Removing a Shard

The removeshard command will remove a shard from an existing cluster. It has two phases which are described below.

Starting

Before a shard can be removed, we have to make sure that all the chunks and databases that once lived there were relocated to other shards. The 'removeshard' command takes care of "draining" the chunks out of a shard for us. To start the shard removal, you can  issue the command

The balancer must be running for this process to work. It does all the work of migrating chunks and removing the shard when that is done.

First run

> db.runCommand( { removeshard : "shard0000" } );
{ msg : "draining started successfully" , state: "started" , shard :"shard0000" , ok : 1

That will put the shard in "draining mode". Its chunks are going to be moved away slowly over time, so not to cause any disturbance to a running system. The command will return right away but the draining task will continue on the background. If you issue the command again during it, you'll get a progress report instead:

> db.runCommand( { removeshard : "shard0000" } );
{ msg: "draining ongoing" ,  state: "ongoing" , remaining : { chunks :23 , dbs : 1 }, ok : 1 }

Whereas the chunks will be removed automatically from that shard, the databases hosted there -- the 'dbs' counter attribute in the above output -- will need to be moved manually. (This has to do with a current limitation that will go away eventually). If you need to figure out which database the removeshard output refers to, you can use the printShardingStatus command. It will tell you what is the "primary" shard for each non-partitioned database. In versions 2.1.0 and higher, this is unnecessary because it will list the dbs in a "dbsToMove" field. You need to remove these with the following command:

Move the primary somewhere else

This is only necessary to do if there are databases listed by printShardingStatus whose primary shard is the shard which has been drained.  MovePrimary should only be run after the shard has finished draining (i.e. all chunks have been removed, the status may not update), and only for those databases with primaries on the drained shard.

> db.runCommand( { movePrimary : "test", to : "shard0001" } );
{ "primary" : "shard0001", "ok" : 1 }

When the shard is empty, you could issue the 'removeshard' command again and that will clean up all metadata information:

Important: You should only issue the movePrimary command after draining has completed - in general you should never use movePrimary if you still have undrained sharded collection data on the primary shard.

Second run

> db.runCommand( { removeshard : "shard0000" } );
{ msg: "remove shard completed succesfully" , stage: "completed", host: "shard0000", ok : 1 }

After the 'removeshard' command reported being done with that shard, you can take that shard down.