TCP retransmission timeout

Each pair of nodes in a cluster communicates via a number of TCP connections which remain open until one of the nodes shuts down or communication between the nodes is disrupted by a failure in the underlying infrastructure.

TCP provides reliable communication over occasionally-unreliable networks by hiding temporary network disruptions from the communicating applications. Your operating system will retransmit any lost messages a number of times before informing the sender of any problem. Most Linux distributions default to retransmitting any lost packets 15 times. Retransmissions back off exponentially, so these 15 retransmissions take over 900 seconds to complete. This means it takes Linux many minutes to detect a network partition or a failed node with this method. Windows defaults to just 5 retransmissions which corresponds with a timeout of around 6 seconds.

The Linux default allows for communication over networks that may experience very long periods of packet loss, but this default is excessive for production networks within a single data centre as is the case for most Elasticsearch clusters. Highly-available clusters must be able to detect node failures quickly so that they can react promptly by reallocating lost shards, rerouting searches and perhaps electing a new master node. Linux users should therefore reduce the maximum number of TCP retransmissions.

You can decrease the maximum number of TCP retransmissions to 5 by running the following command as root. Five retransmissions corresponds with a timeout of around six seconds.

sysctl -w net.ipv4.tcp_retries2=5

To set this value permanently, update the net.ipv4.tcp_retries2 setting in /etc/sysctl.conf. To verify after rebooting, run sysctl net.ipv4.tcp_retries2.

This setting applies to all TCP connections and will affect the reliability of communication with systems outside your cluster too. If your cluster communicates with external systems over an unreliable network then you may need to select a higher value for net.ipv4.tcp_retries2. For this reason, Elasticsearch does not adjust this setting automatically.

Related configuration

Elasticsearch also implements its own internal health checks with timeouts that are much shorter than the default retransmission timeout on Linux. Since these are application-level health checks their timeouts must allow for application-level effects such as garbage collection pauses. You should not reduce any timeouts related to these application-level health checks.

You must also ensure your network infrastructure does not interfere with the long-lived connections between nodes, even if those connections appear to be idle. Devices which drop connections when they reach a certain age are a common source of problems to Elasticsearch clusters, and must not be used.