CS 4321/5321 Project 1
This project is out of 90 points and counts for 22% of your grade.
1 Goals and important points
This is the first project where you actually develop some database functionality. The goals are:
• to teach you how to translate from SQL queries to a relational algebra query plan
• to familiarize you with the iterator model for relational operator evaluation, as well as na¨ıve implementations for the most common operators (selection, projection, join, sort)
• to provide you with a codebase that you will build on in future projects.
You will start using a skeleton codebase that we provide.
The reference implementation of this project is about 1100 lines of code, not including comments. Whether or not you consider this a lot, it is not a project that should be left to the last minute.
No solution code will be given out. The code you implement for Project 1 will be the code you reuse/refactor for Projects 2, 3, 4 and 5. Subsequent projects will assume you have implemented all the functionality required for Project 1.
2 Overview
In this project, you will implement a simple interpreter for SQL statements. That is, you will build a program that takes in a database (a set of files with data) and a file containing several SQL queries. It will process and evaluate each SQL query on the database.
2.1 Supported language features
Your interpreter will not support all of SQL, but it will handle a lot of relatively complex queries. Here we give information about the queries you must support.
Your interpreter will process SELECT-FROM-WHERE queries, which may optionally also have a DISTINCT, an ORDER BY, or both. You do not need to support nested subqueries, set operators (UNION etc.), GROUP BY, aggregates like COUNT, or any other features. In addition, we make a few simplifying assumptions as below. When we say a query is valid, we mean it is a permitted input to your interpreter which you should be able to handle. When we talk about a base table, we mean a real table that exists in the database.
• You may assume all valid queries follow correct SQL syntax and that they only refer to tables that exist in the database. Also, when a query refers to a table column such as Sailors .name, you may assume the column name is valid for that table.
• You may assume there will be at least one table in the FROM clause.
• Valid queries may use aliases (range variables) such as Sailors S or they may just use the names of base tables. If a query does not use aliases, all column references are fully qualified by the base table name. If a query does use aliases, all tables use aliases and all column references are qualified by alias. Here are two examples of valid queries, the first one without aliases and the second with aliases:
– SELECT Sailors .name, Reservations .date FROM Sailors, Reservations WHERE Sailors .id = Reservations .sid;
– SELECT S .name, R.date FROM Sailors S, Reservations R WHERE S .id = R .sid;
You may assume that any string used as an alias will not also be the name of a base table.
• Self-joins, i.e. joining a table with itself, are valid and must be supported (and require the use of aliases)
• The WHERE clause, if present, is a conjunction (i.e. an AND) of expressions of the form. A op B, where op is one f =, ! =,<,>,<=,>= and A and B are either integers or column references. Thus
Sailors .id = Reservations .sid, Sailors .id < 3 and 42 = 42 are all valid expressions for the
WHERE cause, while for example Sailors .id < Boats .id - 1 is not a valid expression even though it would be ok in “real SQL” .
• The SELECT clause will either specify a subset of columns or have the form. SELECT *. For SELECT *, order the columns in your answer following the FROM clause. Thus for SELECT * FROM R, S each answer row has all the columns of R followed by all the columns of S.
• There may be an ORDER BY clause which specifies a subset of columns for ordering. You may assume that we only want to sort in ascending order. You may also assume that the attributes mentioned in the ORDER BY are a subset of those retained by the SELECT. This allows you to do the sorting last, after projection. Note that this does not mean that every attribute in ORDER BY must be mentioned in the SELECT - a query like SELECT * FROM Sailors S ORDER BY S .name is valid.
• There may be a DISTINCT right after the SELECT, and it should be processed appropriately. Yes, SELECT DISTINCT * FROM . . . is valid.
We require a particular behavior. for ORDER BY for consistency in grading. You must order the tuples first by the attributes specified in the ORDER BY, and then by any remaining attributes in the order in which they appear in the tuples that are input to the sort. Thus suppose a relation R has attributes A, B, C, D. If processing the query SELECT * FROM R ORDER BY R.C, R.A we expect you to sort by C, A, B, D, in that order. (If two tuples agree on C, break ties based on A, if they also agree on A, break ties based on B, and so on).
2.2 Getting Started
2.2.1 Create your Private Repository
We have created a skeleton code for the class. Follow the instructions on our GitHub repository, in order to create your own private repository, and mirror the official one in order to be able to fetch future updates. You can find the repository at the following link: https://github.com/CornellDB/db practicum. Make sure you follow all the instructions provided carefully.
2.2.2 Running the Unit Tests
You can run the unit tests using the following command. For a submission to be considered as a valid one, all of the unit tests should pass. Thus, the following command should be able to run successfully:
./gradlew test
The first time you will run this command, the unit tests are expected to fail with a NotImplementedException. This is expected, as all the tests are calling the buildPlan(Statement stmt) method of the QueryPlanBuilder class, which you need to implement.
2.2.3 Code Formatting
To keep the code formatting consistent, we are using the google-java-format plugin. We have provided you the plugin already, which automates the whole process. In order to check whether your code complies, you need to run the following command:
./gradlew verGJF
If this fails, then the following command will auto-format your code and the issues will be resolved: ./gradlew goGF
2.2.4 Create the Jars
In order to create the required jar for the deliverable, simply run the following:
./gradlew build
This should create the jar under the ./build/libs directory. If build fails, there is either an issue with the formatting ( ./gradlew goGF should fix this) or some of the unit tests are failing.
2.3 Data and output formats
We have provided you some sample data and some sample queries. Take a look at the db practicum/src/test/resources/samples directory. There is an input and an expected output subdirectory. These are exactly the same inputs/outputs as the ones that we use for the unit tests provided.
In the input directory you will find a queries .sql file containing some example queries. You will also find a db subdirectory. This contains a schema .txt file specifying the schema for your database as well as a data subdirectory, where the data itself is stored.
The schema .txt file contains one line per table in the database. Every line contains several strings separated by spaces. The first string on each line is the table name and all the remaining ones are attribute (column) names, in the order in which they appear in the table.
The data subdirectory contains one file per database table, and the name of the file is the same as the name of the database table. Every file contains zero or more tuples; a tuple is a line in the file with fields (attributes) separated by commas. All attributes are integers, as they will be for the remainder of the course. Using integer attributes simplifies your job and allows you to focus on implementing “interesting” functionality rather than boilerplate code to handle different data types. Also, you do not have to handle null values in this or any other 4321 project (but you do need to handle empty relations).
In the expected output directory, you will find the expected output files for the queries we provided. File query1, for example, contains the expected output for the first query in the queries .sql file. The format for the output is the same as the format for the data.
Your SQL interpreter should read from the db directory and from the queries .sql file, and write output to a suitable output directory. We expect you to follow the same scheme for naming output files as we do, i.e., put the answer to each query in its own file and number it with the number of the query (starting at 1).
When we run your code, we will use the command line. Thus you need to submit a runnable .jar of your code. As perviously mentioned, to export a runnable .jar run the following command:
./gradlew build
This should create the .jar file under the ./build/libs directory. Your runnable .jar should run on the command line as follows:
java -jar db practicum team name 1.jar inputdir outputdir
In our testing, inputdir will have the same structure as described above, but it will contain our own test queries and data, and outputdir will be an empty directory. This means your top-level Java class has to accept inputdir and outputdir as command-line arguments and handle them appropriately.
Both inputdir and outputdir will be given as absolute paths to the appropriate directory with no final / symbol. If you are concerned about Windows vs *nix file separator issues, consider using File .separator and/or making a note in your README specifying where file separators appear in your code, so that the grader can modify those portions of your code if needed.
After we run your code, the grading harness will compare your outputdir directory to a directory containing expected outputs for each query and check if the files with the same names match. For queries without an ORDER BY, it is ok if your answer file has the answer tuples in a different order to ours; for queries with an ORDER BY, your ordering must match our ordering.
As you can imagine, it is very important for you to respect the expected input and output format. Also, make sure that if your interpreter throws an error on some query, it will still process the subsequent queries in the file and place their answers in appropriately-named files. This is already implemented in the Compiler class, but make sure your keep that functionality as is, in case you do any modifications.
We have already implemented a runnable class which contains a main method that produces the output, namely compiler.Compiler. Thus, once you implement all the required functionality, as well as the Operator buildPlan(Statement stmt) method of the Operator class, you should be able to build a runnable solution right away using the ./gradlew build command.
Avoid altering any already implemented methods unless necessary, as this maintains a consistent output format for all solutions and prevents grading issues. If you need to modify any of these methods, please consult a TA either during office hours or via Ed.
2.4 Operators and the iterator model
A key abstraction in this assignment will be the iterator model for relational operators. You will implement several operators:
• the bag relational algebra select, project and (tuple nested loop) join.
• sort and duplicate elimination operators, which are not part of the basic relational algebra you saw in 4320 but must be added to support ORDER BY and DISTINCT.
• a scan operator which is the leaf operator for any query plan. This is really a physical operator rather than something you would add to the relational algebra, but for now we will put it in the same category as the above.
The standard way to implement all relational operators is to use an iterator API. You should create an abstract class Operator, and all your operators will extend that. Certain operators may have one or two child operators. A scan operator has no children, a join has two children, and the remaining operators have one child. Your end goal is to build a query plan that is a tree of operators.
Every operator must implement the methods getNextTuple() and reset() (put these in your abstract Operator class). The idea is that once you create a relational operator, you can call getNextTuple() repeatedly to get the next tuple of the operator’s output. This is sometimes called “pulling tuples” from the operator. If the operator still has some available output, it will return the next tuple, otherwise it should return null.
The reset() method tells the operator to reset its state and start returning its output again from the beginning; that is, after calling reset() on an operator, a subsequent call to getNextTuple() will return the first tuple in that operator’s output, even though the tuple may have been returned before. This functionality is useful if you need to process an operator’s output multiple times, e.g. for scanning the inner relation multiple times during a join.
Implementing each of the above operators will involve implementing both getNextTuple() and reset(). Remember that if your operator has a child operator, the getNextTuple() of your operator can - and probably will - call getNextTuple() on the child operator and do something useful with the output it receives from the child.
A big advantage of the iterator model, and one of the reasons it is popular, is that it supports pipelined evaluation of multi-operator plans, i.e., evaluation without materializing (writing to disk) intermediate results.
The bulk of this project involves implementing each of the above six operators, as well as writing code to translate an SQL query - i.e. a line of text - to a query plan - i.e. a suitable tree of operators. Once you have the query plan, you can actually compute the answer to the query by repeatedly calling getNextTuple() on the root operator and putting the tuples somewhere as they come out.
We suggest you add a dump() method to your abstract Operator class. This method repeatedly calls getNextTuple() until the next tuple is null (no more output) and writes each tuple to a suitable
PrintStream. That way you can dump() the results of any operator - including the root of your query plan - to your favorite PrintStream, whether it leads to a file or whether it is System .out.
3 Implementation instructions
We recommend that you implement and test one feature at a time. Our instructions below are given in suggested implementation order.
We also recommend you set up a test infrastructure/harness early on. You should do two kinds of testing - unit tests for individual components and end-to-end tests where you run your interpreter on queries and look at the output files produced to see if they match a set of expected output files. As you add more features, rerun all your tests to check that you didn’t introduce bugs that affect earlier functionality.
After you implement and test each feature, make a copy of your code and save it so if you mess up later you still have a version that works (and that you can submit for partial credit if all else fails!).
3.1 Set up JSqlParser
We will use JSqlParser, which takes care of parsing your SQL and creating a Java object. The skeleton code comes with the latest version of the JSqlParser (check the build.gradle file). You need to correctly parse all valid queries as defined in Section 2.1.
The documentation is a little bare-bones but it will be sufficient for our purposes. You should expect to play around with JSqlParser on your own and read the documentation to understand the structure of the objects that it outputs.
To get you started, we have provided a sample test parserExampleTest that uses JSqlParser to read SQL queries from a file and print them out. This test illustrates the use of JSqlParser and some methods to access fields of the Statement object, such as getSelectBody() if the Statement is a Select.
You may assume all the Statements we will work with are Selects, and have a PlainSelect as the selectBody. Take a look at the PlainSelect Javadocs; the first table in the FROM clause will be in the fromItem, the remaining ones in joins. Also of relevance to you is the where field for the WHERE clause, and eventually the distinct and orderByElements fields. Write some SQL queries and write code to access and print out all these objects/fields for your queries to get an idea of “what goes where” .
The where field of a PlainSelect contains an Expression; take a look at the docs for that. For this project, you only need to worry about AndExpression, Column, LongValue, EqualsTo, NotEqualsTo,
GreaterThan, GreaterThanEquals, MinorThan and MinorThanEquals. These capture the recursive structure of an expression, which should bring back memories of Project 1. The last six of the expression types mentioned are comparison expressions, the AndExpression is a conjunction of two other Expressions, the LongValue is a numeric literal, and Column is a column reference (such as the S .id in S.id < 5). Every Column object has a column name, as well as an embedded Table object. Every Table object has a name and an alias (if aliases are used).
JSqlParser also provides a number of Visitor interfaces, which you may or may not choose to use. In particular, ExpressionVisitor is highly recommended to use one you get to the selection operator.
The above should be enough to get you started, but you should expect to do further explorations on your own as you implement more and more SQL features.
3.2 Implement scan
Your first goal is to support queries that are full table scans, e.g. SELECT * FROM Sailors (for now assume the queries do not use aliases). To achieve this, you will need to implement your first operator - the scan operator.
Implement a ScanOperator that extends your Operator abstract class. Every ScanOperator knows which base table it is scanning. Upon initialization, it opens a file scan on the appropriate data file; when getNextTuple() is called, it reads the next line from the file and returns the next tuple. You probably want to have a Tuple class to handle the tuples as objects.
The ScanOperator needs to know where to find the data file for its table. It is recommended to handle this by implementing a database catalog in a separate class. The catalog can keep track of information such as where a file for a given table is located, what the schema of different tables is, and so on. Because the catalog is a global entity that various components of your system may want to access, you should consider using the singleton pattern for the catalog (see Wikpedia and many other online references on this).
Once you have written your ScanOperator, test it thoroughly to be sure getNextTuple() and reset() both work as expected. Then, hook up your ScanOperator to your interpreter. Assuming that all your queries are of the from SELECT * FROM MyTable, write code that grabs MyTable from the fromItem and constructs a ScanOperator from it.
In summary the top-level structure of your code at this point should be:
• while more queries remain, parse the next query in the queries file
• construct a ScanOperator for the table in the fromItem
• call dump() on your ScanOperator to send the results somewhere helpful, like a file or your console.
3.3 Implement selection
The next order of business is single-table selection, still with fully specified table names (no aliases). That is, you are aiming to support queries like SELECT * FROM Boats WHERE Boats .id = 4.
This means you need to implement a second Operator, which is a SelectOperator. Your query plan will now have two operators - the SelectOperator as the root and the ScanOperator as its child. During evaluation, the SelectOperator’s getNextTuple() method will grab the next tuple from its child (i.e.
from the scan), check if that tuple passes the selection condition, and if so output it. If the tuple doesn’t pass the selection condition, the selection operator will continue pulling tuples from the scan until either it finds one that passes or it receives null (i.e. the scan runs out of output).
The tricky part will be implementing the logic to check if a tuple passes the selection condition. The selection condition is an Expression which you will find in the WHERE clause of your query. The
SelectOperator needs to know that Expression.
You will need to write a class to test whether a Expression holds on a given tuple. For example, if you
have a table R with fields A, B and C, you may encounter a tuple 1, 42, 4 and an expression R.A < R.C AND R.B = 42, and you need to determine whether the expression is true or false on this tuple.
This is best achieved using a visitor on the Expression. You should have a class that implements the
JSqlParser ExpressionVisitor. The class will take as input a tuple and recursively walk the expression to evaluate it to true or false on that tuple. The expression may contain column references - in our example R.A < R.C AND R.B = 42 refers to all three columns of R. The visitor class needs some way to resolve the references; i.e., if our input tuple is 1 , 42, 4, it needs a way to determine that R.A is 1, etc. So, your visitor class also needs to take in some schema information. It is up to you how you structure your schema
information, but obviously it must allow mapping from column references like R.A to indexes into the tuple.
Once you have written your visitor class, unit-test it thoroughly. Start with simple expressions that have
no column references, like 1 < 2 AND 3 = 17. Then test it with column references until you are 100% sure it works. Once your expression evaluation logic is solid, you can plug it into the getNextTuple() method of your SelectOperator.
3.4 Implement projection
Your next task is to implement projection, i.e. you will be able to handle queries of the form. SELECT
Sailors .id FROM Sailors WHERE Sailors .age = 20. We still assume that the queries do not use aliases.
In comparison with selection, implementing projection is relatively easy. You need a third Operator that is a ProjectOperator. When getNextTuple() is called, it grabs the next tuple from its child. It extracts
only the desired attributes, makes them into a new tuple and returns that. Note that the child could be either a SelectOperator or a ScanOperator, depending on whether your SQL query has a WHERE clause.
You get the projection columns from the selectItems field of your PlainSelect. selectItems is a list of SelectItems, where each one is either AllColumns (for a SELECT * ) or a SelectExpressionItem. You
may assume the Expression in a SelectExpressionItem will always be a Column. Once you grab these Columns you need to translate that information into something useful to the ProjectOperator.
Note that the attribute order in the SELECT does not have to match the attribute order in the table. The queries SELECT R.A, R.B FROM R and SELECT R.B, R.A FROM R are both valid, and they are different queries with different output.
By this point you should have code takes in a SQL query and produces a query plan containing:
• an optional projection operator, having as a child
• an optional selection operator, having as a child
• a non-optional scan operator.
Thus your query plan could have one, two or three operators. Make sure you are supporting all
possibilities; try queries with/without a projection/selection. If the query is SELECT *, do not create a projection operator, and if the query has no WHERE clause, do not create a selection operator.
You are now producing relatively complex query plans; however, things are about to get much more
exciting and messy as we add joins. This is a good time to pull out the logic for constructing the query plan into its own class, if you haven’t done so already. Thus, you should have a top-level
interpreter/harness class that reads the statements from the queries file. You should also have a second
class that knows how to construct a query plan for each Statement, and returns the query plan back to the interpreter so the interpreter can dump() the results of the query plan somewhere.