couchdb-cluster-admin

utility for managing multi-node couchdb 2.x clusters

First, put together a config file for your setup

This will make the rest of the commands simpler to run. Copy the example

cp config/conf.example.yml config/mycluster.yml

and then edit it with the details of your cluster.

Setting up a local cluster to test on

If you have docker installed you can just run

docker build -t couchdb-cluster - < docker-couchdb-cluster/Dockerfile

to build the cluster image (based on klaemo/couchdb:2.0-dev) and then run

docker run --name couchdb-cluster \
  -p 15984:15984 \
  -p 15986:15986 \
  -p 25984:25984 \
  -p 25986:25986 \
  -p 35984:35984 \
  -p 35986:35986 \
  -p 45984:45984 \
  -p 45986:45986 \
  -v $(pwd)/data:/usr/src/couchdb/dev/lib/ \
  -t couchdb-cluster \
  --with-admin-party-please \
  -n 4

to start a cluster with 4 nodes. The nodes' data will be persisted to ./data.

To run the tests (which require this docker setup), download and install https://github.com/sstephenson/bats

git clone https://github.com/sstephenson/bats.git
cd bats
./install.sh /usr/local  # or wherever on your PATH you want to install this

and then

docker start couchdb-cluster  # make sure this is running and localhost:15984 is receiving pings
bats test/

Optional: Set password in environment

If you do not wish to specify your password every time you run a command, you may put its value in the COUCHDB_CLUSTER_ADMIN_PASSWORD environment variable like so:

read -sp Password: PW

Then, for all commands below prefex the command with COUCHDB_CLUSTER_ADMIN_PASSWORD=$PW, e.g.

COUCHDB_CLUSTER_ADMIN_PASSWORD=$PW python couchdb-admin-cluster/describe.py --conf mycluster.yml

Get a quick overview of your cluster

Now you can run

python couchdb_cluster_admin/describe.py --conf config/mycluster.yml

to see an overview of your cluster nodes and shard allocation. For example, in the following output:

Membership
	cluster_nodes:	couch3	couch1	couch4	couch2
	all_nodes:	couch3	couch1	couch4	couch2
Shards
	                   00000000-1fffffff  20000000-3fffffff  40000000-5fffffff  60000000-7fffffff  80000000-9fffffff  a0000000-bfffffff  c0000000-dfffffff  e0000000-ffffffff
	mydb                    couch1             couch1             couch1             couch1             couch1             couch1             couch1             couch1
	my_second_database      couch1             couch1             couch1             couch1             couch1             couch1             couch1             couch1

you can see that while there are four nodes, all shards are currently assigned only to the first node.

Help estimating shard allocation

In order to plan out a shard reallocation, you can run the following command:

python couchdb_cluster_admin/suggest_shard_allocation.py --conf config/mycluster.yml --allocate couch1:1 couch2,couch3,couch4:2

The values for the --allocate arg in the example above should be interpreted as "Put 1 copy on couch1, and put 2 copies spread across couch2, couch3, and couch4".

The output looks like this:

couch1	57.57 GB
couch2	42.15 GB
couch3	36.5 GB
couch4	36.5 GB
                     00000000-1fffffff     20000000-3fffffff     40000000-5fffffff     60000000-7fffffff     80000000-9fffffff     a0000000-bfffffff     c0000000-dfffffff     e0000000-ffffffff
mydb                couch1,couch2,couch4  couch1,couch2,couch3  couch1,couch3,couch4  couch1,couch2,couch4  couch1,couch2,couch3  couch1,couch3,couch4  couch1,couch2,couch4  couch1,couch2,couch3
my_second_database  couch1,couch3,couch4  couch1,couch3,couch4  couch1,couch3,couch4  couch1,couch3,couch4  couch1,couch3,couch4  couch1,couch3,couch4  couch1,couch3,couch4  couch1,couch3,couch4

Note, the reallocation does not take into account the current location of shards, so it is much more useful in the situation that you're moving from a single-node cluster to a multi-node cluster than it is in the situation where you're adding one more node to a multi-node cluster. In the example above, couch1 would be the single-node cluster and couch2, couch3, and couch4 form are the multi-node cluster–to-be. You can imagine that after implementing the shard allocation suggested here, we might remove all shards from couch1 and remove it from the cluster.

Note also that there is no guarantee that the "same" shard of different databases will go to the same node; each (db, shard)-pair is treated as an independent unit when making computing an even shard allocation. In this example there are only a few dbs and shards; when shards * dbs is high, this process can be quite good at evenly balancing your data across nodes.