Cassandra storage engine

This page is a documentation for Cassandra storage engine which is a work in progress.

Data mapping

MariaDB's table represents a column family in Cassandra. The table must follow this pattern:

create table cassandra_tbl               -- name can be chosen at will
(
  rowkey  char(N),                       -- First field must be named 'rowkey'. This
                                         --   is what Cassandra's row key is mapped to.

  column1    varchar(N),                 -- columns go here. Names should match
                                         -- Cassandra's column names 
  ... 
  PRIMARY KEY(rowkey)                    -- Primary key over 'rowkey' is mandatory
) engine=cassandra 
  thrift_host='192.168.1.0' 
  keyspace= 'cassandra_key_space'
  column_family='column_family_name';

MariaDB's table is a view of Cassandra's column family.

• column family must exist before the MariaDB table can be created
• dropping MariaDB's table will not drop the column family.

Mapping for columns

Cassandra has three kinds of columns

1. The row key. It is not considered to be a column by Cassandra, but MySQL/MariaDB can only store data in columns, so we consider it a special kind of column.
2. Columns that were present in static column family declaration. These columns have a defined name/data type.
3. "Ad-hoc" columns that can be encountered in individual rows.

note: Cassandra's supercolumns are not handled at the moment. If we need to handle them, we'll make them dynamic-column blob. Cassandra's counter columns will be handled, and will be mapped to regular SQL columns.

When these three kinds of columns are mapped to SQL:

1. The row key is mapped into a regular column named rowkey
2. Static column family members are mapped to regular SQL columns (i.e, column named 'foo' in Cassandra will map to column named 'foo' in SQL. If SQL's CREATE TABLE doesn't match Cassandra's columns, a warning is issued.
3. Ad-hoc columns are all put into a regular SQL column which has a Dynamic Columns blob. This allows us to return arbitrary sets of columns within one row.

Mapping for datatypes

There is no direct 1-to-1 mapping between Cassandra data types and MySQL datatypes. Also, Cassandra's limitations are generally more lax than MySQL/MariaDB's: for example, Cassandra's limit on rowkey is about 2G, while MySQL's is about 1.5Kb.

Mapping between Cassandra types and MySQL types are as follows:

SQL's NULL means that "corresponding column is not present in Cassandra column family".

(TODO: what about the scenario where we already have an SQL DDL and just want to use Cassandra for storage? Perhaps, we could allow arbitrary SQL types, which are stored as strings (or even as binary) in cassandra's blob columns?)

{TODO: Mapping for rowkey.}

Command mapping

INSERT

Cassandra doesn't provide any practical way to make INSERT different from UPDATE. Therefore, INSERT works as INSERT-or-UPDATE, it will overwrite the data, if necessary.

INSERT ... SELECT and multi-line INSERT will try to write data in batches. Batch size is controlled by @@cassandra_insert_batch_size system variable, which specifies max. batch size in records.

Status variables Cassandra_row_inserts and Cassandra_row_insert_batches allow one to see whether inserts are actually batched.

UPDATE

Currently, all UPDATE commands will fail (you can use INSERT instead) (TODO: make UPDATE work)

DELETE

• DELETE FROM cassandra_table will map to truncate(column_family)
• DELETE with WHERE clause will do regular row deletions, and, just like del command in cassandra-cli, may leave a row in a weird state where the row is present but all columns are NULL. (TODO: does this live till the nearest compaction?) (TODO: CQL will ignore rows that do not have all needed columns. Maybe, we should do so, too?)

SELECTs

Generally, all SELECT statement should like one expects in SQL. Conditions in form rowkey=.... allow the server to construct query plan with efficient lookups.

Full table scan

These are done in a memory-efficient way. Cassandra SE will split table scan into batches which retrieve @@cassandra_rnd_batch_size records at a time.

Batched Key Access support

Cassandra supports Batched Key Access in no-association mode. This means that it requires the SQL layer to do hashing, which means the following settings are required:

• optimizer_switch='join_cache_hashed=on'
• join_cache_level=7|8

Cassandra SE is unable to use parts of join buffer (size controlled by @@join_buffer_size). This buffer will be used exclusively by the SQL layer. It is also possible (and recommended) to limit size of read batches by setting @@cassandra_multiget_batch_size. It specifies max. read size in records.

It is possible to track number of batches, how many keys were looked-up, and how many results were produced with these status variables:

Variable_name	Value
Cassandra_multiget_reads	0
Cassandra_multiget_keys_scanned	0
Cassandra_multiget_rows_read	0

How to try it out

The code is at lp:maria-captains/maria/5.5-cassandra .

Build process is not fully streamlined yet. You'll need to

• Install Cassandra (I used 1.1.2, judging from changelog, 1.1.3 should work too)
• Install Thrift (I used 0.8.0), only C++ backend is needed.
• edit storage/cassandra/CMakeLists.txt and modify INCLUDE_DIRECTORIES directive to point to Thrift's include directory.
• export LIBS="-lthrift", on another machine it was "-lthrift -ldl"
• export LDFLAGS=-L/path/to/thrift/libs
• Build the server

Testing can be done as follows:

• Start Cassandra on localhost
• Look through mysql-test/t/cassandra.test see the required CQL initialization commands and run them in cqlsh -3.
• Run ./mysql-test-run t/cassandra.test

See also

• MDEV-431 - jira task for the cassandra storage engine
• cassandra-storage-engine-future-plans
• cassandra-storage-engine-use-example
• cassandra-storage-engine-issues
• hbase-storage-engine