Replicating multiple masters to one slave

As standard, MySQL allows replication from one master to multiple slaves, and that is a common scale-out scenario, but there have been a few comments recently, and some longer standing queries about having a setup that works the other way round, that is, multiple slaves replicating into a single master. This a common enough scenario in data logging systems, where the data is collected locally and then distributed up to a central database, or in EPOS (Electronic Point of Sale) systems where you want the transactions logs from the tills logged up to the database at head office. There are many other situations where you want that merging of information. Although MySQL doesn’t support what is called ‘multiple master, single slave’ solution, you can simulate the general approach by using a combination of replication and federated tables. Replication allows for different table types on the master (the source of the data) and the slave. There are many advantages to this, for example, using InnoDB on the master to take advantage of transactions, while using MyISAM on the slave for read performance. Federation allows you to access the tables of a remote server as if it were a local table. You can set up a federated table to access a remote table from as many machines as you like. That means that you can have two, or more, MySQL instances set up to use the remote table using the federated engine. You can execute any queries you like on the remote table, but you need to take care when using multiple hosts to access the remote table. Particularly when doing INSERT from multiple hosts, using InnoDB, Falcon, Maria or another table that supports multiple writers can be a good idea, although I’ll cover some workarounds for that later. Using federated gives us the ability to write to the same table from multiple hosts, but you dont want to read and write from the same remote table all the time, especially if on your local machine (your till, or data collector) you want to be able to run your own queries. This is where the replication fits in, if you set up replication from the master to another instance of MySQL, let’s call it ‘Fed Slave’ (which works both ways). On the Fed Slave, you configure the table or tables that you want to merge on the final ‘Slave’ machine to be federated tables. What happens is that data is replicated from the master to the Fed Slave, and on Fed Slave the queries are sent to the Merge Slave via federation. You can probably see this more clearly in the figure below.


To re-iterate:

  1. INSERT on Master 1 is replicated to Fed Slave 1
  2. Fed Slave 1 executes the INSERT on a Federated table which points to Merge Slave
  3. Merge Slave executes the federated statement on its local table

Each Fed Slave is relatively lightweight – all it’s doing is executing a statement and sending the statement over the network to the Merge Slave, so you could run it on the same machine as Master 1. There are few problems with this design:

  1. Updating the same federated table from multiple hosts can get messy. There are a few ways you can get get round this, one is to stop the query execution on the slaves and only allow them to run during a set period of time. For example, let Fed Slave 1 execute the queries in the log from 1am to 2am, and Fed Slave 2 from 2am to 3am, and so on.
  2. Federation doesn’t get round the problems of duplicate IDs – if you try to run a statement on a federated table that inserts a duplicate ID it will fail just as will locally. You can get round this by making sure that the tables that hold the merge data on your Merge Slave dont have unique ID constraints, and that your Masters and all the table definitions contain a field to identify the source of the data in each case.
  3. Load can be an issue. One of the reasons I suggested InnoDB/Falcon/Maria is to help get round the multiple-insert and locking that is normally applied, but the very nature of the system means that locks and delays might still occur. You can’t eliminate it, but you can ease it.

I’ve tried and used this method in a number of situations, actually not for the reasons given above, but for performance logging from multiple hosts onto one. I’ll be honest and say that I’ve never seen a problem, but, at the same time, the type of data that I am collecting means that I would have been unlikely to notice a missing data point or two.