Skip to content

MariaDB storage engine based on cap'n proto storage

License

Notifications You must be signed in to change notification settings

Shelnutt2/crunch

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Crunch

Build Status

MariaDB storage engine based on cap'n proto storage

Requirements

Requires MariaDB 10.2 or newer. It is untested on older versions.

Installation

There are two methods to building crunch.

Inside MariaDB Source Tree (Recommended)

The first is inside a MariaDB source tree (recommended).

git clone [email protected]:MariaDB/server.git
cd server
git submodule add [email protected]:Shelnutt2/crunch.git storage/crunch
mkdir build && cd build
cmake ..
make -j4

Standalone compilation

The second method is building it standalone and installing the library to the mariadb plugin folder. This method is not recommended, as the compiler and compiler flags must match for mariadb to load the plugin.

git clone [email protected]:Shelnutt2/crunch.git
cd crunch
mkdir build && cd build
cmake .. -DCRUNCH_COMPILE_STANDALONE=ON
make -j4

Features

Inplace Alter Table

In place alter tables are supported for column renames, column additions and some column data type changes. Adding of columns is supported at any position (LAST or AFTER supported).

When altering tables, the on disk data format is not changed. A new schema is produced and when reading data it is converted on read to the new schema.

If a default value is set for a column and the column is also set to null then the default value is not used, but the column will return null for existing data. This is a limitation that will be addressed in the future.

For column data type changes, if the underlying type is similar, int16->int32 or float64 -> float32, the change can be made online. There is no protection against decreased percision or overflow. Changing from a int64 to a int16 is allowed, but could result in data loss if any values are larger than an int16. This is how mariadb normally works, so we've allowed it with online alters.

Architecture

On Disk Format

The ondisk format is based on cap'n proto. Each table gets a cap'n proto schema file which represent a row in the table. See examples/t1.capnp for a sample cap'n proto schema representing the following table:

CREATE TABLE t1 (
  column1 integer,
  column2 varchar(64)
) ENGINE=crunch;

On Disk File Hierarchy

Below is a hierarchy of the ondisk structure, assuming the database is test and table is t1.

mysql_datadir
└── test
    ├── t1
    │   ├── 1515252170775990244-e8e7a69e-923d-471f-b0ca-8e54568a3fef.capnpd
    │   ├── t1.capnp
    │   ├── t1.capnpd
    │   ├── t1.deleted.capnpd
    │   └── transactions
    │       ├── 1515251992233237213-b0929f33-a67c-4d7e-83e1-594b983ce299.capnpd
    │       └── 1515251992233237213-b0929f33-a67c-4d7e-83e1-594b983ce299.deleted.capnpd
    └── t1.frm

Transactions

Transactions are supported and handled with the crunchTxn class. Each transaction will perform it's operations in a "transactions" folder, with dedicated files per table/transactions. This allows for ondisk isolation for of transactions before commit. It also removes the immediate need for an undo log.

A rollback simply deletes the transaction files and its done. When a commit is made, the transaction files are renamed and moved to the main table folder. The rename function call is atomic according to the ISO c standard assuming the transaction folder lies on the same filesystem as the main table folder.

The transaction folder is a subdirectory in the table directory. The transaction files are in the format of unix_timestamp_nanoseconds-uuid.ext

The use of unix timestamp is to order the files on disk so full table scans can sequentially read all files and keep the ordering the same as original inserts.

A consequence of using independent files for each transaction is that the number of data files will grow without bounds. During a full table scan each data file must be opened and mmaped one at a time. Performance degrades in a linear manor. To combat this a consolidation of datafiles has been implemented. Currently this only acts on table close. The long term goal is to implement a daemon plugin which will implement consolidation in the background independent of the mysql server. See #44.

Consolidation can also be manually forced by running OPTIMIZE TABLE tbl_name

System and Table Variables

Below are a list of server and table configuration variables.

Variable Bame description Data Type Default Value Range System Variable Table Variable
consolidation_threshold Threshold for number of data files to start consolidation Integer 100 0 to MAX_INT Yes Yes

Releases

No releases published

Packages

No packages published

Languages