Skip to content

The CONNECT Storage engine implement the concept of a table made of multiple tables. These underlying tables can be distributed remotely. For example the underlying remote tables can be of ODBC or MySQL table type. this allows to execute distributed queries. What is nice is that we can execute this distributed query with parallelism.

How does it work ?

To explain it let us suppose we have 4 nodes : Node0 and Node1 Node2 Node3
Node0 has a MariaDB 10.0.4 installed with the CONNECT storage engine activated.

MariaDB [dbt3]> install plugin connect  soname 'ha_connect.so';

On the 3 other nodes we have MariaDB or plain MySQL installed.
the 'lineitem' table of the dbt3 benchmark has been created and loaded with one chunk of the 'lineitem' table on each node. This is a sharded table. For example to generate the second chunk of 3 chunks 'lineitem' table:

./dbgen -T L -fF -q -b dists.dss -s 1 -C 3 -S 2

Now on node0 we can create a CONNECT table. This table definition embed a SRCDEF parameter that force the computation of an aggregate on the remote server :

create Table lineitem1   ENGINE=CONNECT TABLE_TYPE=MYSQL
SRCDEF='select l_suppkey, sum(l_quantity) qt from dbt3.lineitem3 group by l_suppkey'
connection='mysql://proxy:manager1@node1:3306/dbt3/lineitem3';

if we do a select on this table on node0 we get the result data that has been aggregated on node1. We access only one shard of the data.

MariaDB [dbt3]> select * from lineitem1;
+-----------+------+
| l_suppkey | qt   |
+-----------+------+
|         1 | 2302 |
|         2 | 1960 |
...
|      9999 | 1908 |
|      1000 | 1756 |
+-----------+------+
10000 rows in set (26.82 sec)

Let us now do the same on for node2 and node3 :

create Table lineitem2   ENGINE=CONNECT TABLE_TYPE=MYSQL
SRCDEF='select l_suppkey, sum(l_quantity) qt from dbt3.lineitem3 group by l_suppkey'
connection='mysql://proxy:manager1@node2:3306/dbt3/lineitem3';
 
create Table lineitem3   ENGINE=CONNECT TABLE_TYPE=MYSQL
SRCDEF='select l_suppkey, sum(l_quantity) qt from dbt3.lineitem3 group by l_suppkey'
connection='mysql://proxy:manager1@node3:3306/dbt3/lineitem3';

Now we can use the ability of the the CONNECT storage engine to build a table that allows to see the 3 shards as a single table :

create Table alllineitem   (`l_suppkey` INT(4) NOT NULL,`qt` DOUBLE ) ENGINE=CONNECT
TABLE_TYPE=TBL
table_list='lineitem1,lineitem2,lineitem3';

We can now query this table. All the aggregation computation will be done on the remote nodes and not on node0:

MariaDB [dbt3]> select * from alllineitem;
+-----------+------+
| l_suppkey | qt   |
+-----------+------+
|         1 | 2302 |
|         2 | 1960 |
...
|      9998 | 2395 |
|      9999 | 1908 |
|      1000 | 1756 |
+-----------+------+
30000 rows in set (1 min 19.23 sec)

This basically takes 3 times longer than querying a single shard. This is because the request is executed sequentially on node 1 2 and 3.

We can request the CONNECT storage engine to do the job in parallel by adding the extra option : option_list='thread=1';

create Table alllineitem2   (`l_suppkey` INT(4) NOT NULL,`qt` DOUBLE ) ENGINE=CONNECT
TABLE_TYPE=TBL
table_list='lineitem1,lineitem2,lineitem3' option_list='thread=1';

If we rerun the query we get

MariaDB [dbt3]> select * from alllineitem2;
+-----------+------+
| l_suppkey | qt   |
+-----------+------+
|         1 | 2302 |
|         2 | 1960 |
...
|      9999 | 1908 |
|      1000 | 1756 |
+-----------+------+
30000 rows in set (26.69 sec)

We go back to the same execution time as with a single shard ! Is not that cool. I am sure you have a lot of objections, ideas about this approach. I do. But it is nice anyways.

If you want to know more do not forget to attend to the MariaDB CONNECT Storage Engine webinar :
November 7, 2013 - 5 pm CET / 4 PM UTC / 8 AM PST Register

and if you want to give it a try you are welcome and the doc to help is here:
MariaDB Connect Storage Engine documentation
Your help is welcome : bug reports, documentation fixes, usage feedback, suggestions.
Be indulgent it is still Alpha software 😉

MySQL Fabric is a very promising sharding framework. If I take Ulf Wendel definition of MySQL Fabric :

MySQL Fabric is an administration tool to build large “farms” of MySQL servers. In its most basic form, a farm is a collection of MySQL Replication clusters. In its most advanced form, a farm is a collection of MySQL Replication clusters with sharding on top.

So MySQL Fabric takes care of two very orthogonal features :

  • High availability of servers
  • Sharding of data

Let us forget about sharding and look at the High availability infrastructure.

Servers are included in groups, called "High Availability Groups" when we talk about HA.
Each Server has an associated Status (or Role): primary secondary, spare
Each Server has also a mode : Offline, Read-only, and Read-Write.
The implementation has been made to allow various HA implementation patterns.
The most common HA pattern is the Master/Slave HA group ( in that case we should call it a "replica set" which is the terminology used in MongoDB or Facebook MySQL Pool Scanner (MPS).

Mats Kindahl in his blog post on MySQL Fabric High Availability Groups mentioned that other HA solutions are possible for an availability group :

  • Shared Storage with SAN or NAS
  • Replicated storage like DRBD
  • MySQL Cluster shared nothing cluster

In the case of a HA group based on MySQL Cluster the group is self-managing regarding HA and MySQL Fabric does not handles the failover. With the "Shared Storage" and "Replicated storage" availability groups the secondary servers will be offline.

So one of my ideas that I hope is feasible would be to use MariaDB Galera Cluster as another HA solution with MySQL Fabric. The main advantage of this solution relates to the characteristics of MariaDB Galera Cluster. MariaDB Galera Cluster is an Active-active multi-master topology with synchronous replication. MariaDB Galera Cluster being innoDB based does not carry all the usage limitations associated with MySQL Cluster (main one being limited join capabilities).

Regarding to MySQL fabric the behavior of an availability group based on MariaDB Galera Cluster is identical to MySQL Cluster. It is a self-managing availability group.

MariaDB Galera Cluster

MariaDB Galera Cluster
Getting Started with MariaDB Galera Cluster

MySQL Fabric

MySQL Fabric: A new kid in the MySQL sharding world  2013-10-09 Serge Frezefond
MySQL Fabric: High Availability Groups  2013-10-21 Mats Kindahl
A Brief Introduction to MySQL Fabric  2013-09-21 Mats Kindahl
MySQL Fabric - Sharding - Introduction  2013-09-21 VN (Narayanan Venkateswaran)
MySQL Fabric - Sharding - Simple Example  2013-09-22 VN (Narayanan Venkateswaran)
MySQL Fabric - Sharding - Shard Maintenance  2013-09-27 VN
MySQL Fabric - Sharding - Migrating From an Unsharded to a Sharded Setup  2013-09-22 VN
Installing MySQL Fabric on Windows  2013-10-03 Todd Farmer
MySQL 5.7 Fabric: any good?  2013-09-23 Ulf Wendel

Writing a Fault-tolerant Database Application using MySQL Fabric  2013-09-21 Alfranio Junior
Sharding PHP with MySQL Fabric  2013-10-09 Johannes Schlüter
MySQL Fabric support in Connector/Python  2013-09-22 Geert Vanderkelen
MySQL Connector/J with Fabric Support  2013-09-21 Jess Balint

As alway with new technology there is always different approaches regarding the adoption.
You can try to use the bleeding edge features or start with a very standard configuration.
My personal advise to new users is to start with the most basic configuration.

This allow you to get familiar with the fundamentals :
- how to install
- how to operate
- how to monitor

For MariaDB Galera Cluster he most basic configuration is a 3 nodes cluster. You can chose to use it :
as an HA solution
Galera Cluster is currently the easiest way to solve the HA problem.
when you think HA think Galera Cluster it is so much simpler. Failover is totally transparent and you have nothing to do like you would have with standard replication. if you have a lodbalancer in front you just have to push out from the configuration THE FAILED NODE and that is done.

as a scale out solution
A usual MySQL scale out architecture is based on master/slaves architecture. This solution incurs to the application the choice of where to send the read (Master or most up to date slave…). Scale out of read with Galera cluster synchronous replication is much simpler. Nothing need to be done at the application layer.
You have a synchronous data up to data available on all nodes . You do not have the risk to read stale data when the replication lags. Nothing need to be done at the application level like taking care of reading were you write to have correct data.

Contrary to usual HA solutions or compare to MySQL cluster Galera Cluster is very simple to setup and operate.
Getting Started with MariaDB Galera Cluster

Then of course next step is to push Galera usage a little bit further. One main area is of course write scalability.
A few benchmark have been produced but we do not have yet much experience and  you have to be more careful. With write/write configurations you have to be careful about hot spots in the database. This can lead to deadlock and the behavior has to be correctly understood.

A good distribution for download to do your testing is : MariaDB Galera Cluster 5.5.29 Stable

Some useful pointers to understand various behaviours of Galera Cluster.

 

------- upcoming events :

Tomorrow Seppo Jaakola, Codership, will present Galera Cluster for MySQL & MariaDB at the Meetup SkySQL & MariaDB - Paris

Henrik Ingo, Max Mether and Colin Charles will present "MariaDB Galera Cluster Overview" at the free MySQL & Cloud Solutions Day taking place in Santa Clara the 26th of April. You can register for free.

 

Following my discussion with Mikael Ronstrom regarding Parallel databases through comments on a old blog post ( mysql-for-a-massively-parallel-database/#comments ) I want to clarify why i like the Shard-Query tool.

Shard-Query is a open source distributed parallel query engine for MySQL. It offers a way to answer the question of parallel query execution across shards http://code.google.com/p/shard-query/.

But first what about MySQL Cluster as a parallel database.
MySQL Cluster has effectively a very efficient MPP architecture. It is a remarkable product for massive write and primary key lookups. Its fast failover, thanks to it true shared nothing architecture is quite unique. With respect to these two points it is much better that other architectures. Oracle RAC with its shared disk architecture, distributed cache and distributed Lock manager is much less efficient on both points.

Bust as usual you have the default associated with your qualities. Regarding complex read queries MySQL cluster has always been weak.I agree that some great progress have been with condition pushdown and Adaptive Query localization that allow to push joins and filtering to the data nodes. Both mechanism alleviate the sql node from doing work (filtering, joins) and also reduce network bandwidth consumption.

But this is far from what do parallel databases like Greenplum or Microsoft Parallel Datawarehouse Solution. To have a truly parallel database you really need to analyze the query and split it in pieces. There is a lot of intelligence to be put at the SQL node to split the query plan into subtasks. You sometime even have to implement a multi path approach(a map reduce pattern) to fully execute the request. In that case the sql node need to implement a lot of intelligence regarding parallel query execution. The MySQL Cluster sql node does not currently do it. If this work was ever to happen in the SQL layer it would be great to have it work with various storage engines. it should be abstracted from any particular storage engine like NDB.

Shard-Query presented by Justin Swanhart at FOSDEM 2013 (slides) does it at the client layer. It is using php and Gearman to spread the work in parallel to various MySQL shards. This is not perfect but this is a nice approach that can greatly help in many cases. It can parallelize across shards or using MySQL partitioning. Another interesting point about Shard-Query is that it can also works with MySQL compliant column store.

I would prefer to have this intelligence implemented in the MySQL server with full guaranty of correctly supporting the client/server protocol but this is a much harder approach.