Skip to content

Serverless SQL database stored in a single file

Notifications You must be signed in to change notification settings

LucasBN/sqlite-clone

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

60 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

SQLite3 Clone - A Serverless Single File Database

A simple SQL database engine written in Go, which reads and writes data on a single file in the SQLite3 file format.

You can run the database engine with the following command:

./sqlite sample.db

This project is in it's early stages of development, which means that it does not yet support many features.

The section on architecture below will go into more depth about the various layers that make up the engine, but at present there are only implementations for the Machine, BTreeEngine and Pager layers.

This means that you cannot actually pass arbitrary SQL - the Parser and Generator layers are mocked and return hard coded instructions.

Architecture

Birds Eye View

Birds Eye View

The image above is a visual depiction of the DatabaseEngine, which is responsible for glueing together independent layers and providing an interface for executing SQL statements and meta commands (e.g .dbinfo).

There are 5 layers: Pager, BTreeEngine, Machine, Generator and Parser. Each of these layers is designed to be completely independent of other layers. A layer is responsible for defining the interface by which it expects to make calls to another layer.

This means that layers are decoupled, and don't really on the implementation details of any other layers. The reasoning for this is so that layers can be independently tested and so that the entire functionality of a layer resides in the same package (although they are allowed to have sub-packages) which helps to create a 'separation of concerns'.

In the future, I'd like to potentially create different implementations of some layers and figure out a way to profile them (i.e compare CPU and memory usage).

There is a global package for types which contains an Entry which is a union of a null value, a string and a int. These are the types of values that can be stored in the database, which is a detail that only the BTreeEngine really needs to be concerned with.

Therefore, we can initialise the Machine with a generic type T which represents the 'output' type of the machine - the Machine can take any type here. However, the BTreeEngine takes a generic type T that conforms to the following interface:

type resultTypeConstructor[T any] interface {
    Number(uint64) T
    Text(string) T
    Null() T
}

This is the BTreeEngine's way of saying 'I need you to provide me a type that knows how to construct a T of all of the values that I store'.

The intention here is to ensure that all layers are independent entities that don't rely on external packages - they are valid packages in isolation that fulfill their single responsiblity. I'm not confident that this is the right thing to do here, and I'm open for criticism/alternatives.

I can think of some argument against this which is like: "Entry is basically a primitive type like int or string and so it is completely acceptable to import it everywhere and have packages depend on it.

The "each layer needs to be independent and not rely on external packages" goal / p attern is not so that I can publish these as separate packages that other developers can rely on - it's to help make the code more readable and maintainable for myself and anyone else that works on this project. This makes me wonder if the resultTypeConstructor actually helps with readability/maintainability, or if I'm just doing it to remove any external dependencies for packages. I have a feeling it might be the latter and I'll want to reconsider this change

Layer 1: Pager

The pager is responsible for efficiently reading and writing pages to/from a file, and is not concerned with things like:

  1. The contents of a page
  2. The format of the file
  3. The 'context' of the file (e.g a database)

It could be used as a means for interacting with any page-based file, and conforms to the following interface:

type Pager interface {
    Close() error
    PageSize() uint64
    ReservedSpace() uint64
    GetPage(pageNum uint64) ([]byte, error)
}

It can be instantied with the following constructor:

func NewPager(filepath string, config PagerConfig) (*Pager, error)

where the config contains the page size and the number of reserved bytes at the end of each page.

The pager has a cache, which is implemented as a hash map that relates page numbers to a byte slice (which is the length of a page).

There's likely a huge number of optimisations that could be done here, and I'm deliberately doing none of them now (other than the very simple cache described above) because my primary aim is to get an end-to-end simple, extensible and functional system.

Currently, the pager takes a filepath and opens the file itself. I'd like to spend some time thinking about what consequences this has on testing - would it be better to pass in a file that is opened elsewhere so that I could potentially mock a file in unit tests?

Layer 2: BTreeEngine

The BTreeEngine is responsible for providing an interface for callers to interact with B-Tree structures encoded by the database file.

type BTreeEngine[T any] interface {
    NewCursor(id uint64, rootPageNum uint64) (bool, error)
    RewindCursor(id uint64) (bool, error)
    AdvanceCursor(id uint64) (bool, error)
    ReadColumn(id uint64, column uint64) (T, error)
}

This is done through 'cursors' which are pointers to a particular entry in a B-Tree.

When a cursor is initialised with NewCursor() it points to the first byte on the root page, and is set to point at the first B-Tree entry via the RewindCursor() function. AdvanceCursor() moves the cursor along to the next entry, and ReadColumn() reads a particular column in the entry that is currently being pointed to.

This is the only layer that knows about the SQLite3 file format, which means that this is the only layer that would need to change if the file format were to change.

It can be instantied with the following constructor:

func NewBTreeEngine[T any](pager pager, resultConstructor resultTypeConstructor[T]) (*BTreeEngine[T], error)

The resultTypeConstructor has been discussed in detail above, but it is something which tells the BTreeEngine how to construct objects of type T out of (in this case) ints strings or nulls - the BTreeEngine is responsible for specifying which primitive types it stores and hence needs to create Ts out of.

Layer 3: Machine

The Machine is a simple virtual machine that executes a set of bytecode instructions. The bytecode instruction set is defined in the instructions package, and each instruction must conform to the following interface:

type Instruction[T any] interface {
	Execute(s *state.MachineState[T], b common.BTreeEngine[T]) [][]T
}

Arguments are defined directly on implementations of this interface, and are used to update the machine state s and are able to access the database via b which is the BTreeEngine.

The return type is [][]T which means an instruction can return multiple tuples of type T.

A machine is instantiated with the following constructor:

func NewMachine[T any](config MachineConfig[T]) *Machine[T]

The config contains a list of instructions and a pointer to a BTreeEngine. The machine then initialises its state which stores the current address, the registers and whether or not the machine is halted.

When Run is called on a machine, it will pick the instruction at the current address and call Execute on it. This will happen continuously in a loop until the machine is halted (by setting halted to true on the state).

A register is a map from an int (register number) to a T which means that a register stores individual entries from a tuple.

Layer 4: Bytecode Generator

TODO

Layer 5: Parser

TODO

Useful links

About

Serverless SQL database stored in a single file

Topics

Resources

Stars

Watchers

Forks