operation, the rule will not fire. Sometimes an operation can occur in more than one place, and ultimately the best plan is chosen according to cost.

A calling convention is a class that implements interface Convention, an auxiliary interface (for instance interface CassandraRel), and a set of sub-classes of class RelNode that implement that interface for the core relational operators (Project, Filter, Aggregate, and so forth).

Built-in SQL implementation

How does Calcite implement SQL, if an adapter does not implement all of the core relational operators?

The answer is a particular built-in calling convention, EnumerableConvention. Relational expressions of enumerable convention are implemented as “built-ins”: Calcite generates Java code, compiles it, and executes inside its own JVM. Enumerable convention is less efficient than, say, a distributed engine running over column-oriented data files, but it can implement all core relational operators and all built-in SQL functions and operators. If a data source cannot an implement a relational operator, enumerable convention is a fall-back.

Statistics and cost

Calcite has a metadata system that allow you to define cost functions and statistics about relational operators, collectively referred to as metadata. Each kind of metadata has an interface with (usually) one method. For example, selectivity is defined by interface RelMdSelectivity and the method getSelectivity(RelNode rel, RexNode predicate).

There are many built-in kinds of metadata, including collation, column origins, column uniqueness, distinct row count, distribution, explain visibility, expression lineage, max row count, node types, parallelism, percentage original rows, population size, predicates, row count, selectivity, size, table references, unique keys, and selectivity; you can also define your own.

You can then supply a metadata provider that computes that kind of metadata for particular sub-classes of RelNode. Metadata providers can handle built-in and extended metadata types, and built-in and extended RelNode types. While preparing a query Calcite combines all of the applicable metadata providers and maintains a cache so that a given piece of metadata (for example the selectivity of the condition x > 10 in a particular Filter operator) is computed only once.