Clustering Design

This is archived documentation for InfluxData product versions that are no longer maintained. For newer documentation, see the latest InfluxData documentation.

The current clustering implementation is experimental and we’re completely reworking it for the next major release (either 0.9 or 0.10). However, there are a few things that will remain the same in the new implementation.

How Data is Distributed

Shards are distributed throughout an InfluxDB cluster. A shard holds the data for any number of time series for a contiguous block of time. The length of time and which series are mapped to a shard is determined by the shard spaces.

Currently, if you want to distribute data across a cluster, you need to set the split on your shard spaces to be > 1. The ideal number for split is probably number_of_servers / replication_factor. Later releases will remove the split setting completely and automatically distribute data for you.

Shards are created for each new block of time. Say you have a replication factor of 3 and a split of 4 and shard durations of 7 days. At the beginning of a new 7 day period we’d create 4 shards and make sure each shard exists on 3 different servers. So if we have a cluster of 12 machines, each one would get a shard.

Locating a series in a cluster

The algorithm for determining where the data for a series lives in a cluster is very simple.

1. For the time period, look up the shards (if split is 1 there will be 1 shard. If 2, 2, etc.) 2. Hash the series name % number of shards 3. Use that shard index to get the shard

Write Hot Spots

You may have noticed that all the data for a given series for a block of time will end up mapping itself to a single server (or replication group). The way to scale data ingest in InfluxDB is to create many series. You can later combine these together at query time through merge queries.

If you have an event stream with many thousands of events per second, a simple way to distribute it out is just to do something like events. + random(10) so you’d split it into 10 different series that can be merged together at query time.