This is a step-by-step tutorial that shows how to build and connect to Calcite. It uses a simple adapter that makes a directory of CSV files appear to be a schema containing tables. Calcite does the rest, and provides a full SQL interface.
Calcite-example-CSV is a fully functional adapter for Calcite that reads text files in CSV (comma-separated values) format. It is remarkable that a couple of hundred lines of Java code are sufficient to provide full SQL query capability.
CSV also serves as a template for building adapters to other data formats. Even though there are not many lines of code, it covers several important concepts:
- user-defined schema using SchemaFactory and Schema interfaces;
- declaring schemas in a model JSON file;
- declaring views in a model JSON file;
- user-defined table using the Table interface;
- determining the record type of a table;
- a simple implementation of Table, using the ScannableTable interface, that enumerates all rows directly;
- a more advanced implementation that implements FilterableTable, and can filter out rows according to simple predicates;
- advanced implementation of Table, using TranslatableTable, that translates to relational operators using planner rules.
Download and build
You need Java (version 8, 9 or 10) and Git.
Now let’s connect to Calcite using sqlline, a SQL shell that is included in this project.
(If you are running Windows, the command is
Execute a metadata query:
(JDBC experts, note: sqlline’s
!tables command is just executing
behind the scenes.
It has other commands to query JDBC metadata, such as
As you can see there are 5 tables in the system: tables
SDEPTS in the current
SALES schema, and
TABLES in the system
metadata schema. The
system tables are always present in Calcite, but the other tables are
provided by the specific implementation of the schema; in this case,
SDEPTS tables are based on the
SDEPTS.csv files in the
Let’s execute some queries on those tables, to show that Calcite is providing a full implementation of SQL. First, a table scan:
Now JOIN and GROUP BY:
Last, the VALUES operator generates a single row, and is a convenient way to test expressions and SQL built-in functions:
Calcite has many other SQL features. We don’t have time to cover them here. Write some more queries to experiment.
Now, how did Calcite find these tables? Remember, core Calcite does not know anything about CSV files. (As a “database without a storage layer”, Calcite doesn’t know about any file formats.) Calcite knows about those tables because we told it to run code in the calcite-example-csv project.
There are a couple of steps in that chain. First, we define a schema based on a schema factory class in a model file. Then the schema factory creates a schema, and the schema creates several tables, each of which knows how to get data by scanning a CSV file. Last, after Calcite has parsed the query and planned it to use those tables, Calcite invokes the tables to read the data as the query is being executed. Now let’s look at those steps in more detail.
On the JDBC connect string we gave the path of a model in JSON format. Here is the model:
The model defines a single schema called ‘SALES’. The schema is
powered by a plugin class,
which is part of the
calcite-example-csv project and implements the Calcite interface
create method instantiates a
schema, passing in the
directory argument from the model file:
A schema’s job is to produce a list of tables. (It can also list sub-schemas and
table-functions, but these are advanced features and calcite-example-csv does
not support them.) The tables implement Calcite’s
CsvSchema produces tables that are instances of
and its sub-classes.
Here is the relevant code from
CsvSchema, overriding the
method in the
AbstractSchema base class.
The schema scans the directory, finds all files with the appropriate extension,
and creates tables for them. In this case, the directory
sales and contains files
SDEPTS.csv, which these become
Tables and views in schemas
Note how we did not need to define any tables in the model; the schema generated the tables automatically.
You can define extra tables,
beyond those that are created automatically,
tables property of a schema.
Let’s see how to create an important and useful type of table, namely a view.
A view looks like a table when you are writing a query, but it doesn’t store data. It derives its result by executing a query. The view is expanded while the query is being planned, so the query planner can often perform optimizations like removing expressions from the SELECT clause that are not used in the final result.
Here is a schema that defines a view:
type: 'view' tags
FEMALE_EMPS as a view,
as opposed to a regular table or a custom table.
Note that single-quotes within the view definition are escaped using a
back-slash, in the normal way for JSON.
JSON doesn’t make it easy to author long strings, so Calcite supports an alternative syntax. If your view has a long SQL statement, you can instead supply a list of lines rather than a single string:
Now we have defined a view, we can use it in queries just as if it were a table:
Custom tables are tables whose implementation is driven by user-defined code. They don’t need to live in a custom schema.
There is an example in
We can query the table in the usual way:
The schema is a regular one, and contains a custom table powered by
which implements the Calcite interface
create method instantiates a
passing in the
file argument from the model file:
Implementing a custom table is often a simpler alternative to implementing
a custom schema. Both approaches might end up creating a similar implementation
Table interface, but for the custom table you don’t
need to implement metadata discovery. (
CsvScannableTable, just as
but the table implementation does not scan the filesystem for .csv files.)
Custom tables require more work for the author of the model (the author needs to specify each table and its file explicitly) but also give the author more control (say, providing different parameters for each table).
Comments in models
Models can include comments using
/* ... */ and
(Comments are not standard JSON, but are a harmless extension.)
Optimizing queries using planner rules
The table implementations we have seen so far are fine as long as the tables don’t contain a great deal of data. But if your customer table has, say, a hundred columns and a million rows, you would rather that the system did not retrieve all of the data for every query. You would like Calcite to negotiate with the adapter and find a more efficient way of accessing the data.
This negotiation is a simple form of query optimization. Calcite supports query optimization by adding planner rules. Planner rules operate by looking for patterns in the query parse tree (for instance a project on top of a certain kind of table), and replacing the matched nodes in the tree by a new set of nodes which implement the optimization.
Planner rules are also extensible, like schemas and tables. So, if you have a data store that you want to access via SQL, you first define a custom table or schema, and then you define some rules to make the access efficient.
To see this in action, let’s use a planner rule to access a subset of columns from a CSV file. Let’s run the same query against two very similar schemas:
What causes the difference in plan? Let’s follow the trail of evidence. In the
smart.json model file, there is just one extra line:
This causes a
CsvSchema to be created with
flavor = TRANSLATABLE,
createTable method creates instances of
rather than a
CsvTranslatableTable implements the
method to create
Table scans are the leaves of a query operator tree.
The usual implementation is
but we have created a distinctive sub-type that will cause rules to fire.
Here is the rule in its entirety:
The default instance of the rule resides in the
CsvRules holder class:
The call to the
withOperandSupplier method in the default configuration
DEFAULT field in
interface Config) declares the pattern of relational
expressions that will cause the rule to fire. The planner will invoke the rule
if it sees a
LogicalProject whose sole input is a
CsvTableScan with no
Variants of the rule are possible. For example, a different rule instance
might instead match a
EnumerableProject on a
onMatch method generates a new relational expression and calls
to indicate that the rule has fired successfully.
The query optimization process
There’s a lot to say about how clever Calcite’s query planner is, but we won’t say it here. The cleverness is designed to take the burden off you, the writer of planner rules.
First, Calcite doesn’t fire rules in a prescribed order. The query optimization process follows many branches of a branching tree, just like a chess playing program examines many possible sequences of moves. If rules A and B both match a given section of the query operator tree, then Calcite can fire both.
Second, Calcite uses cost in choosing between plans, but the cost model doesn’t prevent rules from firing which may seem to be more expensive in the short term.
Many optimizers have a linear optimization scheme. Faced with a choice between rule A and rule B, as above, such an optimizer needs to choose immediately. It might have a policy such as “apply rule A to the whole tree, then apply rule B to the whole tree”, or apply a cost-based policy, applying the rule that produces the cheaper result.
Calcite doesn’t require such compromises. This makes it simple to combine various sets of rules. If, say you want to combine rules to recognize materialized views with rules to read from CSV and JDBC source systems, you just give Calcite the set of all rules and tell it to go at it.
Calcite does use a cost model. The cost model decides which plan to ultimately use, and sometimes to prune the search tree to prevent the search space from exploding, but it never forces you to choose between rule A and rule B. This is important, because it avoids falling into local minima in the search space that are not actually optimal.
Also (you guessed it) the cost model is pluggable, as are the table and query operator statistics it is based upon. But that can be a subject for later.
The JDBC adapter maps a schema in a JDBC data source as a Calcite schema.
For example, this schema reads from a MySQL “foodmart” database:
(The FoodMart database will be familiar to those of you who have used the Mondrian OLAP engine, because it is Mondrian’s main test data set. To load the data set, follow Mondrian’s installation instructions.)
Current limitations: The JDBC adapter currently only pushes down table scan operations; all other processing (filtering, joins, aggregations and so forth) occurs within Calcite. Our goal is to push down as much processing as possible to the source system, translating syntax, data types and built-in functions as we go. If a Calcite query is based on tables from a single JDBC database, in principle the whole query should go to that database. If tables are from multiple JDBC sources, or a mixture of JDBC and non-JDBC, Calcite will use the most efficient distributed query approach that it can.
The cloning JDBC adapter
The cloning JDBC adapter creates a hybrid database. The data is sourced from a JDBC database but is read into in-memory tables the first time each table is accessed. Calcite evaluates queries based on those in-memory tables, effectively a cache of the database.
For example, the following model reads tables from a MySQL “foodmart” database:
Another technique is to build a clone schema on top of an existing
schema. You use the
source property to reference a schema
defined earlier in the model, like this:
You can use this approach to create a clone schema on any type of schema, not just JDBC.
The cloning adapter isn’t the be-all and end-all. We plan to develop more sophisticated caching strategies, and a more complete and efficient implementation of in-memory tables, but for now the cloning JDBC adapter shows what is possible and allows us to try out our initial implementations.
There are many other ways to extend Calcite not yet described in this tutorial. The adapter specification describes the APIs involved.